feat: duration seconds; required fields; type-validated insertions/updates/retrievals

- Duration literals now accept s/m/h/d/w (was m/h/d/w): 'BEHAVES { ttl 30s }' is 30s, and 'within 30s' temporal expressions parse. - Definition fields are REQUIRED by default ('field: type'); 'field: type?' marks a field optional. addNode enforces presence on insert, getSchema exposes per-field requiredness, and updateNodeData still validates the provided fields' types. - Provider-returned edges are validated against the fact's declared typing: a value-carrying fact must return { value, possibility } with a value of the declared type (bare-number shorthand is rejected); a non-value fact must not carry a value; and every possibility must lie in [0, 1]. Violations throw a clear provider-authoring error instead of silently injecting malformed edges. Tests: DSLRuntimeTyping (duration units, required-field enforcement, schema requiredness, value-type + shape + possibility validation).
feat: provider-cache bypass knob; reject non-final comparator chain steps
2026-08-03 16:00:49 -07:00 · 2026-08-03 15:43:59 -07:00 · 2026-08-03 15:25:33 -07:00 · 2026-08-03 14:32:15 -07:00 · 2026-08-03 13:36:16 -07:00 · 2026-08-03 12:45:16 -07:00
25 changed files with 4646 additions and 1365 deletions
@@ -26,6 +26,8 @@ jobs:
          echo "//hub.kl1.tenere.ai/api/packages/Arbiter/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc
          echo "@tenere:registry=https://hub.kl1.tenere.ai/api/packages/Tenere/npm/" >> .npmrc
          echo "//hub.kl1.tenere.ai/api/packages/Tenere/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc
+          echo "@rigor:registry=https://hub.kl1.tenere.ai/api/packages/Rigor/npm/" >> .npmrc
+          echo "//hub.kl1.tenere.ai/api/packages/Rigor/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc

      - run: npm ci

@@ -50,6 +52,8 @@ jobs:
          echo "//hub.kl1.tenere.ai/api/packages/Arbiter/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc
          echo "@tenere:registry=https://hub.kl1.tenere.ai/api/packages/Tenere/npm/" >> .npmrc
          echo "//hub.kl1.tenere.ai/api/packages/Tenere/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc
+          echo "@rigor:registry=https://hub.kl1.tenere.ai/api/packages/Rigor/npm/" >> .npmrc
+          echo "//hub.kl1.tenere.ai/api/packages/Rigor/npm/:_authToken=${{ secrets.PACKAGE_TOKEN }}" >> .npmrc

      - run: npm ci

@@ -2,3 +2,5 @@
 //hub.kl1.tenere.ai/api/packages/Arbiter/npm/:_authToken=${PACKAGE_TOKEN}
@tenere:registry=https://hub.kl1.tenere.ai/api/packages/Tenere/npm/
 //hub.kl1.tenere.ai/api/packages/Tenere/npm/:_authToken=${PACKAGE_TOKEN}
+@rigor:registry=https://hub.kl1.tenere.ai/api/packages/Rigor/npm/
+//hub.kl1.tenere.ai/api/packages/Rigor/npm/:_authToken=${PACKAGE_TOKEN}
@@ -1,24 +1,25 @@
 {
  "name": "@arbiter/evidence-dsl",
-  "version": "1.0.0",
+  "version": "1.6.0",
  "lockfileVersion": 3,
  "requires": true,
  "packages": {
    "": {
      "name": "@arbiter/evidence-dsl",
-      "version": "1.0.0",
+      "version": "1.6.0",
      "license": "ISC",
      "dependencies": {
-        "@arbiter/core": "^1.0.1"
+        "@arbiter/core": "^1.0.4"
      },
      "devDependencies": {
+        "@rigor/core": "^3.1.0",
        "peggy": "^5.0.6"
      }
    },
    "node_modules/@arbiter/core": {
-      "version": "1.0.1",
-      "resolved": "https://hub.kl1.tenere.ai/api/packages/Arbiter/npm/%40arbiter%2Fcore/-/1.0.1/core-1.0.1.tgz",
-      "integrity": "sha512-BfIv6vRKsuJR39WBkBGxl2PQ/LDoW/J1u4UeFrog0AbeTSTvRqhxG1+its2pAcUvcQXq99ZOccFlK3XgNRKt4Q==",
+      "version": "1.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Arbiter/npm/%40arbiter%2Fcore/-/1.0.4/core-1.0.4.tgz",
+      "integrity": "sha512-1zXy3mZACjwELptsV8QpYvZJmMU7BhUQ4FsKNfp3/oJDwjfMqAImcAF3mJ+HNZpQnLQXScmELTwow5VduvN27g==",
      "license": "ISC",
      "dependencies": {
        "@tenere/pltc-core": "^0.6.3",
@@ -39,6 +40,268 @@
        "node": ">=20.8"
      }
    },
+    "node_modules/@rigor/analysis": {
+      "version": "0.0.7",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fanalysis/-/0.0.7/analysis-0.0.7.tgz",
+      "integrity": "sha512-CH9g9gU5P+aWZfneL7l91Q8wgZ0kwteYTpOdShqi6VFg2aoOSk4HMl8RTzkO4suxWa2VQ/mMnFJ/TAiHwJOobA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/instrument": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/artifact": {
+      "version": "0.1.3",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fartifact/-/0.1.3/artifact-0.1.3.tgz",
+      "integrity": "sha512-/SP5veJOpZ7y0dEBYOm+iBKQnBqZbTQzCXGd/LvQvmDtMGL7KJv2UbU7blnWp9CIwD8qU8t9U9c9oKO+qupUcA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/benchmark": {
+      "version": "0.0.6",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fbenchmark/-/0.0.6/benchmark-0.0.6.tgz",
+      "integrity": "sha512-uCCtEhCeY0t81aSgvV1Q7D3cgNU1WufEGJ/vcljhjXnx05hdMWnlK0NPdI/qkWRbRiz+4sSpcGOGO9NBFk7ExQ==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/artifact": "*",
+        "@rigor/complexity": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/complexity": {
+      "version": "0.0.7",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fcomplexity/-/0.0.7/complexity-0.0.7.tgz",
+      "integrity": "sha512-RfrkXLykb37WFwmB6OfZoXYMEZJLG78r2HqfDB27jWQmGt5l/4Hb9Zh77dEKl2CsUx2fgd5NafdDl9qFtTE6bg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/core": {
+      "version": "3.1.0",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fcore/-/3.1.0/core-3.1.0.tgz",
+      "integrity": "sha512-gUuq2YuaM0LNjdmHABcRbEW6FvPhO+U1ErokQ6ftPKpPBb38vDc+pPqu7KgGtr3R9NDU3Vj9OBV1Wq9AlzfraA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/analysis": "*",
+        "@rigor/artifact": "*",
+        "@rigor/benchmark": "*",
+        "@rigor/complexity": "*",
+        "@rigor/fault": "*",
+        "@rigor/fuzzer": "*",
+        "@rigor/gen": "*",
+        "@rigor/instrument": "*",
+        "@rigor/linearizability": "*",
+        "@rigor/model": "*",
+        "@rigor/network": "*",
+        "@rigor/probe": "*",
+        "@rigor/prop": "*",
+        "@rigor/reporters": "*",
+        "@rigor/rng": "*",
+        "@rigor/search": "*",
+        "@rigor/shrink": "*",
+        "@rigor/spec": "*",
+        "@rigor/storage": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/fault": {
+      "version": "0.0.5",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Ffault/-/0.0.5/fault-0.0.5.tgz",
+      "integrity": "sha512-MxzAEk8tx/Nc9CL3dQZZwKXj7eqBgnfTC9dI9TiU9b1u3zW7EIoL28vGqtrr1ucqSTeNwD8BGGAenan2tvb1ZQ==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/rng": "*",
+        "@rigor/shrink": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/fuzzer": {
+      "version": "0.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Ffuzzer/-/0.0.4/fuzzer-0.0.4.tgz",
+      "integrity": "sha512-cJu9bL0GqpkTsIne9PLMcsLtRfBlDz18DqPiarSi+MetH9R0joh7C+fp/e4guo0y+Q2RBJZwulA8gOh9AwgEaQ==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/artifact": "*",
+        "@rigor/rng": "*",
+        "@rigor/shrink": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/gen": {
+      "version": "0.0.11",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fgen/-/0.0.11/gen-0.0.11.tgz",
+      "integrity": "sha512-PJG40PtiRJRbNDajdpGaachLYYR+tkSK/0ToIbG5BlG+l/enw7g80hkH674Vw4jeWoftzuwe2NM4dC9rj9rHbg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/rng": "*"
+      }
+    },
+    "node_modules/@rigor/instrument": {
+      "version": "0.0.5",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Finstrument/-/0.0.5/instrument-0.0.5.tgz",
+      "integrity": "sha512-GXJa6kcEAbd86yM57hR+2g0vcxJlpBfBOoMldWk3mzLmAcCkyYXNSjJekTAwNovfUrtIcnYgQ187Zj3A2d8Tcw==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/probe": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/linearizability": {
+      "version": "0.0.5",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Flinearizability/-/0.0.5/linearizability-0.0.5.tgz",
+      "integrity": "sha512-Ruw+/Zc1hiMfs6i/4qYxYuvUsBtCm5SUmpuXrrixSRglvJ9PYfL/H8xb/Pk3Ve1dFBj8TQN6M3AnwCeaJF2/tg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE"
+    },
+    "node_modules/@rigor/model": {
+      "version": "0.2.13",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fmodel/-/0.2.13/model-0.2.13.tgz",
+      "integrity": "sha512-eZ/888p1zGJGCBFVMA9suvS2mUNX+iHhZn+Uwl9lWa9CSHVRUWH9dTYYB+Qr1hNrCSGIDjV7rOC34wzT+QA4fw==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/artifact": "*",
+        "@rigor/complexity": "*",
+        "@rigor/fault": "*",
+        "@rigor/gen": "*",
+        "@rigor/probe": "*",
+        "@rigor/rng": "*",
+        "@rigor/shrink": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/network": {
+      "version": "0.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fnetwork/-/0.0.4/network-0.0.4.tgz",
+      "integrity": "sha512-iNmiQA5QiCNkXnjAMPibGYzDk3Lf5hsgwXQGwRcSSVlZNBfVVw5fURzTsNen3TeAnqh9nozJnhD52yp3zdfP6Q==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/fault": "*",
+        "@rigor/rng": "*",
+        "@rigor/scheduler": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/probe": {
+      "version": "0.0.6",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fprobe/-/0.0.6/probe-0.0.6.tgz",
+      "integrity": "sha512-vUGY6vwciwYqxMaUYa3L2IIwipnnXdPANG8GONx+PrlII2eSbMCjb7BLZGso4yG/iaheRR0JwAnjV8mvtvUOOA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/fault": "*",
+        "@rigor/gen": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/prop": {
+      "version": "0.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fprop/-/0.0.4/prop-0.0.4.tgz",
+      "integrity": "sha512-z3+STJPeNpz080ZfFbw5h6LpNFjCNnSymoQPUsjyaSovq3euyJrGx9sqtosVgVVE7Rzld6K4QwchJSJOqIi73w==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/artifact": "*",
+        "@rigor/complexity": "*",
+        "@rigor/gen": "*",
+        "@rigor/probe": "*",
+        "@rigor/rng": "*",
+        "@rigor/shrink": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/reporters": {
+      "version": "0.0.5",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Freporters/-/0.0.5/reporters-0.0.5.tgz",
+      "integrity": "sha512-+YYfOMlYDsmGIi9rKiJm7ZAptQ0rTeTU1Sp2uVG9UJteuVKW2eKZIC6TeJuM1qvjQmh7/IzBkl3PZVX9pKnYaw==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/artifact": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/rng": {
+      "version": "0.0.3",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Frng/-/0.0.3/rng-0.0.3.tgz",
+      "integrity": "sha512-OFERK5HlI6eV1yUadHoz9n88twqsd/lstAxMzBWFJiN1CCM0xG7IHUYMBCnYK6Di3TTcAarFNFL16Bchm3LrKg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE"
+    },
+    "node_modules/@rigor/scheduler": {
+      "version": "0.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fscheduler/-/0.0.4/scheduler-0.0.4.tgz",
+      "integrity": "sha512-tY7KASEfufOw4L1vIGeD1bTArCv+zAEPqT/XqiXFz+fckGxY8Nux6zDhGcjkur3qWOUNoWap6KOjSZSORHhXiw==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/fault": "*",
+        "@rigor/rng": "*",
+        "@rigor/shrink": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/search": {
+      "version": "0.1.3",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fsearch/-/0.1.3/search-0.1.3.tgz",
+      "integrity": "sha512-NAq5zbUSc6QQO4YIUFq16+hPtrtkNrjAn75lH/4VDPpbCdQPbKvS6zsQ3v954u1u7I6x2jI5uilpX6/xfr69yA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/gen": "*",
+        "@rigor/shrink": "*"
+      }
+    },
+    "node_modules/@rigor/shrink": {
+      "version": "0.0.7",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fshrink/-/0.0.7/shrink-0.0.7.tgz",
+      "integrity": "sha512-z0yaZqmyysn40BBqAnqTQlymLuaIIU5u2BijjayX0+XZxhhcPIyvs4PISzJ4nmfkx2yE6qKABe2HHub9wVVUQw==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE"
+    },
+    "node_modules/@rigor/spec": {
+      "version": "2.0.1",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fspec/-/2.0.1/spec-2.0.1.tgz",
+      "integrity": "sha512-0/lg19KenJy1e8D87+WqaMRh8Xad4h209/CgK6ofCQapwz/IcriMPHgOZZbHMA4C8G3B6/luVbSVT0EHZoJ8qg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/gen": "*"
+      }
+    },
+    "node_modules/@rigor/storage": {
+      "version": "0.0.4",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Fstorage/-/0.0.4/storage-0.0.4.tgz",
+      "integrity": "sha512-7XapdgmvEg/C/kERldqFlw3mT1JkpjPMNK6fEcB2Ggx7+UR3ZeE9GyaodJ29my61b28SrQUWxLBHT5bm0bIVaA==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE",
+      "dependencies": {
+        "@rigor/fault": "*",
+        "@rigor/rng": "*",
+        "@rigor/scheduler": "*",
+        "@rigor/trace": "*"
+      }
+    },
+    "node_modules/@rigor/trace": {
+      "version": "0.0.7",
+      "resolved": "https://hub.kl1.tenere.ai/api/packages/Rigor/npm/%40rigor%2Ftrace/-/0.0.7/trace-0.0.7.tgz",
+      "integrity": "sha512-vCLF+WTSdy0pwe9sTonPt8HnNeSkK9U5SAxNpdIeiT+IBP74t9K7JoHMmfo7ka+S0puWecDjjNtWCSqtvKkjNg==",
+      "dev": true,
+      "license": "SEE LICENSE IN LICENSE"
+    },
    "node_modules/@tenere/graph-core": {
      "version": "1.0.1",
      "resolved": "https://hub.kl1.tenere.ai/api/packages/Tenere/npm/%40tenere%2Fgraph-core/-/1.0.1/graph-core-1.0.1.tgz",
@@ -1,6 +1,6 @@
 {
  "name": "@arbiter/evidence-dsl",
-  "version": "1.0.0",
+  "version": "1.10.0",
  "description": "Evidence DSL v2 compiler: translates the natural Evidence DSL (ADR-000) into @arbiter/core relation configurations.",
  "license": "ISC",
  "type": "module",
@@ -13,6 +13,7 @@
    "./parser/GeneratedParser": "./src/parser/GeneratedParser.js",
    "./generator/RuleGenerator": "./src/generator/RuleGenerator.js",
    "./validation/DSLValidation": "./src/validation/DSLValidation.js",
+    "./runtime/DSLRuntime": "./src/runtime/DSLRuntime.js",
    "./interpreter/BuiltInFunctions": "./src/interpreter/BuiltInFunctions.js"
  },
  "files": [
@@ -23,9 +24,10 @@
    "generate:parser": "node scripts/generate-parser.js"
  },
  "dependencies": {
-    "@arbiter/core": "^1.0.1"
+    "@arbiter/core": "^1.0.4"
  },
  "devDependencies": {
+    "@rigor/core": "^3.1.0",
    "peggy": "^5.0.6"
  }
 }
@@ -10,6 +10,7 @@ export class RuleGenerator {
    this.generatedRules = new Map();
    this.errors = [];
    this.dependencyIndex = new Map();
+    this.evidenceNames = new Set();
  }

  /**
@@ -38,6 +39,14 @@ export class RuleGenerator {
        this.generateFactConfig(fact);
      });

+      // Resolve evidence composition: a rule that references another derived
+      // evidence (WHEN can_read(user, doc) where can_read is an evidence) is
+      // lowered in place to that evidence's own config — compile-time inlining
+      // (a linker pass), so the engine evaluates a fully-resolved config tree
+      // and never needs a sub-query traversal mechanism. Forward references are
+      // handled because every evidence config is built before this pass runs.
+      this.resolveEvidenceReferences();
+
      // Apply generated rules to arbiter
      this.applyRulesToArbiter();

@@ -62,6 +71,7 @@ export class RuleGenerator {
   */
  generateEvidenceRules(evidence) {
    const relationName = evidence.name;
+    this.evidenceNames.add(relationName);
    const ruleConfig = this.buildRuleConfig(evidence);
    
    if (ruleConfig) {
@@ -145,6 +155,7 @@ export class RuleGenerator {
        for (const step of rule.steps) {
          if (typeof step === 'string') targetSet.add(step);
          else if (step && typeof step.relation === 'string') targetSet.add(step.relation);
+          else if (step && step.rule) collect(step.rule, targetSet);
        }
      }
      if (rule.type === 'relational_comparator') {
@@ -162,6 +173,7 @@ export class RuleGenerator {
          for (const child of ruleList) collect(child, targetSet);
        }
        if (node?.rule) collect(node.rule, targetSet);
+        if (node?.direct) collect(node.direct, targetSet);
      }
    };

@@ -258,41 +270,42 @@ export class RuleGenerator {
    
    // Handle single statement evidence
    if (statements.length === 1) {
-      return this.buildSingleStatementRule(statements[0]);
+      return this.buildSingleStatementRule(statements[0], evidence);
    }

    // Handle multiple statements with logical operators
-    return this.buildLogicalRule(statements);
+    return this.buildLogicalRule(statements, evidence);
  }

  /**
   * Build rule configuration for a single statement
   * @param {BaseNode} statement - Statement to build rule for
+   * @param {Object} evidence - Evidence definition (params inform pattern/operand lowering)
   * @returns {Object|null} Rule configuration or null
   */
-  buildSingleStatementRule(statement) {
+  buildSingleStatementRule(statement, evidence) {
    switch (statement.type) {
      case 'DirectEvidence':
-        return this.buildDirectRule(statement);
+        return this.buildDirectRule(statement, evidence);
      case 'PatternMatch':
-        return this.buildPatternMatchRule(statement);
+        return this.buildPatternMatchRule(statement, evidence);
      case 'DefeasibleLogic':
-        return this.buildDefeasibleRule(statement);
+        return this.buildDefeasibleRule(statement, evidence);
      case 'Fusion':
        return this.buildFusionRule(statement);
      case 'PredicateCall':
-        return this.buildPredicateRule(statement);
+        return this.buildPredicateRule(statement, evidence);
      case 'UnaryExpression':
        return this.buildUnaryRule(statement);
      case 'BinaryExpression':
        // Top-level comparator — emit a relational_comparator rule. RF-24 closure.
-        return this.buildRuleFromExpressionNode(statement);
+        return this.buildRuleFromExpressionNode(statement, evidence);
      case 'Expression':
        // Handle expressions that might be predicate calls
        if (statement.type === 'PredicateCall') {
-          return this.buildPredicateRule(statement);
+          return this.buildPredicateRule(statement, evidence);
        }
-        return this.buildRuleFromExpressionNode(statement);
+        return this.buildRuleFromExpressionNode(statement, evidence);
      default:
        this.errors.push(`Unsupported statement type: ${statement.type}`);
        return null;
@@ -331,11 +344,24 @@ export class RuleGenerator {
   * @param {BaseNode[]} statements - Statements to combine
   * @returns {Object|null} Rule configuration or null
   */
-  buildLogicalRule(statements) {
+  buildLogicalRule(statements, evidence) {
+    const defeasible = statements.filter(s => s && s.type === 'DefeasibleLogic');
+    const others = statements.filter(s => s && s.type !== 'DefeasibleLogic');
+
+    // Defeasible levels (NEVER / REQUIRES / ALWAYS / WHEN / UNLESS) form ONE
+    // five-level hierarchy (ADR-000), not separate ANDed rules. A standalone
+    // NEVER-only rule contributes 0 whether or not it fires, so ANDing the
+    // levels separately would always yield 0. Merge all defeasible statements
+    // into a single config; any non-defeasible statements become the base
+    // grant (when) that the defeaters and requirements gate.
+    if (defeasible.length > 0) {
+      return this.buildMergedDefeasibleRule(defeasible, others, evidence);
+    }
+
    const rules = [];
-    
-    statements.forEach(statement => {
-      const rule = this.buildSingleStatementRule(statement);
+
+    others.forEach(statement => {
+      const rule = this.buildSingleStatementRule(statement, evidence);
      if (rule) {
        rules.push(rule);
      }
@@ -361,12 +387,91 @@ export class RuleGenerator {
    };
  }

+  /**
+   * Merge defeasible-level statements into a single five-level rule config.
+   * The ADR-000 hierarchy is never > requires > strict (always) > when > unless;
+   * each level accumulates its conditions and the whole thing evaluates as one
+   * defeasible rule rather than a conjunction of level-only rules.
+   */
+  buildMergedDefeasibleRule(defeasibleStatements, otherStatements, evidence) {
+    const neverRules = [];
+    const alwaysRules = [];
+    const requiresRules = [];
+    const whenRules = [];
+    const unlessRules = [];
+
+    for (const st of defeasibleStatements) {
+      const condition = this.buildRuleFromExpression(st.condition, evidence);
+      const defeater = st.defeater ? this.buildRuleFromExpression(st.defeater, evidence) : null;
+      switch (st.logicType) {
+        case 'NEVER':
+          if (condition) neverRules.push(condition);
+          break;
+        case 'ALWAYS':
+          if (condition) alwaysRules.push(condition);
+          break;
+        case 'REQUIRES':
+          if (condition) requiresRules.push(condition);
+          break;
+        case 'WHEN':
+          if (condition) whenRules.push(condition);
+          if (defeater) unlessRules.push(defeater);
+          break;
+        case 'UNLESS':
+          if (condition) unlessRules.push(condition);
+          break;
+        default:
+          this.errors.push(`Unsupported defeasible logic type: ${st.logicType}`);
+      }
+    }
+
+    // Non-defeasible statements in the same body act as the base grant
+    // (when) that NEVER/REQUIRES/UNLESS gate.
+    if (otherStatements.length > 0) {
+      const baseRules = otherStatements
+        .map(s => this.buildSingleStatementRule(s, evidence))
+        .filter(Boolean);
+      if (baseRules.length === 1) {
+        whenRules.push(baseRules[0]);
+      } else if (baseRules.length > 1) {
+        whenRules.push({
+          type: 'logical',
+          intersection: { rules: baseRules, aggregator: 'min' }
+        });
+      }
+    }
+
+    const rule = { type: 'logical' };
+    if (neverRules.length > 0) {
+      rule.never = { union: { rules: neverRules, aggregator: 'max' } };
+    }
+    if (requiresRules.length > 0) {
+      rule.requires = { union: { rules: requiresRules, aggregator: 'min' } };
+    }
+    if (alwaysRules.length > 0) {
+      rule.always = {
+        direct: alwaysRules.length === 1
+          ? alwaysRules[0]
+          : { type: 'logical', intersection: { rules: alwaysRules, aggregator: 'min' } },
+        aggregator: 'min'
+      };
+    }
+    if (whenRules.length > 0) {
+      rule.when = { intersection: { rules: whenRules, aggregator: 'min' } };
+    }
+    if (unlessRules.length > 0) {
+      rule.unless = { union: { rules: unlessRules, aggregator: 'max' } };
+    }
+
+    return rule;
+  }
+
  /**
   * Build direct rule configuration
   * @param {DirectEvidenceNode} directEvidence - Direct evidence statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildDirectRule(directEvidence) {
+  buildDirectRule(directEvidence, evidence) {
    if (!directEvidence.predicate) {
      this.errors.push('Direct evidence must have a predicate');
      return null;
@@ -374,12 +479,31 @@ export class RuleGenerator {

    const predicate = directEvidence.predicate;
    const relation = predicate.name;
-    
-    return {
+
+    const rule = {
      type: 'direct',
      relation: relation,
      reverse: false
    };
+
+    // Unary predicate calls check the relation as a self-edge on the call's
+    // subject entity (the graph stores unary facts as self-edges). The subject
+    // entity may be the evidence's SUBJECT or its OBJECT parameter — mark the
+    // matching rewrite flag.
+    const evidenceParams = (evidence && evidence.params) || [];
+    const objectVar = evidenceParams[1] && evidenceParams[1].name;
+    const argName = a => a && (a.name !== undefined ? a.name : a.value);
+    const args = predicate.arguments || [];
+    if (objectVar !== undefined) {
+      const hasObjectArg = args.some(a => argName(a) === objectVar);
+      if (args.length === 1 && argName(args[0]) === objectVar) {
+        rule._subjectIsObject = true;
+      } else if (!hasObjectArg) {
+        rule._subjectAsObject = true;
+      }
+    }
+
+    return rule;
  }

  /**
@@ -387,7 +511,7 @@ export class RuleGenerator {
   * @param {PatternMatchNode} patternMatch - Pattern match statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildPatternMatchRule(patternMatch) {
+  buildPatternMatchRule(patternMatch, evidence) {
    if (!patternMatch.predicate) {
      this.errors.push('Pattern match must have a predicate');
      return null;
@@ -396,27 +520,119 @@ export class RuleGenerator {
    const predicate = patternMatch.predicate;
    const relation = predicate.name;

-    // Membership/hierarchy predicates map to TupleToUsersetRule / ParentRule
-    // regardless of body shape — those have priority over chain detection.
+    // Structural classification (ADR-000 §Mapping to Engine Rule Types): the
+    // two-hop pattern "P(args) { Q(args) }" binds an intermediate via a Wildcard.
+    // The predicate whose args include the OBJECT parameter is the object-side
+    // hop. Object-side = outer → tuple_to_userset (user → computed → intermediate
+    // → tupleset → object). Object-side = inner → chain (user → outer →
+    // intermediate → inner → object). This mirrors Zanzibar's tuple-to-userset
+    // vs. two-hop path semantics and fixes the previous heuristic that routed
+    // every outer-wildcard pattern to chain (owner(*g, doc) { member_of(user, g) }
+    // was emitted as a chain and could never match).
+    const inner = this._singleInnerPredicate(patternMatch);
+    const evidenceParams = (evidence && evidence.params) || [];
+    const userVar = evidenceParams[0] && evidenceParams[0].name;
+    const objectVar = evidenceParams[1] && evidenceParams[1].name;
+
+    // Nested PatternMatch bodies ("P(user, *a) { Q(a, *b) { R(b, doc) } }") are
+    // fixed-length multi-hop PATHS — a chain whose steps are the flattened
+    // predicate sequence [P, Q, R], not transitive-closure multi_hop over one
+    // relation. Chain handles N steps; multi_hop only walks a single relation.
+    if (this._isNestedPattern(patternMatch)) {
+      return this.buildNestedChainRule(patternMatch, inner);
+    }
+
+    if (inner) {
+      const argName = a => a && (a.name !== undefined ? a.name : a.value);
+      const outerHasObject = objectVar !== undefined &&
+        (predicate.args || []).some(a => argName(a) === objectVar);
+      const innerHasObject = objectVar !== undefined &&
+        (inner.args || []).some(a => argName(a) === objectVar);
+
+      if (outerHasObject && !innerHasObject) {
+        return this.buildTupleToUsersetRule(patternMatch, inner);
+      }
+      if (innerHasObject && !outerHasObject) {
+        return this.buildChainRule(patternMatch, inner);
+      }
+    }
+
+    // Fallback: membership/hierarchy naming hints (evidence params unavailable
+    // or both predicates reference the object — keep legacy behavior).
    if (this.isMembershipPredicate(predicate)) {
-      return this.buildTupleToUsersetRule(patternMatch);
+      return this.buildTupleToUsersetRule(patternMatch, inner);
    }
    if (this.isHierarchyPredicate(predicate)) {
      return this.buildParentRule(patternMatch);
    }

-    // Chain detection: ADR-000 ChainRule shape is "works_in(p, *d) { has_access(d, r) }".
-    // The outer PatternMatch has a Wildcard binding, and its body contains a single
-    // PredicateCall (no DefeasibleLogic wrapping, no nested PatternMatch). Treat that
-    // as a chain: two-hop traversal through the wildcard intermediate. RF-24 closure
-    // (parallel to RF-22/RF-23 — DSL→engine mapping gap surfaced by rigor coverage).
    if (this._isChainPattern(patternMatch)) {
-      return this.buildChainRule(patternMatch);
+      return this.buildChainRule(patternMatch, inner);
    }

    return this.buildMultiHopRule(patternMatch);
  }

+  /**
+   * Extract the single inner PredicateCall of a PatternMatch body (null if the
+   * body has multiple statements, is nested, or is wrapped in logic operators).
+   */
+  _singleInnerPredicate(patternMatch) {
+    if (!patternMatch.body || !Array.isArray(patternMatch.body.statements)) return null;
+    const stmts = patternMatch.body.statements;
+    if (stmts.length !== 1) return null;
+    if (stmts[0].type === 'PredicateCall') return stmts[0];
+    return null;
+  }
+
+  /**
+   * True when the PatternMatch body is itself a nested PatternMatch
+   * ("P(user, *a) { Q(a, *b) { R(b, doc) } }") — a multi-hop path.
+   */
+  _isNestedPattern(patternMatch) {
+    if (!patternMatch.body || !Array.isArray(patternMatch.body.statements)) return false;
+    const stmts = patternMatch.body.statements;
+    return stmts.length === 1 && stmts[0].type === 'PatternMatch';
+  }
+
+  /**
+   * Flatten a nested PatternMatch into its linear predicate sequence
+   * [P, Q, ..., R] where R is the object-side hop.
+   */
+  _flattenPatternSteps(patternMatch, acc = []) {
+    const predicate = patternMatch.predicate;
+    if (!predicate || !predicate.name) return acc;
+    acc.push(predicate.name);
+    if (patternMatch.body && Array.isArray(patternMatch.body.statements) &&
+        patternMatch.body.statements.length === 1) {
+      const child = patternMatch.body.statements[0];
+      if (child && child.type === 'PredicateCall' && child.name) {
+        acc.push(child.name);
+      } else if (child && child.type === 'PatternMatch') {
+        this._flattenPatternSteps(child, acc);
+      }
+    }
+    return acc;
+  }
+
+  /**
+   * Build a chain rule from a nested (multi-hop path) PatternMatch.
+   * "P(user, *a) { Q(a, *b) { R(b, doc) } }" → { type: 'chain', steps: [P, Q, R] }.
+   */
+  buildNestedChainRule(patternMatch) {
+    const steps = this._flattenPatternSteps(patternMatch);
+    if (steps.length < 2) {
+      this.errors.push('Nested pattern match must yield at least two steps');
+      return null;
+    }
+    return {
+      type: 'chain',
+      steps,
+      aggregator: 'max',
+      collectValues: true
+    };
+  }
+
  /**
   * Detect the ChainRule shape: a PatternMatch whose body contains exactly one
   * PredicateCall and uses a Wildcard arg to bind the intermediate. The predicate
@@ -435,17 +651,19 @@ export class RuleGenerator {

  /**
   * Build chain rule configuration (ADR-000 ChainRule).
-   * Compiles "works_in(p, *d) { has_access(d, doc) }" into
+   * Compiles "works_in(user, *d) { has_access(d, doc) }" into
   *   { type: 'chain', steps: ['works_in', 'has_access'] }
-   * The intermediate wildcard binds the two predicates' arguments.
+   * The intermediate wildcard binds the two predicates' arguments. Step order is
+   * [userSide, objectSide]: the outer predicate connects user → intermediate,
+   * the inner predicate connects intermediate → object.
   */
-  buildChainRule(patternMatch) {
+  buildChainRule(patternMatch, inner) {
    const steps = [];
    steps.push(patternMatch.predicate.name);

-    const inner = patternMatch.body.statements[0];
-    if (inner && inner.type === 'PredicateCall' && inner.name) {
-      steps.push(inner.name);
+    const innerPredicate = inner || (patternMatch.body.statements[0]);
+    if (innerPredicate && innerPredicate.type === 'PredicateCall' && innerPredicate.name) {
+      steps.push(innerPredicate.name);
    }

    return {
@@ -457,18 +675,31 @@ export class RuleGenerator {
  }

  /**
-   * Build tuple-to-userset rule configuration
-   * @param {PatternMatchNode} patternMatch - Pattern match statement
-   * @returns {Object|null} Rule configuration or null
+   * Build tuple-to-userset rule configuration (ADR-000 TupleToUsersetRule).
+   * Compiles "owner(*g, doc) { member_of(user, g) }" into
+   *   {
+   *     type: 'tuple_to_userset',
+   *     tuplesetRelation: 'owner',     // object-side hop: intermediate → object
+   *     computedRelation: 'member_of', // user-side hop: user → intermediate
+   *     tuplesetDirection: 'in',       // intermediates hold the tupleset edge TO the object
+   *     reverse: false
+   *   }
+   * The tupleset edge direction follows the wildcard position in the outer
+   * predicate: wildcard as first arg (owner(*g, doc)) means the intermediate is
+   * the edge source ('in' — fetch edges with dst = object); wildcard as second
+   * arg (owner(doc, *g)) means the object is the source ('out').
   */
-  buildTupleToUsersetRule(patternMatch) {
+  buildTupleToUsersetRule(patternMatch, inner) {
    const predicate = patternMatch.predicate;
    const relation = predicate.name;
-    
+    const computedRelation = inner && inner.name ? inner.name : relation;
+    const wildcardIndex = (predicate.args || []).findIndex(a => a && a.type === 'Wildcard');
+
    return {
      type: 'tuple_to_userset',
-      tuplesetRelation: 'owner', // Default, could be inferred from context
-      computedRelation: relation,
+      tuplesetRelation: relation,
+      computedRelation,
+      tuplesetDirection: wildcardIndex === 0 ? 'in' : 'out',
      reverse: false,
      earlyExitThreshold: 0.95,
      maxIntermediates: patternMatch.limit || 10
@@ -518,19 +749,19 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildDefeasibleRule(defeasibleLogic) {
+  buildDefeasibleRule(defeasibleLogic, evidence) {
    const logicType = defeasibleLogic.logicType;

    if (logicType === 'NEVER') {
-      return this.buildNeverRule(defeasibleLogic);
+      return this.buildNeverRule(defeasibleLogic, evidence);
    } else if (logicType === 'ALWAYS') {
-      return this.buildStrictRule(defeasibleLogic);
+      return this.buildStrictRule(defeasibleLogic, evidence);
    } else if (logicType === 'WHEN') {
-      return this.buildDefeasibleRuleWithDefeater(defeasibleLogic);
+      return this.buildDefeasibleRuleWithDefeater(defeasibleLogic, evidence);
    } else if (logicType === 'UNLESS') {
-      return this.buildDefeaterRule(defeasibleLogic);
+      return this.buildDefeaterRule(defeasibleLogic, evidence);
    } else if (logicType === 'REQUIRES') {
-      return this.buildRequirementRule(defeasibleLogic);
+      return this.buildRequirementRule(defeasibleLogic, evidence);
    }

    this.errors.push(`Unsupported defeasible logic type: ${logicType}`);
@@ -542,8 +773,8 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildNeverRule(defeasibleLogic) {
-    const condition = this.buildRuleFromExpression(defeasibleLogic.condition);
+  buildNeverRule(defeasibleLogic, evidence) {
+    const condition = this.buildRuleFromExpression(defeasibleLogic.condition, evidence);

    return {
      type: 'logical',
@@ -561,8 +792,8 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildStrictRule(defeasibleLogic) {
-    const condition = this.buildRuleFromExpression(defeasibleLogic.condition);
+  buildStrictRule(defeasibleLogic, evidence) {
+    const condition = this.buildRuleFromExpression(defeasibleLogic.condition, evidence);
    
    return {
      type: 'logical',
@@ -578,9 +809,9 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildDefeasibleRuleWithDefeater(defeasibleLogic) {
-    const condition = this.buildRuleFromExpression(defeasibleLogic.condition);
-    const defeater = this.buildRuleFromExpression(defeasibleLogic.defeater);
+  buildDefeasibleRuleWithDefeater(defeasibleLogic, evidence) {
+    const condition = this.buildRuleFromExpression(defeasibleLogic.condition, evidence);
+    const defeater = this.buildRuleFromExpression(defeasibleLogic.defeater, evidence);
    
    const rule = {
      type: 'logical',
@@ -609,8 +840,8 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildDefeaterRule(defeasibleLogic) {
-    const condition = this.buildRuleFromExpression(defeasibleLogic.condition);
+  buildDefeaterRule(defeasibleLogic, evidence) {
+    const condition = this.buildRuleFromExpression(defeasibleLogic.condition, evidence);
    
    return {
      type: 'logical',
@@ -628,8 +859,8 @@ export class RuleGenerator {
   * @param {DefeasibleLogicNode} defeasibleLogic - Defeasible logic statement
   * @returns {Object|null} Rule configuration or null
   */
-  buildRequirementRule(defeasibleLogic) {
-    const condition = this.buildRuleFromExpression(defeasibleLogic.condition);
+  buildRequirementRule(defeasibleLogic, evidence) {
+    const condition = this.buildRuleFromExpression(defeasibleLogic.condition, evidence);
    
    return {
      type: 'logical',
@@ -698,19 +929,21 @@ export class RuleGenerator {
   * @param {BaseNode} expression - Expression to build rule from
   * @returns {Object|null} Rule configuration or null
   */
-  buildRuleFromExpression(expression) {
+  buildRuleFromExpression(expression, evidence) {
    if (!expression) {
      return null;
    }

    if (expression.type === 'Predicate') {
-      return this.buildDirectRuleFromPredicate(expression);
+      return this.buildDirectRuleFromPredicate(expression, evidence);
    } else if (expression.type === 'Expression') {
-      return this.buildRuleFromExpressionNode(expression);
+      return this.buildRuleFromExpressionNode(expression, evidence);
    } else if (expression.type === 'PredicateCall') {
-      return this.buildPredicateRule(expression);
+      return this.buildPredicateRule(expression, evidence);
    } else if (expression.type === 'UnaryExpression') {
      return this.buildUnaryRule(expression);
+    } else if (expression.type === 'BinaryExpression') {
+      return this.buildRuleFromExpressionNode(expression, evidence);
    }

    this.errors.push(`Unsupported expression type: ${expression.type}`);
@@ -722,47 +955,190 @@ export class RuleGenerator {
   * @param {PredicateNode} predicate - Predicate to build rule from
   * @returns {Object|null} Rule configuration or null
   */
-  buildDirectRuleFromPredicate(predicate) {
-    const expanded = this._expandPredicate(predicate.name);
-    if (expanded) return expanded;
-
-    return {
+  buildDirectRuleFromPredicate(predicate, evidence) {
+    const rule = {
      type: 'direct',
      relation: predicate.name,
      reverse: false
    };
+
+    const evidenceParams = (evidence && evidence.params) || [];
+    const objectVar = evidenceParams[1] && evidenceParams[1].name;
+    if (objectVar !== undefined && !(predicate.args || []).some(a =>
+      a && a.type === 'Variable' && a.name === objectVar)) {
+      rule._subjectAsObject = true;
+    }
+
+    return rule;
  }

-  _expandPredicate(predicateName) {
-    const existingConfig = this.generatedRules.get(predicateName) || this.arbiter?.relationConfigs?.get(predicateName);
-    if (!existingConfig) return null;
-    if (!existingConfig.union && !existingConfig.intersection && !existingConfig.exclusion) return null;
+  _expandPredicate() {
+    // Replaced by resolveEvidenceReferences() (the compile-time evidence
+    // composition pass). Predicate references are now emitted as direct rules
+    // and inlined during resolution, which also handles forward references and
+    // preserves the correct _subjectAsObject scoping.
+  }

-    const logicalKey = existingConfig.union ? 'union' : existingConfig.intersection ? 'intersection' : 'exclusion';
-    const subRules = Array.isArray(existingConfig[logicalKey]?.rules)
-      ? existingConfig[logicalKey].rules
-      : Array.isArray(existingConfig[logicalKey]) ? existingConfig[logicalKey] : [];
+  /**
+   * Evidence composition pass. Every rule that references a DERIVED evidence
+   * (e.g. `WHEN can_read(user, doc)` where can_read is itself an evidence) is
+   * rewritten to inline that evidence's own config. This runs after all
+   * evidence configs are generated, so forward references resolve; cycles are
+   * detected and reported. The engine therefore evaluates a fully-resolved,
+   * acyclic config tree — no runtime sub-query traversal is needed.
+   */
+  resolveEvidenceReferences() {
+    for (const name of this.evidenceNames) {
+      if (!this.generatedRules.has(name)) continue;
+      const stack = new Set([name]);
+      const resolved = this._resolveRule(this.generatedRules.get(name), stack);
+      this.generatedRules.set(name, resolved);
+      this._annotateDependencies(name, resolved);
+    }
+  }

-    if (subRules.length === 0) return null;
+  /**
+   * Recursively rewrite a rule tree, inlining references to derived evidence
+   * configs. `stack` holds the evidence names currently being expanded so a
+   * cyclic reference (A → B → A) is detected and reported.
+   */
+  _resolveRule(rule, stack) {
+    if (!rule || typeof rule !== 'object') return rule;
+    if (Array.isArray(rule)) return rule.map(r => this._resolveRule(r, stack));

-    const expandedRules = subRules.map(r => {
-      if (r && r.type === 'direct') return { type: 'direct', relation: r.relation, reverse: !!r.reverse };
-      if (typeof r === 'string') return { type: 'direct', relation: r, reverse: false };
-      return null;
-    }).filter(Boolean);
+    // Direct rule referencing a derived evidence → inline its resolved config.
+    if (rule.type === 'direct' && rule.relation) {
+      const ref = rule.relation;
+      if (this.evidenceNames.has(ref)) {
+        const referencedConfig = this.generatedRules.get(ref);
+        if (referencedConfig) {
+          if (stack.has(ref)) {
+            this.errors.push(`Cyclic evidence reference involving '${ref}'. Evidence composition must be acyclic.`);
+            return rule;
+          }
+          const refStack = new Set(stack);
+          refStack.add(ref);
+          const resolvedRef = this._resolveRule(referencedConfig, refStack);
+          if (resolvedRef) {
+            const clone = this._deepCloneRule(resolvedRef);
+            if (rule._subjectAsObject) clone._subjectAsObject = true;
+            return clone;
+          }
+        }
+      }
+      return rule;
+    }

-    if (expandedRules.length === 0) return null;
+    // Recurse into logical / defeasible / nested containers: rule-lists
+    // (union/intersection/exclusion/never/requires/when/unless .rules) and
+    // single nested rules (always.direct, comparator operands).
+    const out = { ...rule };
+    for (const key of ['union', 'intersection', 'exclusion', 'never', 'always', 'requires', 'when', 'unless', 'direct', 'rule']) {
+      const node = out[key];
+      if (!node || typeof node !== 'object') continue;
+      if (Array.isArray(node)) {
+        out[key] = node.map(r => this._resolveRule(r, stack));
+        continue;
+      }
+      const next = { ...node };
+      if (Array.isArray(next.rules)) {
+        next.rules = next.rules.map(r => this._resolveRule(r, stack));
+      }
+      if (next.union && Array.isArray(next.union.rules)) {
+        next.union = { ...next.union, rules: next.union.rules.map(r => this._resolveRule(r, stack)) };
+      }
+      if (next.intersection && Array.isArray(next.intersection.rules)) {
+        next.intersection = { ...next.intersection, rules: next.intersection.rules.map(r => this._resolveRule(r, stack)) };
+      }
+      if (next.direct && typeof next.direct === 'object') {
+        next.direct = this._resolveRule(next.direct, stack);
+      }
+      if (next.rule && typeof next.rule === 'object') {
+        next.rule = this._resolveRule(next.rule, stack);
+      }
+      out[key] = next;
+    }
+    if (out.type === 'relational_comparator') {
+      if (out.left?.rule) out.left = { ...out.left, rule: this._resolveRule(out.left.rule, stack) };
+      if (out.right?.rule) out.right = { ...out.right, rule: this._resolveRule(out.right.rule, stack) };
+    }
+    // Chain steps may reference a derived evidence; expand those steps
+    // (direct evidence → underlying relation, chain evidence → spliced steps).
+    if (rule.type === 'chain' && Array.isArray(out.steps)) {
+      out.steps = this._expandChainSteps(out.steps, stack);
+    }
+    return out;
+  }

-    return {
-      type: 'logical',
-      [logicalKey]: {
-        rules: expandedRules,
-        aggregator: existingConfig[logicalKey]?.aggregator || 'min'
-      },
-      // Flag to tell the evaluator: this expanded sub-predicate is unary —
-      // use the subject as the object instead of inheriting the parent's object.
-      _subjectAsObject: true
-    };
+  /**
+   * Expand chain steps that reference a derived evidence:
+   *   - direct evidence  → rename the step to the underlying relation
+   *     (member_of(user,*g){ group_read(g,doc) } where group_read = can_view
+   *      becomes step 'can_view');
+   *   - chain evidence   → splice its steps into this chain (flattening)
+   *     (a step that is itself a sub-path becomes its steps, preserving the
+   *      linear source→…→object traversal);
+   *   - logical / defeasible / comparator evidence → only expressible as a
+   *     FINAL condition-gated step (the object is known, so the engine can
+   *     verify the condition at (intermediate, object) instead of traversing
+   *     an edge). Emitted as a `{ rule: <config> }` step the ChainRule
+   *     evaluates as a condition hop. Non-final such steps are a compile
+   *     error: a condition cannot discover intermediate nodes.
+   */
+  _expandChainSteps(steps, stack) {
+    const out = [];
+    for (let idx = 0; idx < steps.length; idx++) {
+      const step = steps[idx];
+      const stepName = typeof step === 'string' ? step : step.relation;
+      if (stepName && this.evidenceNames.has(stepName)) {
+        if (stack.has(stepName)) {
+          this.errors.push(`Cyclic evidence reference involving '${stepName}'. Evidence composition must be acyclic.`);
+          out.push(step);
+          continue;
+        }
+        const referencedConfig = this.generatedRules.get(stepName);
+        if (referencedConfig) {
+          const refStack = new Set(stack);
+          refStack.add(stepName);
+          const resolved = this._resolveRule(referencedConfig, refStack);
+          if (resolved.type === 'direct' && resolved.relation && resolved.relation !== stepName) {
+            out.push(typeof step === 'string'
+              ? resolved.relation
+              : { ...step, relation: resolved.relation });
+            continue;
+          }
+          if (resolved.type === 'chain' && Array.isArray(resolved.steps)) {
+            out.push(...this._expandChainSteps(resolved.steps, refStack));
+            continue;
+          }
+          if (resolved.type === 'relational_comparator' && idx !== steps.length - 1) {
+            // A comparator compares values at (src, candidate) but provides no
+            // candidate set — it cannot enumerate intermediate nodes, so only
+            // a FINAL comparator step (verified at the known object) lowers.
+            this.errors.push(`Chain step '${stepName}' references a comparator evidence at a non-final position. ` +
+              'Comparators can only be the final chain step (the object is known); intermediate positions are not enumerable.');
+            out.push(step);
+            continue;
+          }
+          // Condition step: inline the evidence's config as a rule step. As the
+          // FINAL step the engine verifies it at (intermediate, object); as an
+          // INTERMEDIATE step the engine EXPANDS it from the current node
+          // (rule-based reachability) and continues from each discovered node.
+          out.push({ rule: this._deepCloneRule(resolved), conditionStep: true });
+          continue;
+        }
+      }
+      out.push(step);
+    }
+    return out;
+  }
+
+  _deepCloneRule(rule) {
+    try {
+      return structuredClone(rule);
+    } catch {
+      return JSON.parse(JSON.stringify(rule));
+    }
  }

  /**
@@ -770,7 +1146,7 @@ export class RuleGenerator {
   * @param {ExpressionNode} expression - Expression to build rule from
   * @returns {Object|null} Rule configuration or null
   */
-  buildRuleFromExpressionNode(expression) {
+  buildRuleFromExpressionNode(expression, evidence) {
    if (expression.type === 'AttributeAccess') {
      return this.buildAttributeRule(expression);
    } else if (expression.type === 'PredicateCall') {
@@ -782,7 +1158,7 @@ export class RuleGenerator {
      // Route BinaryExpression with comparator operators here so the
      // evaluator can run a fuzzy interval comparison instead of treating
      // them as logical truth values. RF-24 closure.
-      return this.buildRelationalComparatorRule(expression);
+      return this.buildRelationalComparatorRule(expression, evidence);
    }

    this.errors.push(`Unsupported expression type: ${expression.type}`);
@@ -812,10 +1188,10 @@ export class RuleGenerator {
   *     minRulePossibility: 0
   *   }
   */
-  buildRelationalComparatorRule(binaryExpression) {
+  buildRelationalComparatorRule(binaryExpression, evidence) {
    const comparator = binaryExpression.operator;
-    const left = this._buildComparatorOperand(binaryExpression.left);
-    const right = this._buildComparatorOperand(binaryExpression.right);
+    const left = this._buildComparatorOperand(binaryExpression.left, evidence);
+    const right = this._buildComparatorOperand(binaryExpression.right, evidence);
    if (!left || !right) {
      this.errors.push(`Comparator operands must resolve to predicate calls (operator=${comparator})`);
      return null;
@@ -833,29 +1209,52 @@ export class RuleGenerator {
  }

  /**
-   * Wrap a BinaryExpression side into a relational_comparator operand. The
-   * operand's `rule` field is the original predicate call (preserving reference
-   * semantics so the inner rule's evaluator can resolve its values). `extractValue`
-   * tells the evaluator to read the relation's `value` field rather than its
-   * `possibility`, which is what `personAge(p)` / `docMinAge(d)` semantics require.
+   * Lower a comparator operand to a core-evaluable config. The operand's `rule`
+   * MUST be a real rule configuration (the engine's RuleEvaluator only accepts
+   * configs — raw AST nodes evaluate to 0). Per the core's operand contract:
+   *
+   *   userRisk(user)  → { rule: { type: 'direct', relation: 'userRisk' }, extractValue: true }   (user perspective, auto)
+   *   riskLimit(doc)  → { rule: { type: 'direct', relation: 'riskLimit' }, extractValue: true, evaluateFrom: 'object' }
+   *
+   * The evaluateFrom side is derived from the evidence parameter positions:
+   * params[0] is the subject (user), params[1] is the object. A predicate call
+   * whose first arg is the object variable reads its value from the object
+   * perspective; anything else defaults to the user perspective.
+   *
+   * Literal value args (userRisk(user, 5)) annotate the operand with
+   * `expectedValue` — the declared value the caller expects the relation to
+   * carry. The engine compares resolved relation values; the DSLRuntime wrapper
+   * may enforce expectedValue as an additional gate.
   */
-  _buildComparatorOperand(side) {
+  _buildComparatorOperand(side, evidence) {
    if (!side) return null;
    if (side.type === 'PredicateCall') {
-      return {
-        rule: side,
-        extractValue: true,
-        evaluatorFrom: 'auto'
-      };
-    }
-    if (side.type === 'AttributeAccess') {
-      // user.age — treat the attribute path as a "measure" reference
-      return {
-        rule: side,
+      const operand = {
+        rule: { type: 'direct', relation: side.name, reverse: false },
        extractValue: true,
        evaluatorFrom: 'auto',
-        attributePath: side.getAttributePath ? side.getAttributePath() : null
+        valueRelation: side.name
      };
+      const evidenceParams = (evidence && evidence.params) || [];
+      const userVar = evidenceParams[0] && evidenceParams[0].name;
+      const objectVar = evidenceParams[1] && evidenceParams[1].name;
+      const firstArg = (side.args || [])[0];
+      const firstArgName = firstArg && (firstArg.name !== undefined ? firstArg.name : firstArg.value);
+      if (objectVar !== undefined && firstArgName === objectVar) {
+        operand.evaluateFrom = 'object';
+      }
+      const literalArg = (side.args || []).find(a => a && a.type === 'Literal');
+      if (literalArg) {
+        operand.expectedValue = literalArg.value;
+      }
+      return operand;
+    }
+    if (side.type === 'AttributeAccess') {
+      // Attribute paths cannot lower to a relation config — the engine's
+      // operand machinery reads relation `value` fields, not node attributes.
+      // Reject loudly instead of emitting an unevaluable rule.
+      this.errors.push('Comparator operands cannot be attribute accesses (use a value-carrying relation instead)');
+      return null;
    }
    return null;
  }
@@ -880,22 +1279,40 @@ export class RuleGenerator {
   * @param {ExpressionNode} expression - Function expression
   * @returns {Object|null} Rule configuration or null
   */
-  buildPredicateRule(expression) {
+  buildPredicateRule(expression, evidence) {
    const predicateName = expression.name;
    if (expression.challenge) {
      return this.buildChallengeRule(expression, null);
    }

-    // Expand composite (logical) predicate references into their direct
-    // leaf components so the optimizer can flatten to a correct direct_list.
-    const expanded = this._expandPredicate(predicateName);
-    if (expanded) return expanded;
-
-    return {
+    const rule = {
      type: 'direct',
      relation: predicateName,
      reverse: false
    };
+
+    // Unary predicate calls check the relation as a self-edge on the call's
+    // subject entity (the graph stores unary facts as self-edges). The subject
+    // entity may be the evidence's SUBJECT or its OBJECT parameter:
+    //   banned(user)   in can_open(user, doc)   -> self-edge on the user
+    //   trusted(other) in peer_trusted(user, other) -> self-edge on the other
+    // Mark _subjectAsObject (subject-as-object on the subject entity) or
+    // _subjectIsObject (the subject entity IS the object parameter) so the
+    // engine rewrites the pair accordingly.
+    const evidenceParams = (evidence && evidence.params) || [];
+    const objectVar = evidenceParams[1] && evidenceParams[1].name;
+    const argName = a => a && (a.name !== undefined ? a.name : a.value);
+    if (objectVar !== undefined) {
+      const args = expression.args || [];
+      const hasObjectArg = args.some(a => a && a.type === 'Variable' && a.name === objectVar);
+      if (args.length === 1 && argName(args[0]) === objectVar) {
+        rule._subjectIsObject = true;
+      } else if (!hasObjectArg) {
+        rule._subjectAsObject = true;
+      }
+    }
+
+    return rule;
  }

  buildWithinRule(expression) {
@@ -54,11 +54,14 @@ Definition "A type definition"
    }

 Field
-  = name:Identifier _ ":" _ fieldType:Type _ isArray:("[]")? _ behavior:Behavior? _ cache:CacheDirective? {
+  = name:Identifier _ ":" _ fieldType:Type optional:("?")? _ isArray:("[]")? _ behavior:Behavior? _ cache:CacheDirective? {
      return {
        type: "Field",
        name,
        fieldType,
+        // `field: type` is REQUIRED on node insert; `field: type?` is optional.
+        // Presence is enforced by the DSLRuntime when a node is created.
+        required: !optional,
        isArray: !!isArray,
        behavior: behavior || null,
        cache: cache || null
@@ -248,6 +251,13 @@ BehaviorAnnotation
  = "BEHAVES" __ "AS" __ behavior:("edge" / "transitive" / "hierarchical" / "symmetrical_graph") {
      return { type: "BehaviorAnnotation", behavior };
    }
+  / "BEHAVES" __ "{" _ behavior:(TTLBehavior) _ "}" {
+      // Fact-level freshness: `fact balance(user, amount) BEHAVES { ttl 1h }`
+      // declares the relation's value-freshness window, which the runtime uses
+      // as the provider-result cache TTL. The behavior is wrapped like the
+      // `BEHAVES AS` form so consumers read `behavior.behaviorType`.
+      return { type: "BehaviorAnnotation", behavior };
+    }

 FactProperty
  = "transitive" { return "transitive"; }
@@ -385,7 +395,7 @@ Boolean "A boolean literal"
  = value:("true" / "false") { return { type: "Literal", value: value === "true" }; }

 Duration "A time duration literal"
-  = value:([0-9]+ ("h" / "d" / "w" / "m")) { return { type: "Literal", value: text(), unit: text().slice(-1) }; }
+  = value:([0-9]+ ("s" / "m" / "h" / "d" / "w")) { return { type: "Literal", value: text(), unit: text().slice(-1) }; }


 // -- Core Tokens & Whitespace --
@@ -15,6 +15,9 @@ export { RuleGenerator } from './generator/RuleGenerator.js';
 // Validation
 export { validateDslText } from './validation/DSLValidation.js';

+// Runtime
+export { DSLRuntime } from './runtime/DSLRuntime.js';
+
 // All AST nodes
 export * from './nodes/index.js';

@@ -0,0 +1,726 @@
+import { DSLCompiler } from '../DSLCompiler.js';
+
+const VALUE_TYPES = new Set(['string', 'number', 'boolean', 'timestamp', 'duration', 'object', 'any']);
+
+/**
+ * DSLRuntime — higher-order wrapper combining the Evidence DSL with an
+ * @arbiter/core Arbiter.
+ *
+ * The DSL declares a typed schema: `definition` blocks (entity types with
+ * typed fields), `fact` declarations (relations with typed params, optional
+ * `*` injectable marker), and `evidence` rules (relations the runtime can
+ * check). A raw Arbiter accepts untyped inserts; this wrapper adds the
+ * DSL-informed layer:
+ *
+ *   - schema introspection: getSchema() exposes the compiled type system
+ *     (entity types/fields, facts, evidence, dependencies, providers);
+ *   - typed mutations: addNode / updateNodeData / addRelation / updateRelation
+ *     validate their arguments against the compiled schema — known types,
+ *     known relations, matching param types, typed field values — before
+ *     mutating the arbiter; removeNode / removeRelation pass through;
+ *   - per-relation data retrieval: registerFact(relation, asyncFn) registers a
+ *     provider that retrieves the missing partial-graph edges for a fact; a
+ *     bounded retrieval loop runs providers to a fixed point so a provider's
+ *     edges can satisfy another required fact;
+ *   - DSL-informed check: derives the evidence's injectable facts, retrieves
+ *     them via providers, injects them into a partial graph, and delegates to
+ *     the arbiter; require() throws on denial for middleware use.
+ *
+ * Trust boundary follows the core: caller-supplied evidence (partial graph /
+ * provider results) is trusted, never policed; only structure is validated.
+ */
+export class DSLRuntime {
+  /**
+   * @param {object} arbiter - An @arbiter/core Arbiter instance.
+   * @param {object} options
+   * @param {object} options.factProviders - relation → async fn(subject, object, ctx)
+   *   returning a boolean, possibility number, { possibility, value }, or an
+   *   array of { src, relation, dst, possibility, value } partial-graph edges.
+   * @param {object} options.policy
+   * @param {boolean} options.policy.strictTypes - throw on unknown types/relations
+   *   (default true; false degrades to arbiter behavior for undeclared names).
+   */
+  constructor(arbiter, options = {}) {
+    this.arbiter = arbiter;
+    this.compiler = new DSLCompiler(this.arbiter);
+    this.factProviders = options.factProviders || {};
+    this.strictTypes = options.policy?.strictTypes !== false;
+    this.program = null;
+    this.types = new Map();      // typeName -> { fields: Map(field -> {type,isArray}) }
+    this.relations = new Map();  // relation -> { kind: 'fact'|'evidence', params, injectable, ttlMs }
+    this.dependsOn = new Map();  // evidence relation -> Set(fact relations)
+
+    // Provider-result cache: relation|subject|object -> { edges, fetchedAt }.
+    // Provider retrieval is a data-store read (balance lookups, session
+    // checks, etc.) — caching results with a time expiry avoids hammering the
+    // underlying store on every check. The clock is injectable (default wall
+    // clock) and drives cache freshness, mirroring the core's unpinned-clock
+    // contract.
+    this.providerCache = new Map();
+    this.clock = typeof options.clock === 'function' ? options.clock : (() => Date.now());
+    // Default provider-result TTL in ms (0 disables caching).
+    this.defaultProviderCacheTTL = options.policy?.providerCacheTTL ?? options.providerCacheTTL ?? 30_000;
+    // Provider caching is a STORE-RETRIEVAL cache (wall-clock), deliberately
+    // independent of the caller's decision `{ now }` — a provider returns the
+    // store's current data, not a time-travel snapshot. Callers who pin time
+    // or otherwise want fresh retrieval can disable it per-check
+    // (options.cacheProviderResults: false) or globally (policy).
+    this.cacheProviderResults = options.policy?.cacheProviderResults ?? options.cacheProviderResults ?? true;
+    // Per-fact overrides (ms). DSL-declared ttl behaviors are indexed here too.
+    this.factTTLs = new Map(Object.entries(options.factTTLs || {}));
+  }
+
+  /**
+   * Compile a DSL program and index its schema. Returns this for chaining.
+   * @param {string} dsl
+   * @param {string} name
+   */
+  compile(dsl, name) {
+    const result = this.compiler.compile(dsl, name);
+    if (!result.success) {
+      const error = new Error(`DSLRuntime compile failed: ${(result.errors || []).join('; ')}`);
+      error.errors = result.errors || [];
+      throw error;
+    }
+    this.program = result.program;
+    this._indexSchema();
+    return this;
+  }
+
+  // ---------------------------------------------------------------------------
+  // Schema introspection
+  // ---------------------------------------------------------------------------
+
+  /**
+   * A serializable snapshot of the compiled type system: entity types with
+   * typed fields, facts, evidence (with their dependencies), and registered
+   * providers. Callers can use this to render forms, build clients, or audit
+   * a compiled program without reaching into the internal Maps.
+   */
+  getSchema() {
+    const types = [...this.types.entries()].map(([name, { fields }]) => ({
+      name,
+      fields: [...fields.entries()].map(([fieldName, f]) => ({
+        name: fieldName,
+        type: f.type,
+        isArray: f.isArray,
+        required: f.required !== false
+      }))
+    }));
+    const facts = [...this.relations.entries()]
+      .filter(([, r]) => r.kind === 'fact')
+      .map(([name, r]) => ({ name, params: r.params, injectable: r.injectable }));
+    const evidence = [...this.relations.entries()]
+      .filter(([, r]) => r.kind === 'evidence')
+      .map(([name, r]) => ({
+        name,
+        params: r.params,
+        dependsOn: [...(this.dependsOn.get(name) || [])]
+      }));
+    return { types, facts, evidence, providers: this.registeredFacts() };
+  }
+
+  /** All relation names declared by the program (facts + evidence). */
+  relationNames() {
+    return [...this.relations.keys()];
+  }
+
+  _indexSchema() {
+    this.types.clear();
+    this.relations.clear();
+    this.dependsOn.clear();
+
+    for (const def of this.program.definitions || []) {
+      const fields = new Map();
+      for (const field of def.fields || []) {
+        fields.set(field.name, { type: field.fieldType, isArray: !!field.isArray, required: field.required !== false });
+      }
+      this.types.set(def.name, { fields });
+    }
+
+    for (const fact of this.program.facts || []) {
+      const ttlMs = this._ttlFromBehavior(fact.behavior);
+      this.relations.set(fact.name, {
+        kind: 'fact',
+        params: (fact.params || []).map(p => ({ name: p.name, type: p.paramType, isArray: !!p.isArray })),
+        injectable: !!fact.injectable,
+        ttlMs
+      });
+    }
+
+    for (const ev of this.program.evidence || []) {
+      this.relations.set(ev.name, {
+        kind: 'evidence',
+        params: (ev.params || []).map(p => ({ name: p.name, type: p.paramType, isArray: !!p.isArray })),
+        injectable: false
+      });
+    }
+
+    // Index each evidence's fact dependencies from the compiled arbiter configs.
+    for (const ev of this.program.evidence || []) {
+      const config = this.arbiter.relationConfigs.get(ev.name);
+      const deps = new Set();
+      const collect = (rule) => {
+        if (!rule || typeof rule !== 'object') return;
+        if (rule.type === 'direct' && rule.relation) deps.add(rule.relation);
+        if (rule.type === 'tuple_to_userset') {
+          if (rule.tuplesetRelation) deps.add(rule.tuplesetRelation);
+          if (rule.computedRelation) deps.add(rule.computedRelation);
+        }
+        if (rule.type === 'chain' && Array.isArray(rule.steps)) {
+          for (const s of rule.steps) {
+            if (typeof s === 'string') deps.add(s);
+            else if (s && s.relation) deps.add(s.relation);
+            else if (s && s.rule) collect(s.rule);
+          }
+        }
+        if (rule.type === 'parent' && rule.parentRelation) deps.add(rule.parentRelation);
+        if (rule.type === 'multi_hop' && rule.relation) deps.add(rule.relation);
+        if (rule.type === 'relational_comparator') {
+          collect(rule.left?.rule);
+          collect(rule.right?.rule);
+          if (rule.left?.valueRelation) deps.add(rule.left.valueRelation);
+          if (rule.right?.valueRelation) deps.add(rule.right.valueRelation);
+        }
+        for (const key of ['union', 'intersection', 'exclusion', 'never', 'always', 'requires', 'when', 'unless']) {
+          const node = rule[key];
+          if (!node) continue;
+          if (Array.isArray(node.rules)) for (const c of node.rules) collect(c);
+          if (Array.isArray(node.union?.rules)) for (const c of node.union.rules) collect(c);
+          if (Array.isArray(node.intersection?.rules)) for (const c of node.intersection.rules) collect(c);
+          if (node.direct) collect(node.direct);
+          if (node.rule) collect(node.rule);
+        }
+      };
+      if (config && Array.isArray(config.dependsOn)) {
+        for (const d of config.dependsOn) deps.add(d);
+      } else {
+        collect(config);
+      }
+      this.dependsOn.set(ev.name, deps);
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Provider registration (per-relation data retrieval)
+  // ---------------------------------------------------------------------------
+
+  /**
+   * Register (or replace) an async provider for a relation name. When a check
+   * needs that relation's facts and they are not in the graph, the provider is
+   * invoked to retrieve the missing partial-graph edges.
+   *
+   * @param {string} relation - fact relation name
+   * @param {Function} provider - async (subject, object, ctx) => edges
+   */
+  registerFact(relation, provider) {
+    if (typeof provider !== 'function') {
+      throw new Error(`DSLRuntime: provider for '${relation}' must be a function`);
+    }
+    this.factProviders[relation] = provider;
+    // A new provider supersedes any cached retrieval for this fact.
+    this.invalidateProviderCache(relation);
+    return this;
+  }
+
+  /** Remove a registered provider. */
+  unregisterFact(relation) {
+    delete this.factProviders[relation];
+    this.invalidateProviderCache(relation);
+    return this;
+  }
+
+  /** Relation names that currently have a registered provider. */
+  registeredFacts() {
+    return Object.keys(this.factProviders);
+  }
+
+  // ---------------------------------------------------------------------------
+  // Provider-result caching
+  // ---------------------------------------------------------------------------
+
+  /**
+   * Set a per-fact provider-result TTL (ms). Overrides the policy default and
+   * the DSL-declared ttl behavior for that fact.
+   */
+  setFactTTL(relation, ms) {
+    this.factTTLs.set(relation, ms);
+    return this;
+  }
+
+  /** The effective provider-result TTL (ms) for a fact: DSL > per-fact > policy default. */
+  _ttlFor(relation) {
+    if (this.factTTLs.has(relation)) return this.factTTLs.get(relation);
+    const meta = this.relations.get(relation);
+    if (meta && meta.ttlMs != null) return meta.ttlMs;
+    return this.defaultProviderCacheTTL;
+  }
+
+  /**
+   * Invalidate cached provider results — all, or for a single relation.
+   * Callers use this when the underlying data store changes out-of-band.
+   */
+  invalidateProviderCache(relation) {
+    if (relation === undefined) {
+      this.providerCache.clear();
+      return this;
+    }
+    const prefix = `${relation}\u0000`;
+    for (const key of [...this.providerCache.keys()]) {
+      if (key.startsWith(prefix)) this.providerCache.delete(key);
+    }
+    return this;
+  }
+
+  _providerCacheKey(relation, subject, object) {
+    return `${relation}\u0000${subject}\u0000${object}`;
+  }
+
+  _providerCacheGet(relation, subject, object) {
+    const ttl = this._ttlFor(relation);
+    if (ttl <= 0) return null;
+    const entry = this.providerCache.get(this._providerCacheKey(relation, subject, object));
+    if (!entry) return null;
+    if (this.clock() - entry.fetchedAt >= ttl) {
+      this.providerCache.delete(this._providerCacheKey(relation, subject, object));
+      return null;
+    }
+    return entry;
+  }
+
+  _providerCacheSet(relation, subject, object, edges) {
+    const ttl = this._ttlFor(relation);
+    if (ttl <= 0) return;
+    this.providerCache.set(this._providerCacheKey(relation, subject, object), {
+      edges,
+      fetchedAt: this.clock()
+    });
+  }
+
+  /** Convert a DSL `BEHAVES { ttl <duration> }` behavior (or `BEHAVES AS`) into ms. */
+  _ttlFromBehavior(behavior) {
+    if (!behavior || typeof behavior !== 'object') return null;
+    const b = behavior.behavior || behavior;
+    if (b && b.behaviorType === 'ttl' && b.duration) {
+      const n = parseInt(String(b.duration.value), 10);
+      const mult = { s: 1000, m: 60_000, h: 3600_000, d: 86_400_000, w: 604_800_000 }[b.duration.unit];
+      if (!Number.isNaN(n) && mult) return n * mult;
+    }
+    return null;
+  }
+
+  /**
+   * Normalize a provider result (boolean / number / { possibility, value } /
+   * array of edge objects) into an array of partial-graph edge objects. The
+   * destination follows the DSL fact's declared shape: unary and value-carrying
+   * facts are self-edges on the subject; binary entity facts go subject → object.
+   * Provider-returned edges are validated against the fact's declared typing:
+   * a value-carrying fact must return an object with a value of the declared
+   * type, and possibilities must be in [0, 1]. A violation throws — it is a
+   * provider-authoring error, not a denial.
+   */
+  _normalizeProviderEdges(result, factMeta, fact, user, object) {
+    const edges = Array.isArray(result) ? result : [result];
+    const secondParamType = factMeta.params[1] && factMeta.params[1].type;
+    const isValueFact = factMeta.params.length >= 2 && this._isValueType(secondParamType);
+    const defaultDst = isValueFact ? user : (factMeta.params.length >= 2 ? object : user);
+    const label = `provider for '${fact}'`;
+    const out = [];
+    for (const edge of edges) {
+      const normalized = typeof edge === 'boolean' || typeof edge === 'number'
+        ? (() => {
+            if (isValueFact) {
+              throw new Error(`DSLRuntime: ${label} is a value-carrying fact — return { value, possibility } (got a bare ${typeof edge === 'number' ? 'number' : 'boolean'})`);
+            }
+            const possibility = edge === true ? 1 : edge;
+            this._checkPossibility(possibility, label);
+            return { src: user, dst: defaultDst, possibility };
+          })()
+        : (() => {
+            const possibility = edge.possibility ?? 1;
+            this._checkPossibility(possibility, label);
+            if (edge.value !== undefined) {
+              if (!isValueFact) {
+                throw new Error(`DSLRuntime: ${label} returned a value for a non-value fact '${fact}'`);
+              }
+              this._checkScalarValue(secondParamType, edge.value, `${label}.value`);
+            } else if (isValueFact) {
+              throw new Error(`DSLRuntime: ${label} must supply a 'value' of type ${secondParamType}`);
+            }
+            return {
+              ...(edge.relation ? { relation: edge.relation } : {}),
+              src: edge.src ?? user,
+              dst: edge.dst ?? defaultDst,
+              possibility,
+              ...(edge.value !== undefined ? { value: edge.value } : {}),
+              ...(edge.reliability !== undefined ? { reliability: edge.reliability } : {})
+            };
+          })();
+      out.push(normalized);
+    }
+    return out;
+  }
+
+  _checkPossibility(possibility, label) {
+    if (typeof possibility !== 'number' || !Number.isFinite(possibility) || possibility < 0 || possibility > 1) {
+      throw new Error(`DSLRuntime: ${label} returned invalid possibility ${possibility} (expected a number in [0, 1])`);
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Schema validation helpers
+  // ---------------------------------------------------------------------------
+
+  _isValueType(typeName) {
+    return VALUE_TYPES.has(typeName);
+  }
+
+  _nodeType(key) {
+    const nodeId = this.arbiter.resolveNodeId(key);
+    if (nodeId === undefined) return null;
+    const node = this.arbiter.nodes.get(nodeId);
+    return node ? node.type : null;
+  }
+
+  _checkNodeExists(key, position) {
+    if (!this.arbiter.nodeIdByKey.has(key)) {
+      throw new Error(`DSLRuntime: ${position} node '${key}' does not exist`);
+    }
+  }
+
+  _checkNodeType(key, expectedType, position) {
+    if (this._isValueType(expectedType)) return; // value positions are validated separately
+    const actual = this._nodeType(key);
+    if (actual === null) {
+      this._checkNodeExists(key, position);
+      return;
+    }
+    if (actual !== expectedType) {
+      throw new Error(`DSLRuntime: ${position} node '${key}' has type '${actual}', expected '${expectedType}'`);
+    }
+  }
+
+  _checkFieldValue(field, value, path) {
+    if (field.isArray) {
+      if (!Array.isArray(value)) {
+        throw new Error(`DSLRuntime: field '${path}' must be an array of ${field.type}`);
+      }
+      for (const item of value) this._checkScalarValue(field.type, item, path);
+      return;
+    }
+    this._checkScalarValue(field.type, value, path);
+  }
+
+  _checkScalarValue(type, value, path) {
+    const ok = type === 'string' ? typeof value === 'string'
+      : type === 'number' ? typeof value === 'number'
+      : type === 'boolean' ? typeof value === 'boolean'
+      : (type === 'timestamp' || type === 'duration')
+        ? (typeof value === 'number' || typeof value === 'string')
+        : true; // object / any / entity-typed fields accept any value
+    if (!ok) {
+      throw new Error(`DSLRuntime: field '${path}' must be ${type}, got ${typeof value}`);
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Typed mutations
+  // ---------------------------------------------------------------------------
+
+  /**
+   * Insert a node, validating the type exists (when declared) and that `data`
+   * conforms to the definition's typed fields.
+   */
+  addNode(key, typeName, data = {}) {
+    if (this.types.has(typeName)) {
+      const { fields } = this.types.get(typeName);
+      for (const [name, field] of fields) {
+        // Required fields must be present on insert (`field: type` in the DSL;
+        // `field?: type` marks a field optional).
+        if (field.required && data[name] === undefined) {
+          throw new Error(`DSLRuntime: missing required field '${typeName}.${name}' on node insert`);
+        }
+        if (data[name] !== undefined) this._checkFieldValue(field, data[name], `${typeName}.${name}`);
+      }
+    } else if (this.strictTypes) {
+      throw new Error(`DSLRuntime: unknown type '${typeName}'`);
+    }
+    // A graph mutation can make previously-retrieved facts stale.
+    this.invalidateProviderCache();
+    return this.arbiter.addNode(key, typeName, data);
+  }
+
+  /**
+   * Update node data, validating fields against the node's declared type.
+   */
+  updateNodeData(key, data) {
+    const typeName = this._nodeType(key);
+    if (typeName && this.types.has(typeName)) {
+      const { fields } = this.types.get(typeName);
+      for (const [name, field] of fields) {
+        if (data[name] !== undefined) this._checkFieldValue(field, data[name], `${typeName}.${name}`);
+      }
+    }
+    this.invalidateProviderCache();
+    return this.arbiter.updateNodeData(key, data);
+  }
+
+  /** Remove a node (passthrough to the arbiter's node manager). */
+  removeNode(key) {
+    this.invalidateProviderCache();
+    if (this.arbiter.nodeManager && typeof this.arbiter.nodeManager.removeNode === 'function') {
+      return this.arbiter.nodeManager.removeNode(key);
+    }
+    return this.arbiter.removeNode?.(key);
+  }
+
+  _relationOrThrow(relation) {
+    const meta = this.relations.get(relation);
+    if (!meta) {
+      if (this.strictTypes) throw new Error(`DSLRuntime: unknown relation '${relation}'`);
+      return null;
+    }
+    return meta;
+  }
+
+  /**
+   * Insert a relation edge. Validates the relation is declared, that the
+   * subject/object nodes match the declared entity param types, and that any
+   * primitive value param is supplied in attrs.value of the correct type.
+   */
+  addRelation(src, relation, dst, attrs = {}) {
+    const meta = this._relationOrThrow(relation);
+    if (meta) {
+      this._validateRelationEndpoints(relation, meta, src, dst, attrs);
+    }
+    this.invalidateProviderCache();
+    return this.arbiter.addRelation(src, relation, dst, attrs);
+  }
+
+  /**
+   * Update a relation edge (idempotent replace). Validates like addRelation.
+   */
+  updateRelation(src, relation, dst, attrs = {}) {
+    const meta = this._relationOrThrow(relation);
+    if (meta) {
+      this._validateRelationEndpoints(relation, meta, src, dst, attrs);
+    }
+    this.invalidateProviderCache();
+    this.arbiter.removeRelation(src, relation, dst);
+    return this.arbiter.addRelation(src, relation, dst, attrs);
+  }
+
+  /** Remove a relation edge (passthrough to the arbiter). */
+  removeRelation(src, relation, dst) {
+    this.invalidateProviderCache();
+    return this.arbiter.removeRelation(src, relation, dst);
+  }
+
+  _validateRelationEndpoints(relation, meta, src, dst, attrs) {
+    const params = meta.params;
+    if (params.length === 0) {
+      throw new Error(`DSLRuntime: relation '${relation}' declares no parameters`);
+    }
+    // First param is always the subject (entity).
+    const subjectType = params[0].type;
+    if (this._isValueType(subjectType)) {
+      throw new Error(`DSLRuntime: relation '${relation}' subject param must be an entity type, got '${subjectType}'`);
+    }
+    this._checkNodeType(src, subjectType, 'subject');
+
+    if (params.length >= 2) {
+      const secondType = params[1].type;
+      if (this._isValueType(secondType)) {
+        // Value-carrying fact (e.g. session(user, token: string)): the value
+        // lives on the edge's `value` field; the graph edge is a self-edge on
+        // the subject so the value is discoverable by value extraction.
+        if (attrs.value === undefined) {
+          attrs.value = dst;
+        }
+        this._checkScalarValue(secondType, attrs.value, `${relation}.${params[1].name}`);
+        if (dst !== src) {
+          throw new Error(`DSLRuntime: value param '${params[1].name}' must be supplied as attrs.value with dst = src (self-edge), got dst '${dst}'`);
+        }
+      } else {
+        this._checkNodeType(dst, secondType, 'object');
+      }
+    }
+  }
+
+  /**
+   * The partial-graph requirements of an evidence relation: the declared
+   * injectable facts it depends on.
+   */
+  requiredFacts(relation) {
+    const deps = this.dependsOn.get(relation);
+    if (!deps) return [];
+    const required = [];
+    for (const dep of deps) {
+      const meta = this.relations.get(dep);
+      if (meta && meta.kind === 'fact' && meta.injectable) required.push(dep);
+    }
+    return required;
+  }
+
+  // ---------------------------------------------------------------------------
+  // DSL-informed check
+  // ---------------------------------------------------------------------------
+
+  /**
+   * Validate a check request against the DSL schema, derive and retrieve the
+   * evidence's injectable facts, inject them into a partial graph, and delegate
+   * to the arbiter.
+   *
+   * Providers run in a bounded fixed-point loop: each round invokes the
+   * provider for every required fact whose edges are not yet in the partial
+   * graph. Because a provider may return edges for relations other than its
+   * own name, an edge injected in one round can satisfy another required fact
+   * (or unblock another provider) in a later round. The loop stops when a
+   * round injects no new relation or the round budget is exhausted.
+   *
+   * @param {string} user - subject key
+   * @param {string} relation - evidence (or fact) relation name
+   * @param {string} object - object key
+   * @param {object} options
+   * @param {object} options.partialGraph - caller-supplied partial graph edges
+   *   ({ relations: [{ src, relation, dst, possibility, value }], nodes, challenges })
+   * @param {object} options.factProviders - per-call provider overrides
+   *   (merged over registered providers)
+   * @param {number} options.maxProviderRounds - fixed-point loop budget (default 3)
+   * @returns {object} core check result extended with { requiredFacts, providedFacts, missingFacts }
+   */
+  async check(user, relation, object, options = {}) {
+    const meta = this.relations.get(relation);
+    if (!meta) {
+      if (this.strictTypes) throw new Error(`DSLRuntime: unknown relation '${relation}'`);
+    } else if (meta.params.length === 2) {
+      this._checkNodeType(user, meta.params[0].type, 'subject');
+      this._checkNodeType(object, meta.params[1].type, 'object');
+    } else if (meta.params.length === 1) {
+      this._checkNodeType(user, meta.params[0].type, 'subject');
+    }
+
+    // Retrieval set: for an evidence, the injectable facts it depends on; for
+    // a direct FACT check, the fact itself is the retrieval target (its
+    // provider, if registered, supplies the edge — checking `owns` directly
+    // must consult the `owns` provider, not only evidence-mediated checks).
+    const required = new Set(this.requiredFacts(relation));
+    if (meta && meta.kind === 'fact') required.add(relation);
+    const requiredList = [...required];
+    const providers = { ...this.factProviders, ...(options.factProviders || {}) };
+    const maxRounds = options.maxProviderRounds ?? 3;
+    const partialRelations = [];
+    const injectedRelations = []; // { relation, edges, round }
+    const missingFacts = [];
+    const satisfied = new Set();  // facts whose edges are in the partial graph
+
+    if (options.partialGraph && Array.isArray(options.partialGraph.relations)) {
+      for (const rel of options.partialGraph.relations) {
+        partialRelations.push(rel);
+        if (rel && rel.relation) satisfied.add(rel.relation);
+      }
+    }
+
+    // Fixed-point provider retrieval loop.
+    for (let round = 1; round <= maxRounds; round++) {
+      let newRelationsThisRound = 0;
+      for (const fact of requiredList) {
+        if (satisfied.has(fact)) continue;
+        const factMeta = this.relations.get(fact);
+        const provider = providers[fact];
+
+        // Provider-result cache: reuse fresh edges without re-invoking the
+        // data store. A cached entry stores the NORMALIZED edges. Per-check
+        // provider overrides are one-off observations — they bypass the cache
+        // entirely (no read, no write) so a fresh override is never masked by
+        // a cached registered-provider result, nor does it pollute the cache.
+        // options.cacheProviderResults:false (or the policy default) disables
+        // the cache for this check.
+        const cachingEnabled = options.cacheProviderResults ?? this.cacheProviderResults;
+        const isPerCheckOverride = !!(options.factProviders && fact in options.factProviders);
+        const cacheHit = (cachingEnabled && !isPerCheckOverride) ? this._providerCacheGet(fact, user, object) : null;
+        let edges = null;
+        let fromCache = false;
+        if (cacheHit) {
+          edges = cacheHit.edges;
+          fromCache = true;
+        } else if (typeof provider === 'function') {
+          let result = null;
+          let error = null;
+          try {
+            result = await provider(user, object, {
+              relation: fact,
+              params: factMeta.params,
+              runtime: this,
+              options,
+              round,
+              alreadyInjected: [...satisfied]
+            });
+          } catch (err) {
+            error = err;
+          }
+          if (error) {
+            missingFacts.push({ relation: fact, reason: error.message });
+            satisfied.add(fact);
+            continue;
+          }
+          if (result === false || result === null || result === undefined) {
+            missingFacts.push({ relation: fact, reason: 'not_provided' });
+            satisfied.add(fact);
+            continue;
+          }
+          edges = this._normalizeProviderEdges(result, factMeta, fact, user, object);
+          if (cachingEnabled && !isPerCheckOverride) this._providerCacheSet(fact, user, object, edges);
+        } else {
+          missingFacts.push({ relation: fact, reason: 'no_provider' });
+          satisfied.add(fact);
+          continue;
+        }
+
+        // A provider may return edges for relations other than its own; the
+        // injected relation names satisfy those facts too (fixed point).
+        for (const normalized of edges) {
+          const injectedRelation = normalized.relation ?? fact;
+          partialRelations.push({ relation: injectedRelation, ...normalized });
+          satisfied.add(injectedRelation);
+        }
+        injectedRelations.push({ relation: fact, edges: edges.length, round, cacheHit: fromCache });
+        newRelationsThisRound += edges.length;
+        satisfied.add(fact);
+      }
+      if (newRelationsThisRound === 0) break;
+    }
+
+    const checkOptions = { ...options };
+    if (partialRelations.length > 0) {
+      checkOptions.partialGraph = {
+        ...(options.partialGraph || {}),
+        relations: partialRelations
+      };
+    }
+
+    const result = this.arbiter.check(user, relation, object, checkOptions);
+
+    return {
+      ...result,
+      requiredFacts: requiredList,
+      providedFacts: injectedRelations.map(r => r.relation),
+      missingFacts
+    };
+  }
+
+  /**
+   * Check and throw on denial — convenience for middleware / guards.
+   * @returns {object} the check result on success.
+   * @throws {Error} with `.result` attached when the decision denies.
+   */
+  async require(user, relation, object, options = {}) {
+    const result = await this.check(user, relation, object, options);
+    if (result.possibility <= 0) {
+      const error = new Error(`DSLRuntime: authorization denied for '${relation}' (${result.reason || 'denied'})`);
+      error.result = result;
+      throw error;
+    }
+    return result;
+  }
+}
@@ -41,6 +41,7 @@ export function validateDslText(dslText, options = {}) {
  validateSources(program, tables, errors, warnings, dslText);
  validateMeasures(program, tables, errors, warnings, dslText);
  validateEvidence(program, tables, errors, warnings, dslText);
+  validateCrossKindRelationNames(program, errors, warnings, dslText);

  return {
    success: errors.length === 0,
@@ -144,6 +145,7 @@ function validateDefinitions(program, tables, errors, warnings, source) {
 }

 function validateFacts(program, tables, errors, warnings, source) {
+  const seen = new Map();
  for (const fact of program.facts || []) {
    if (tables.builtins?.facts?.has(fact.name)) {
      errors.push(createError({
@@ -154,6 +156,16 @@ function validateFacts(program, tables, errors, warnings, source) {
        context: formatContext(source, findLocation(source, fact.name))
      }));
    }
+    if (seen.has(fact.name)) {
+      errors.push(createError({
+        message: `Duplicate fact definition '${fact.name}'.`,
+        rule: 'Each fact name must be unique within a program.',
+        fix: 'Rename one of the fact definitions to a unique name.',
+        location: findLocation(source, `fact ${fact.name}`),
+        context: formatContext(source, findLocation(source, fact.name))
+      }));
+    }
+    seen.set(fact.name, fact);
    const arity = fact.params ? fact.params.length : 0;
    if (!fact.params || arity === 0) {
      warnings.push(createError({
@@ -285,6 +297,7 @@ function validateMeasures(program, tables, errors, warnings, source) {
 }

 function validateEvidence(program, tables, errors, warnings, source) {
+  const seen = new Map();
  for (const ev of program.evidence || []) {
    if (tables.builtins?.evidence?.has(ev.name)) {
      errors.push(createError({
@@ -295,6 +308,16 @@ function validateEvidence(program, tables, errors, warnings, source) {
        context: formatContext(source, findLocation(source, ev.name))
      }));
    }
+    if (seen.has(ev.name)) {
+      errors.push(createError({
+        message: `Duplicate evidence definition '${ev.name}'.`,
+        rule: 'Each evidence name must be unique within a program.',
+        fix: 'Rename one of the evidence definitions to a unique name.',
+        location: findLocation(source, `evidence ${ev.name}`),
+        context: formatContext(source, findLocation(source, ev.name))
+      }));
+    }
+    seen.set(ev.name, ev);
    const returnType = ev.provides || DEFAULT_EVIDENCE_RETURN;
    if (returnType !== DEFAULT_EVIDENCE_RETURN && !isTypeKnown(returnType, tables)) {
      errors.push(createError({
@@ -313,6 +336,37 @@ function validateEvidence(program, tables, errors, warnings, source) {
  }
 }

+/**
+ * Relation names must be unique across facts, sources, evidence, and measures.
+ * A fact and an evidence sharing a name would silently overwrite each other's
+ * relation config during generation (and read as a false cyclic reference).
+ */
+function validateCrossKindRelationNames(program, errors, warnings, source) {
+  const seen = new Map();
+  const kinds = [
+    ['fact', program.facts],
+    ['source', program.sources],
+    ['evidence', program.evidence],
+    ['measure', program.measures]
+  ];
+  for (const [kind, items] of kinds) {
+    for (const item of items || []) {
+      const prev = seen.get(item.name);
+      if (prev) {
+        errors.push(createError({
+          message: `Name '${item.name}' is already used by a ${prev} declaration.`,
+          rule: 'Relation names must be unique across facts, sources, evidence, and measures.',
+          fix: `Rename the ${kind} or the ${prev} to a unique name.`,
+          location: findLocation(source, `${kind} ${item.name}`),
+          context: formatContext(source, findLocation(source, item.name))
+        }));
+      } else {
+        seen.set(item.name, kind);
+      }
+    }
+  }
+}
+
 function validateEvidenceBody(body, scope, tables, errors, warnings, source, parent) {
  for (const stmt of body.statements || []) {
    switch (stmt.type) {
@@ -0,0 +1,129 @@
+/**
+ * tests/ChainConditionStep.test.js — a chain whose FINAL (object-side) hop
+ * references a defeasible/logical evidence. The compiler lowers it to a
+ * condition step: `{ rule: <config>, conditionStep: true }`, which the engine
+ * verifies at (intermediate, object) rather than traversing an edge.
+ *
+ * Only the final step may be a condition (the object is known); an
+ * intermediate condition cannot discover nodes and is a compile error.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLCompiler } from '../src/DSLCompiler.js';
+
+const DEFS = `
+  definition Employee { id: string }
+  definition Group { id: string }
+  definition Doc { id: string }
+  fact member_of(user: Employee, group: Group)
+  fact can_view(group: Group, doc: Doc)
+  fact banned(group: Group)
+  fact can_edit(group: Group, doc: Doc)
+`;
+
+function compile(dsl, name = 'chain-cond') {
+  const arb = new Arbiter();
+  const compiler = new DSLCompiler(arb);
+  const result = compiler.compile(dsl, name);
+  return { arb, result };
+}
+
+describe('Chain condition step (logical evidence as final hop)', () => {
+  it('lowers a defeasible final step to a condition step and grants', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence gated(group: Group, doc: Doc) { WHEN can_view(group, doc) UNLESS banned(group) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    const steps = arb.relationConfigs.get('can_via').steps;
+    assert.equal(steps[0], 'member_of');
+    assert.equal(steps[1].conditionStep, true);
+    assert.equal(steps[1].rule.type, 'logical');
+    // transitive dependency collection through the condition step
+    assert.deepEqual(arb.relationConfigs.get('can_via').dependsOn, ['member_of', 'can_view', 'banned']);
+
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'can_view', 'doc:9', { possibility: 0.7 });
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0.7);
+
+    // banning the intermediate defeats the condition hop
+    arb.addRelation('g:1', 'banned', 'g:1', { possibility: 1.0 });
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0);
+  });
+
+  it('supports ALWAYS/NEVER evidence as a condition step', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence gated(group: Group, doc: Doc) { ALWAYS can_edit(group, doc) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    const steps = arb.relationConfigs.get('can_via').steps;
+    assert.equal(steps[1].conditionStep, true);
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'can_edit', 'doc:9', { possibility: 0.6 });
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0.6);
+    arb.removeRelation('g:1', 'can_edit', 'doc:9');
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0);
+  });
+
+  it('keeps the condition evidence checkable in its own right', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence gated(group: Group, doc: Doc) { WHEN can_view(group, doc) UNLESS banned(group) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('g:1', 'can_view', 'doc:9', { possibility: 0.8 });
+    assert.equal(arb.check('g:1', 'gated', 'doc:9').possibility, 0.8);
+  });
+
+  it('parallel intermediates aggregate through the condition step', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence gated(group: Group, doc: Doc) { can_view(group, doc) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('g2:2', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 0.5 });
+    arb.addRelation('g:1', 'can_view', 'doc:9', { possibility: 0.7 });
+    arb.addRelation('u:1', 'member_of', 'g2:2', { possibility: 1.0 });
+    arb.addRelation('g2:2', 'can_view', 'doc:9', { possibility: 0.8 });
+    // max over paths: min(0.5,0.7)=0.5, min(1.0,0.8)=0.8 -> 0.8
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0.8);
+  });
+
+  it('expands an INTERMEDIATE condition step via rule-based reachability', () => {
+    const { arb, result } = compile(`
+      definition Employee { id: string }
+      definition Doc { id: string }
+      fact peer(user: Employee, other: Employee)
+      fact trusted(other: Employee)
+      fact can_read(user: Employee, doc: Doc)
+      evidence peer_trusted(user: Employee, other: Employee) { WHEN peer(user, other) UNLESS trusted(other) }
+      evidence can_access(user: Employee, doc: Doc) { peer_trusted(user, *p) { can_read(p, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    const steps = arb.relationConfigs.get('can_access').steps;
+    assert.equal(steps[0].conditionStep, true);
+    assert.equal(steps[0].rule.type, 'logical');
+
+    arb.addNode('u:1', 'Employee'); arb.addNode('p:1', 'Employee'); arb.addNode('p:2', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'peer', 'p:1', { possibility: 1.0 });
+    arb.addRelation('u:1', 'peer', 'p:2', { possibility: 1.0 });
+    arb.addRelation('p:1', 'trusted', 'p:1', { possibility: 1.0 }); // p:1 filtered
+    arb.addRelation('p:1', 'can_read', 'doc:9', { possibility: 0.9 });
+    arb.addRelation('p:2', 'can_read', 'doc:9', { possibility: 0.7 });
+    // only untrusted peer p:2 survives the intermediate condition -> 0.7
+    assert.equal(arb.check('u:1', 'can_access', 'doc:9').possibility, 0.7);
+    // trusting p:2 too removes all intermediates -> 0
+    arb.addRelation('p:2', 'trusted', 'p:2', { possibility: 1.0 });
+    assert.equal(arb.check('u:1', 'can_access', 'doc:9').possibility, 0);
+  });
+});
@@ -0,0 +1,127 @@
+/**
+ * tests/ChainStepComposition.test.js — evidence composition inside CHAIN
+ * steps. A chain step that references a derived evidence is expanded at
+ * compile time:
+ *   - a DIRECT evidence step → renamed to its underlying relation
+ *     (member_of(user,*g){ group_read(g,doc) } where group_read = can_view
+ *      becomes step 'can_view');
+ *   - a CHAIN evidence step → its steps are spliced into the parent chain
+ *     (a sub-path flattens into the linear source→…→object traversal);
+ *   - a DEFEASIBLE / LOGICAL / COMPARATOR evidence step is not an edge
+ *     traversal and is rejected at compile time;
+ *   - cycles and self-references through chain steps are compile errors.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLCompiler } from '../src/DSLCompiler.js';
+
+const DEFS = `
+  definition Employee { id: string }
+  definition Group { id: string }
+  definition Doc { id: string }
+  fact member_of(user: Employee, group: Group)
+  fact group_has(group: Group, sub: Group)
+  fact can_view(group: Group, doc: Doc)
+  fact can_access(group: Group, doc: Doc)
+  fact banned(group: Group)
+`;
+
+function compile(dsl, name = 'chain-compose') {
+  const arb = new Arbiter();
+  const compiler = new DSLCompiler(arb);
+  const result = compiler.compile(dsl, name);
+  return { arb, result };
+}
+
+describe('Chain step composition', () => {
+  it('renames a direct-evidence chain step to its underlying relation', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence group_read(group: Group, doc: Doc) { can_view(group, doc) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { group_read(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    // step 'group_read' → 'can_view'
+    assert.deepEqual(arb.relationConfigs.get('can_via').steps, ['member_of', 'can_view']);
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'can_view', 'doc:9', { possibility: 0.7 });
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0.7);
+  });
+
+  it('splices a chain-evidence step into the parent chain', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence group_enter(group: Group, doc: Doc) { group_has(group, *s) { can_access(s, doc) } }
+      evidence can_deep(user: Employee, doc: Doc) { member_of(user, *g) { group_enter(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    // step 'group_enter' → its steps [group_has, can_access]
+    assert.deepEqual(arb.relationConfigs.get('can_deep').steps, ['member_of', 'group_has', 'can_access']);
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('g2:2', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'group_has', 'g2:2', { possibility: 0.9 });
+    arb.addRelation('g2:2', 'can_access', 'doc:9', { possibility: 0.8 });
+    assert.equal(arb.check('u:1', 'can_deep', 'doc:9').possibility, 0.8);
+  });
+
+  it('expands a chain step whose direct evidence is itself composed', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence group_view(group: Group, doc: Doc) { can_view(group, doc) }
+      evidence group_read(group: Group, doc: Doc) { group_view(group, doc) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { group_read(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    assert.deepEqual(arb.relationConfigs.get('can_via').steps, ['member_of', 'can_view']);
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'can_view', 'doc:9', { possibility: 0.6 });
+    assert.equal(arb.check('u:1', 'can_via', 'doc:9').possibility, 0.6);
+  });
+
+  it('lowers a logical evidence FINAL step to a condition step', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence gated(group: Group, doc: Doc) { WHEN can_view(group, doc) UNLESS banned(group) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    // final-step logical evidence → condition step (verified at the object)
+    const steps = arb.relationConfigs.get('can_via').steps;
+    assert.equal(steps[0], 'member_of');
+    assert.equal(steps[1].conditionStep, true);
+    assert.equal(steps[1].rule.type, 'logical');
+  });
+
+  it('rejects a mutual cycle through chain steps', () => {
+    const { result } = compile(`
+      ${DEFS}
+      evidence cyc_a(group: Group, doc: Doc) { group_has(group, *g) { cyc_b(g, doc) } }
+      evidence cyc_b(group: Group, doc: Doc) { cyc_a(group, doc) }
+    `);
+    assert.equal(result.success, false);
+    assert.ok(result.errors.some(e => /[Cc]yclic/.test(e)), JSON.stringify(result.errors));
+  });
+
+  it('rejects a self-reference through its own chain step', () => {
+    const { result } = compile(`
+      ${DEFS}
+      evidence cyc_c(group: Group, doc: Doc) { group_has(group, *g) { cyc_c(g, doc) } }
+    `);
+    assert.equal(result.success, false);
+    assert.ok(result.errors.some(e => /[Cc]yclic/.test(e)), JSON.stringify(result.errors));
+  });
+
+  it('re-derives transitive dependencies through expanded chain steps', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence group_read(group: Group, doc: Doc) { can_view(group, doc) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { group_read(g, doc) } }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    // dependsOn reflects the expanded step, not the evidence reference
+    assert.deepEqual(arb.relationConfigs.get('can_via').dependsOn, ['member_of', 'can_view']);
+  });
+});
@@ -0,0 +1,208 @@
+/**
+ * tests/DSLRuntime.test.js — higher-order DSL+Core wrapper.
+ *
+ * Covers:
+ *   - schema indexing (types, relations, injectable facts, dependency graph)
+ *   - typed inserts/updates (addNode / updateNodeData / addRelation / updateRelation)
+ *     reject unknown types, wrong node types, and mistyped field values
+ *   - DSL-informed check: derives partial-graph requirements, retrieves missing
+ *     injectable facts through providers, injects them, and delegates
+ *   - missing-fact reporting
+ *
+ * NOTE: referencing a derived evidence relation as a sub-rule of another rule
+ * (e.g. `WHEN can_read(user, doc)` where can_read is an evidence) lowers to a
+ * direct edge lookup and does NOT re-derive the evidence's config. Evidence
+ * composition across rules is a documented gap (use fusion or facts).
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLRuntime } from '../src/runtime/DSLRuntime.js';
+
+const BASE_DSL = `
+  definition Employee { id: string? level: number? active: boolean? }
+  definition Group { id: string? }
+  definition Doc { id: string? }
+  fact member_of(user: Employee, group: Group)
+  fact *owns(user: Employee, doc: Doc)
+  fact *user_score(user: Employee, value: number)
+  fact *granted(user: Employee, doc: Doc)
+  fact can_access(group: Group, doc: Doc)
+  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+  evidence can_enter(user: Employee, doc: Doc) { member_of(user, *g) { can_access(g, doc) } }
+  evidence can_borrow(user: Employee, doc: Doc) { WHEN granted(user, doc) UNLESS user_score(user, 1) }
+`;
+
+function makeRuntime() {
+  return new DSLRuntime(new Arbiter()).compile(BASE_DSL, 'rt-test');
+}
+
+describe('DSLRuntime', () => {
+  it('indexes the DSL schema', () => {
+    const rt = makeRuntime();
+    assert.ok(rt.types.has('Employee'));
+    assert.equal(rt.types.get('Employee').fields.get('level').type, 'number');
+    assert.equal(rt.relations.get('owns').kind, 'fact');
+    assert.equal(rt.relations.get('owns').injectable, true);
+    assert.equal(rt.relations.get('member_of').injectable, false);
+    assert.equal(rt.relations.get('can_read').kind, 'evidence');
+    assert.deepEqual(rt.requiredFacts('can_read'), ['owns']);
+  });
+
+  it('validates typed node inserts', () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', { level: 3, active: true });
+    assert.throws(() => rt.addNode('g:1', 'Ghost', {}), /unknown type/);
+    assert.throws(() => rt.addNode('u:2', 'Employee', { level: 'high' }), /must be number/);
+    assert.throws(() => rt.addNode('u:3', 'Employee', { active: 'yes' }), /must be boolean/);
+  });
+
+  it('validates node updates against the declared type', () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', { level: 3, active: true });
+    rt.updateNodeData('u:1', { level: 5 });
+    assert.throws(() => rt.updateNodeData('u:1', { level: 'x' }), /must be number/);
+  });
+
+  it('validates relation endpoints against declared param types', () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('g:1', 'Group', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    assert.throws(() => rt.addRelation('u:1', 'member_of', 'doc:9', {}), /expected 'Group'/);
+    assert.throws(() => rt.addRelation('u:1', 'ghost_relation', 'g:1', {}), /unknown relation/);
+    // value-param fact: second param is a number value, dst must be the subject
+    rt.addRelation('u:1', 'user_score', 'u:1', { possibility: 1.0, value: 5 });
+    assert.throws(() => rt.addRelation('u:1', 'user_score', 'g:1', { possibility: 1.0, value: 5 }), /self-edge/);
+    assert.throws(() => rt.addRelation('u:1', 'user_score', 'u:1', { possibility: 1.0, value: 'high' }), /must be number/);
+  });
+
+  it('updateRelation validates and replaces', () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('g:1', 'Group', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.addRelation('u:1', 'member_of', 'g:1', { possibility: 0.5 });
+    rt.updateRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    assert.equal(rt.arbiter.check('u:1', 'member_of', 'g:1').possibility, 1.0);
+    assert.throws(() => rt.updateRelation('u:1', 'member_of', 'doc:9', {}), /expected 'Group'/);
+  });
+
+  it('DSL-informed check retrieves injectable facts via providers', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    const res = await rt.check('u:1', 'can_read', 'doc:9', {
+      factProviders: { owns: async () => 0.8 }
+    });
+    assert.equal(res.possibility, 0.8);
+    assert.equal(res.reason, 'allow_rule_matched');
+    assert.deepEqual(res.requiredFacts, ['owns']);
+    assert.deepEqual(res.providedFacts, ['owns']);
+    assert.deepEqual(res.missingFacts, []);
+  });
+
+  it('reports missing facts when a provider declines', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    const res = await rt.check('u:1', 'can_read', 'doc:9', {
+      factProviders: { owns: async () => null }
+    });
+    assert.equal(res.possibility, 0);
+    assert.deepEqual(res.missingFacts, [{ relation: 'owns', reason: 'not_provided' }]);
+  });
+
+  it('merges caller-supplied partial graphs with provider results', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    const res = await rt.check('u:1', 'can_read', 'doc:9', {
+      partialGraph: { relations: [{ src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: 1.0 }] },
+      factProviders: { owns: async () => null }
+    });
+    assert.equal(res.possibility, 1.0);
+  });
+
+  it('unary condition inside binary evidence (subject-as-object) defeats the grant', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    // can_borrow: WHEN granted(user, doc) UNLESS user_score(user, 1).
+    // user_score is injectable+unary; the provider injects a user self-edge
+    // with value 1 -> the unless fires and defeats the grant.
+    const res = await rt.check('u:1', 'can_borrow', 'doc:9', {
+      factProviders: {
+        granted: async () => 0.9,
+        user_score: async () => ({ possibility: 1.0, value: 1 })
+      }
+    });
+    assert.equal(res.possibility, 0);
+    assert.equal(res.reason, 'defeated_by_unless');
+    assert.deepEqual(res.requiredFacts, ['granted', 'user_score']);
+  });
+
+  it('chain evidence across an intermediate validates and checks', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('g:1', 'Group', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    rt.addRelation('g:1', 'can_access', 'doc:9', { possibility: 0.7 });
+    // can_enter: member_of(user, *g) { can_access(g, doc) } — chain [member_of, can_access]
+    const res = await rt.check('u:1', 'can_enter', 'doc:9', {});
+    assert.equal(res.possibility, 0.7);
+  });
+
+  it('rejects checks against unknown relations in strict mode', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    await assert.rejects(() => rt.check('u:1', 'does_not_exist', 'doc:9'), /unknown relation/);
+  });
+
+  it('derives transitive required facts through evidence composition', async () => {
+    const dsl = `
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *owns(user: Employee, doc: Doc)
+      fact *banned(user: Employee)
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_open(user: Employee, doc: Doc) { WHEN can_read(user, doc) UNLESS banned(user) }
+    `;
+    const rt = new DSLRuntime(new Arbiter()).compile(dsl, 'rt-comp');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    // can_open composes can_read, so its requirements reach through to owns.
+    assert.deepEqual(rt.requiredFacts('can_open'), ['owns', 'banned']);
+    const granted = await rt.check('u:1', 'can_open', 'doc:9', {
+      factProviders: { owns: async () => 0.9, banned: async () => 0 }
+    });
+    assert.equal(granted.possibility, 0.9);
+    assert.equal(granted.reason, 'allow_rule_matched');
+    assert.deepEqual(granted.providedFacts, ['owns', 'banned']);
+    const denied = await rt.check('u:1', 'can_open', 'doc:9', {
+      factProviders: { owns: async () => 0.9, banned: async () => 1 }
+    });
+    assert.equal(denied.possibility, 0);
+    assert.equal(denied.reason, 'defeated_by_unless');
+  });
+
+  it('derives transitive required facts through a condition-step chain', () => {
+    const dsl = `
+      definition Employee { id: string? }
+      definition Group { id: string? }
+      definition Doc { id: string? }
+      fact *member_of(user: Employee, group: Group)
+      fact *can_view(group: Group, doc: Doc)
+      fact *banned(group: Group)
+      evidence gated(group: Group, doc: Doc) { WHEN can_view(group, doc) UNLESS banned(group) }
+      evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }
+    `;
+    const rt = new DSLRuntime(new Arbiter()).compile(dsl, 'rt-cond');
+    // The condition step's facts (can_view, banned) reach through to the
+    // evidence's requirements, alongside the edge-traversal fact.
+    assert.deepEqual(rt.requiredFacts('can_via'), ['member_of', 'can_view', 'banned']);
+  });
+});
@@ -0,0 +1,203 @@
+/**
+ * tests/DSLRuntimeCache.test.js — provider-result caching with time expiry.
+ *
+ * Registered providers retrieve missing facts from a data store; caching the
+ * retrieval avoids hammering the store on repeated checks. TTL resolution:
+ * DSL-declared `BEHAVES { ttl <duration> }` on a fact > per-fact setFactTTL >
+ * policy default (30s). Per-check factProviders are cache-transparent (one-off
+ * observations: no cache read, no cache write). Registering a provider or
+ * mutating the graph invalidates the cache.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLRuntime } from '../src/runtime/DSLRuntime.js';
+
+const BASE_DSL = `
+  definition Employee { id: string? }
+  definition Doc { id: string? }
+  fact *owns(user: Employee, doc: Doc)
+  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+`;
+
+function makeRuntime(options = {}) {
+  let t = 0;
+  const clock = () => t;
+  const rt = new DSLRuntime(new Arbiter(), { clock, ...options }).compile(BASE_DSL, 'rt-cache');
+  rt._test_advance = (ms) => { t += ms; };
+  return rt;
+}
+
+describe('DSLRuntime provider-result caching', () => {
+  it('reuses a registered provider result within the TTL', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    await rt.check('u:1', 'can_read', 'doc:9');
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 1, 'provider should be invoked once within TTL');
+  });
+
+  it('re-invokes the provider after the TTL expires', async () => {
+    const rt = makeRuntime();
+    rt.setFactTTL('owns', 100);
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    let value = 0.9;
+    rt.registerFact('owns', async () => { calls++; return value; });
+    const first = await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(first.possibility, 0.9);
+    rt._test_advance(50);
+    await rt.check('u:1', 'can_read', 'doc:9'); // within TTL -> cached
+    assert.equal(calls, 1);
+    rt._test_advance(60); // past TTL (110 total)
+    value = 0.4;
+    const after = await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 2);
+    assert.equal(after.possibility, 0.4);
+  });
+
+  it('per-check factProviders override the cache (fresh observation)', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('owns', async () => 0.9);
+    await rt.check('u:1', 'can_read', 'doc:9');
+    // Per-check override is cache-transparent: it must NOT be masked by the
+    // cached 0.9, and it must NOT overwrite the cached value.
+    const over = await rt.check('u:1', 'can_read', 'doc:9', {
+      factProviders: { owns: async () => 0.2 }
+    });
+    assert.equal(over.possibility, 0.2);
+    const next = await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(next.possibility, 0.9, 'registered provider cache untouched by per-check override');
+  });
+
+  it('registerFact invalidates the cached result for that relation', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('owns', async () => 0.9);
+    await rt.check('u:1', 'can_read', 'doc:9');
+    rt.registerFact('owns', async () => 0.3); // re-register -> cache invalidated
+    const res = await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(res.possibility, 0.3);
+  });
+
+  it('invalidates cached results on graph mutations', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 1);
+    rt.addRelation('u:1', 'owns', 'doc:9', { possibility: 1.0 }); // mutation clears cache
+    const res = await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 2, 'graph mutation should invalidate the provider cache');
+  });
+
+  it('invalidateProviderCache() clears all or per relation', async () => {
+    const dsl = `
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *owns(user: Employee, doc: Doc)
+      fact *banned(user: Employee)
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_open(user: Employee, doc: Doc) { WHEN can_read(user, doc) UNLESS banned(user) }
+    `;
+    const rt = new DSLRuntime(new Arbiter()).compile(dsl, 'rt-cache2');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let ownsCalls = 0, bannedCalls = 0;
+    rt.registerFact('owns', async () => { ownsCalls++; return 0.9; });
+    rt.registerFact('banned', async () => { bannedCalls++; return 0; });
+    await rt.check('u:1', 'can_open', 'doc:9');
+    assert.equal(ownsCalls, 1);
+    assert.equal(bannedCalls, 1);
+    // Invalidate a non-dependency relation: can_open's cache (owns+banned) survives.
+    rt.invalidateProviderCache('does_not_exist');
+    await rt.check('u:1', 'can_open', 'doc:9');
+    assert.equal(ownsCalls, 1);
+    assert.equal(bannedCalls, 1);
+    // Invalidate owns only: banned survives, owns re-fetched.
+    rt.invalidateProviderCache('owns');
+    await rt.check('u:1', 'can_open', 'doc:9');
+    assert.equal(ownsCalls, 2, 'owns cache cleared by per-relation invalidation');
+    assert.equal(bannedCalls, 1, 'banned cache survives per-relation invalidation');
+    // Clear all.
+    rt.invalidateProviderCache();
+    await rt.check('u:1', 'can_open', 'doc:9');
+    assert.equal(bannedCalls, 2, 'full invalidation clears every relation');
+  });
+
+  it('policy default TTL applies when no per-fact TTL is set', async () => {
+    const rt = makeRuntime({ policy: { providerCacheTTL: 50 } });
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    await rt.check('u:1', 'can_read', 'doc:9');
+    rt._test_advance(40);
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 1, 'within 50ms policy TTL -> cached');
+    rt._test_advance(20);
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 2, 'past 50ms policy TTL -> re-invoked');
+  });
+
+  it('uses the DSL-declared fact TTL (BEHAVES { ttl X })', async () => {
+    const dsl = `
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *balance(user: Employee, amount: number) BEHAVES { ttl 1h }
+      evidence can_spend(user: Employee, doc: Doc) { balance(user, 1) }
+    `;
+    let t = 0;
+    const rt = new DSLRuntime(new Arbiter(), { clock: () => t }).compile(dsl, 'rt-dsl-ttl');
+    // The DSL declares a 1h TTL for the balance fact.
+    assert.equal(rt.relations.get('balance').ttlMs, 3600_000);
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('balance', async () => { calls++; return { possibility: 1.0, value: 50 }; });
+    await rt.check('u:1', 'can_spend', 'doc:9');
+    assert.equal(calls, 1);
+    t += 60 * 60 * 1000 - 1; // just under 1h
+    await rt.check('u:1', 'can_spend', 'doc:9');
+    assert.equal(calls, 1, 'cached within DSL-declared 1h TTL');
+    t += 2;
+    await rt.check('u:1', 'can_spend', 'doc:9');
+    assert.equal(calls, 2, 're-invoked past the DSL-declared 1h TTL');
+  });
+
+  it('cacheProviderResults:false bypasses the cache per check', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 1);
+    // Bypass forces a fresh retrieval without clearing the cache.
+    await rt.check('u:1', 'can_read', 'doc:9', { cacheProviderResults: false });
+    assert.equal(calls, 2);
+    // Cache still intact for the next default check.
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 2);
+  });
+
+  it('policy.cacheProviderResults:false disables caching globally', async () => {
+    const rt = makeRuntime({ policy: { cacheProviderResults: false } });
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    await rt.check('u:1', 'can_read', 'doc:9');
+    await rt.check('u:1', 'can_read', 'doc:9');
+    assert.equal(calls, 2, 'no caching when disabled globally');
+  });
+});
@@ -0,0 +1,168 @@
+/**
+ * tests/DSLRuntimeExt.test.js — extended DSLRuntime capabilities:
+ *   - schema introspection (getSchema)
+ *   - per-relation provider registration (registerFact/unregisterFact)
+ *   - provider merging (registered + per-check overrides)
+ *   - bounded fixed-point provider retrieval loop (edges satisfy other facts)
+ *   - require() throw-on-deny
+ *   - removal passthroughs and fact-relation check validation
+ *   - timestamp/duration field typing
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLRuntime } from '../src/runtime/DSLRuntime.js';
+
+const BASE_DSL = `
+  definition Employee { id: string? level: number? active: boolean? }
+  definition Doc { id: string? created: timestamp? }
+  fact *owns(user: Employee, doc: Doc)
+  fact *banned(user: Employee)
+  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+  evidence can_open(user: Employee, doc: Doc) { WHEN can_read(user, doc) UNLESS banned(user) }
+`;
+
+function makeRuntime() {
+  return new DSLRuntime(new Arbiter()).compile(BASE_DSL, 'rt-ext');
+}
+
+describe('DSLRuntime extended', () => {
+  it('exposes a serializable schema snapshot', () => {
+    const rt = makeRuntime();
+    const schema = rt.getSchema();
+    assert.ok(Array.isArray(schema.types));
+    const employee = schema.types.find(t => t.name === 'Employee');
+    assert.ok(employee);
+    assert.ok(employee.fields.some(f => f.name === 'level' && f.type === 'number'));
+    const owns = schema.facts.find(f => f.name === 'owns');
+    assert.equal(owns.injectable, true);
+    assert.equal(owns.params[1].type, 'Doc');
+    const can_open = schema.evidence.find(e => e.name === 'can_open');
+    assert.ok(can_open.dependsOn.includes('owns'));
+    assert.deepEqual(schema.providers, []);
+    assert.ok(rt.relationNames().includes('owns') && rt.relationNames().includes('can_read'));
+  });
+
+  it('registers, lists, and unregisters per-relation providers', () => {
+    const rt = makeRuntime();
+    rt.registerFact('owns', async () => 0.8);
+    assert.deepEqual(rt.registeredFacts(), ['owns']);
+    rt.registerFact('banned', async () => 0);
+    assert.deepEqual(rt.registeredFacts().sort(), ['banned', 'owns']);
+    rt.unregisterFact('banned');
+    assert.deepEqual(rt.registeredFacts(), ['owns']);
+    assert.throws(() => rt.registerFact('owns', 'not a function'), /must be a function/);
+  });
+
+  it('merges registered providers with per-check overrides', async () => {
+    const rt = makeRuntime();
+    rt.registerFact('owns', async () => 0.5);
+    rt.registerFact('banned', async () => 0);
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    // registered owns (0.5) wins over nothing; per-check banned overrides
+    const res = await rt.check('u:1', 'can_open', 'doc:9', {
+      factProviders: { banned: async () => 0 }
+    });
+    assert.equal(res.possibility, 0.5);
+    assert.deepEqual(res.providedFacts.sort(), ['banned', 'owns']);
+  });
+
+  it('runs providers to a fixed point when edges satisfy other required facts', async () => {
+    // can_open needs owns (injectable). A registered owns provider returns an
+    // edge for a DIFFERENT injectable fact that can_open also requires via
+    // composition — here we add a transitive requirement to prove the loop.
+    const dsl = `
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *owns(user: Employee, doc: Doc)
+      fact *granted(user: Employee, doc: Doc)
+      evidence base_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_open(user: Employee, doc: Doc) { WHEN base_read(user, doc) UNLESS granted(user, doc) }
+    `;
+    const rt = new DSLRuntime(new Arbiter()).compile(dsl, 'rt-loop');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let ownsCalls = 0;
+    let grantedCalls = 0;
+    rt.registerFact('owns', async () => {
+      ownsCalls++;
+      // First round the owns provider also supplies the granted edge (a
+      // fixed-point dependency: granted needs owns to have been retrieved).
+      return [
+        { src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: 0.9 },
+        { src: 'u:1', relation: 'granted', dst: 'doc:9', possibility: 0 }
+      ];
+    });
+    rt.registerFact('granted', async () => { grantedCalls++; return 0; });
+    const res = await rt.check('u:1', 'can_open', 'doc:9', { maxProviderRounds: 3 });
+    assert.equal(res.possibility, 0.9);
+    // granted was satisfied by the owns provider's extra edge, so its own
+    // provider was never needed in a later round.
+    assert.equal(grantedCalls, 0);
+    assert.ok(ownsCalls >= 1);
+    assert.deepEqual(res.providedFacts, ['owns']);
+    assert.deepEqual(res.missingFacts, []);
+  });
+
+  it('require() throws on denial and returns the result on grant', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('owns', async () => 0.9);
+    const ok = await rt.require('u:1', 'can_read', 'doc:9');
+    assert.equal(ok.possibility, 0.9);
+    rt.registerFact('owns', async () => 0);
+    await assert.rejects(
+      () => rt.require('u:1', 'can_read', 'doc:9'),
+      (err) => err.result && err.result.possibility === 0 && /denied/.test(err.message)
+    );
+  });
+
+  it('passes through node/relation removal', () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.addRelation('u:1', 'owns', 'doc:9', { possibility: 1.0 });
+    rt.removeRelation('u:1', 'owns', 'doc:9');
+    assert.equal(rt.arbiter.check('u:1', 'owns', 'doc:9').possibility, 0);
+    rt.removeNode('u:1');
+    assert.equal(rt.arbiter.nodeIdByKey.has('u:1'), false);
+  });
+
+  it('validates fact-relation check endpoints like evidence', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    // can_read is evidence; owns is a fact — checking a fact still validates.
+    await assert.rejects(() => rt.check('u:1', 'owns', 'u:1', {}), /expected 'Doc'/);
+  });
+
+  it('accepts timestamp field values and rejects mistyped ones', () => {
+    const rt = makeRuntime();
+    rt.addNode('doc:9', 'Doc', { created: 1720000000000 });
+    rt.updateNodeData('doc:9', { created: '2026-08-03T00:00:00Z' });
+    assert.throws(() => rt.addNode('doc:8', 'Doc', { created: {} }), /must be timestamp/);
+  });
+
+  it('direct FACT checks consult the registered provider', async () => {
+    const rt = makeRuntime();
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    let calls = 0;
+    rt.registerFact('owns', async () => { calls++; return 0.9; });
+    // Checking the fact directly (not via an evidence) must retrieve it.
+    const res = await rt.check('u:1', 'owns', 'doc:9');
+    assert.equal(res.possibility, 0.9);
+    assert.equal(calls, 1);
+    assert.deepEqual(res.requiredFacts, ['owns']);
+    assert.deepEqual(res.providedFacts, ['owns']);
+    // Without a provider and without an edge, it reports the missing fact.
+    const rt2 = new DSLRuntime(new Arbiter()).compile(BASE_DSL, 'rt-fact-miss');
+    rt2.addNode('u:1', 'Employee', {});
+    rt2.addNode('doc:9', 'Doc', {});
+    const missed = await rt2.check('u:1', 'owns', 'doc:9');
+    assert.equal(missed.possibility, 0);
+    assert.deepEqual(missed.missingFacts, [{ relation: 'owns', reason: 'no_provider' }]);
+  });
+});
@@ -0,0 +1,96 @@
+/**
+ * tests/DSLRuntimeTyping.test.js — duration seconds, required fields, and
+ * type validation of insertions / updates / provider retrievals.
+ *
+ * - Duration literals now accept s/m/h/d/w: `BEHAVES { ttl 30s }` is 30s.
+ * - Definition fields are REQUIRED by default (`field: type`); `field: type?`
+ *   marks a field optional. addNode enforces presence on insert.
+ * - Provider-returned edges are validated against the fact's declared typing:
+ *   a value-carrying fact must return { value, possibility } with a value of
+ *   the declared type, and possibilities must lie in [0, 1].
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLRuntime } from '../src/runtime/DSLRuntime.js';
+
+describe('DSLRuntime typing', () => {
+  it('accepts seconds/minutes/hours/days/weeks in duration literals', async () => {
+    const dsl = `
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *a(user: Employee, amount: number) BEHAVES { ttl 30s }
+      fact *b(user: Employee, amount: number) BEHAVES { ttl 2m }
+      fact *c(user: Employee, amount: number) BEHAVES { ttl 1h }
+      fact *d(user: Employee, amount: number) BEHAVES { ttl 3d }
+      fact *e(user: Employee, amount: number) BEHAVES { ttl 1w }
+    `;
+    const rt = new DSLRuntime(new Arbiter()).compile(dsl, 'rt-units');
+    assert.equal(rt.relations.get('a').ttlMs, 30_000);
+    assert.equal(rt.relations.get('b').ttlMs, 120_000);
+    assert.equal(rt.relations.get('c').ttlMs, 3_600_000);
+    assert.equal(rt.relations.get('d').ttlMs, 259_200_000);
+    assert.equal(rt.relations.get('e').ttlMs, 604_800_000);
+  });
+
+  it('enforces required definition fields on node insert', () => {
+    const rt = new DSLRuntime(new Arbiter()).compile(`
+      definition Employee { id: string level: number active: boolean? }
+    `, 'rt-req');
+    // id and level are required (no `?`); active is optional.
+    assert.throws(() => rt.addNode('u:1', 'Employee', { level: 3 }), /missing required field 'Employee.id'/);
+    assert.throws(() => rt.addNode('u:2', 'Employee', { id: 'u:2' }), /missing required field 'Employee.level'/);
+    rt.addNode('u:3', 'Employee', { id: 'u:3', level: 5 }); // both required, no active -> ok
+    rt.addNode('u:4', 'Employee', { id: 'u:4', level: 5, active: true });
+  });
+
+  it('exposes requiredness in the schema snapshot', () => {
+    const rt = new DSLRuntime(new Arbiter()).compile(`
+      definition Employee { id: string level: number? }
+    `, 'rt-schema-req');
+    const employee = rt.getSchema().types.find(t => t.name === 'Employee');
+    assert.equal(employee.fields.find(f => f.name === 'id').required, true);
+    assert.equal(employee.fields.find(f => f.name === 'level').required, false);
+  });
+
+  it('validates a provider-returned value against the declared value type', async () => {
+    const rt = new DSLRuntime(new Arbiter()).compile(`
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *balance(user: Employee, amount: number)
+      evidence can_spend(user: Employee, doc: Doc) { balance(user, 1) }
+    `, 'rt-valuetype');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('balance', async () => ({ possibility: 1.0, value: 'high' }));
+    await assert.rejects(() => rt.check('u:1', 'can_spend', 'doc:9'), /must be number/);
+  });
+
+  it('requires a value for a value-carrying fact (no bare-number shorthand)', async () => {
+    const rt = new DSLRuntime(new Arbiter()).compile(`
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *balance(user: Employee, amount: number)
+      evidence can_spend(user: Employee, doc: Doc) { balance(user, 1) }
+    `, 'rt-valshape');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('balance', async () => 0.9);
+    await assert.rejects(() => rt.check('u:1', 'can_spend', 'doc:9'), /value-carrying fact/);
+    rt.registerFact('balance', async () => ({ possibility: 1.0 })); // missing value
+    await assert.rejects(() => rt.check('u:1', 'can_spend', 'doc:9'), /must supply a 'value'/);
+  });
+
+  it('rejects a provider-returned possibility outside [0, 1]', async () => {
+    const rt = new DSLRuntime(new Arbiter()).compile(`
+      definition Employee { id: string? }
+      definition Doc { id: string? }
+      fact *owns(user: Employee, doc: Doc)
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+    `, 'rt-poss');
+    rt.addNode('u:1', 'Employee', {});
+    rt.addNode('doc:9', 'Doc', {});
+    rt.registerFact('owns', async () => ({ possibility: 2.0 }));
+    await assert.rejects(() => rt.check('u:1', 'can_read', 'doc:9'), /invalid possibility/);
+  });
+});
@@ -19,11 +19,11 @@ describe('Type Definitions', () => {
  test('Basic definitions', () => {
    const testCases = [
      {
-        input: `definition User { role: string }`,
+        input: `definition Employee { role: string }`,
        description: 'Simple definition with one field'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          role: string
          isActive: boolean
        }`,
@@ -57,7 +57,8 @@ describe('Type Definitions', () => {
    ];

    testCases.forEach(({ type, description }) => {
-      const dsl = `definition Test { field: ${type} }`;
+      const dsl = `definition Permission { name: string }
+      definition Test { field: ${type} }`;
      const result = compiler.compile(dsl, `test-field-type-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -73,7 +74,8 @@ describe('Type Definitions', () => {
    ];

    testCases.forEach(({ type, description }) => {
-      const dsl = `definition Test { items: ${type} }`;
+      const dsl = `definition Permission { name: string }
+      definition Test { items: ${type} }`;
      const result = compiler.compile(dsl, `test-array-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -82,61 +84,61 @@ describe('Type Definitions', () => {
  test('Behaviors', () => {
    const testCases = [
      {
-        input: `definition User {
+        input: `definition Employee {
          balance: number BEHAVES { decaying down hourly }
        }`,
        description: 'Decay behavior - down hourly'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          reputation: number BEHAVES { decaying up daily }
        }`,
        description: 'Decay behavior - up daily'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          score: number BEHAVES { decaying neutral weekly }
        }`,
        description: 'Decay behavior - neutral weekly'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          stability: number BEHAVES { decaying stable monthly }
        }`,
        description: 'Decay behavior - stable monthly'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          confidence: number BEHAVES { blurring fixed }
        }`,
        description: 'Blur behavior - fixed'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          accuracy: number BEHAVES { blurring adaptive }
        }`,
        description: 'Blur behavior - adaptive'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          precision: number BEHAVES { blurring confidence confidence_90 }
        }`,
        description: 'Blur behavior - confidence with level'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          session: string BEHAVES { ttl 1h }
        }`,
        description: 'TTL behavior - hours'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          token: string BEHAVES { ttl 24h }
        }`,
        description: 'TTL behavior - 24 hours'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          cache: string BEHAVES { ttl 7d }
        }`,
        description: 'TTL behavior - days'
@@ -152,25 +154,25 @@ describe('Type Definitions', () => {
  test('Caching', () => {
    const testCases = [
      {
-        input: `definition User {
+        input: `definition Employee {
          role: string CACHE eager
        }`,
        description: 'Eager caching'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          score: number CACHE lazy
        }`,
        description: 'Lazy caching'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          balance: number BEHAVES { decaying down hourly } CACHE eager
        }`,
        description: 'Behavior with eager caching'
      },
      {
-        input: `definition User {
+        input: `definition Employee {
          reputation: number BEHAVES { blurring adaptive } CACHE lazy
        }`,
        description: 'Behavior with lazy caching'
@@ -186,7 +188,7 @@ describe('Type Definitions', () => {
  test('Complex definitions', () => {
    const testCases = [
      {
-        input: `definition User {
+        input: `definition Employee {
          role: string
          isActive: boolean
          lastActive: timestamp BEHAVES {
@@ -206,10 +208,12 @@ describe('Type Definitions', () => {
        description: 'Complex definition with multiple behaviors and caching'
      },
      {
-        input: `definition Group {
+        input: `definition Employee { role: string }
+        definition Permission { name: string }
+        definition Group {
          name: string
          permissions: Permission[]
-          members: User[]
+          members: Employee[]
          created: timestamp BEHAVES {
            decaying stable monthly
          } CACHE lazy
@@ -218,9 +222,10 @@ describe('Type Definitions', () => {
        description: 'Definition with arrays and mixed behaviors'
      },
      {
-        input: `definition Document {
+        input: `definition Employee { role: string }
+        definition Document {
          level: string
-          owner: User
+          owner: Employee
          tags: string[]
          content: string BEHAVES {
            blurring fixed
@@ -246,27 +251,27 @@ describe('Type Definitions', () => {
  test('Definition error handling', () => {
    const testCases = [
      {
-        input: `definition User { role: string`,
+        input: `definition Employee { role: string`,
        description: 'Missing closing brace should fail'
      },
      {
-        input: `definition User { role: }`,
+        input: `definition Employee { role: }`,
        description: 'Missing field type should fail'
      },
      {
-        input: `definition User { : string }`,
+        input: `definition Employee { : string }`,
        description: 'Missing field name should fail'
      },
      {
-        input: `definition User { role: string BEHAVES { }`,
+        input: `definition Employee { role: string BEHAVES { }`,
        description: 'Incomplete behavior should fail'
      },
      {
-        input: `definition User { role: string CACHE }`,
+        input: `definition Employee { role: string CACHE }`,
        description: 'Incomplete cache directive should fail'
      },
      {
-        input: `definition User { role: string BEHAVES { invalid } }`,
+        input: `definition Employee { role: string BEHAVES { invalid } }`,
        description: 'Invalid behavior should fail'
      }
    ];
@@ -0,0 +1,154 @@
+/**
+ * tests/EvidenceComposition.test.js — referencing a derived evidence as a
+ * sub-rule of another evidence (WHEN can_read(user, doc) where can_read is
+ * itself an evidence).
+ *
+ * Composition is resolved at COMPILE time: the generator inlines each
+ * evidence reference with the referenced evidence's own config (a linker
+ * pass that handles forward references and rejects cycles), so the engine
+ * evaluates a fully-resolved, acyclic config tree.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { Arbiter } from '@arbiter/core';
+import { DSLCompiler } from '../src/DSLCompiler.js';
+
+const DEFS = `
+  definition Employee { id: string }
+  definition Group { id: string }
+  definition Doc { id: string }
+  fact owns(user: Employee, doc: Doc)
+  fact *trusted(user: Employee)
+  fact member_of(user: Employee, group: Group)
+  fact can_access(group: Group, doc: Doc)
+`;
+
+function compile(dsl, name = 'compose') {
+  const arb = new Arbiter();
+  const compiler = new DSLCompiler(arb);
+  const result = compiler.compile(dsl, name);
+  return { arb, result };
+}
+
+describe('Evidence composition', () => {
+  it('composes a direct evidence into another evidence', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_browse(user: Employee, doc: Doc) { can_read(user, doc) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'owns', 'doc:9', { possibility: 0.8 });
+    const res = arb.check('u:1', 'can_browse', 'doc:9');
+    assert.equal(res.possibility, 0.8);
+    // The reference is inlined to the underlying fact config.
+    assert.equal(arb.relationConfigs.get('can_browse').type, 'direct');
+    assert.equal(arb.relationConfigs.get('can_browse').relation, 'owns');
+  });
+
+  it('composes an evidence inside a defeasible WHEN/UNLESS', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_open(user: Employee, doc: Doc) { WHEN can_read(user, doc) UNLESS trusted(user) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'owns', 'doc:9', { possibility: 0.9 });
+    assert.equal(arb.check('u:1', 'can_open', 'doc:9').possibility, 0.9);
+    arb.addRelation('u:1', 'trusted', 'u:1', { possibility: 1.0 });
+    const denied = arb.check('u:1', 'can_open', 'doc:9');
+    assert.equal(denied.possibility, 0);
+    assert.equal(denied.reason, 'defeated_by_unless');
+  });
+
+  it('composes a chain evidence into another evidence', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence can_enter(user: Employee, doc: Doc) { member_of(user, *g) { can_access(g, doc) } }
+      evidence can_work(user: Employee, doc: Doc) { can_enter(user, doc) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('g:1', 'Group'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'member_of', 'g:1', { possibility: 1.0 });
+    arb.addRelation('g:1', 'can_access', 'doc:9', { possibility: 0.7 });
+    const res = arb.check('u:1', 'can_work', 'doc:9');
+    assert.equal(res.possibility, 0.7);
+    assert.equal(arb.relationConfigs.get('can_work').type, 'chain');
+  });
+
+  it('composes transitively (A → B → fact) and re-derives dependencies', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_browse(user: Employee, doc: Doc) { can_read(user, doc) }
+      evidence can_open(user: Employee, doc: Doc) { can_browse(user, doc) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'owns', 'doc:9', { possibility: 0.6 });
+    assert.equal(arb.check('u:1', 'can_open', 'doc:9').possibility, 0.6);
+    assert.deepEqual(arb.relationConfigs.get('can_open').dependsOn, ['owns']);
+  });
+
+  it('composes a value-carrying evidence and preserves subject-as-object scope', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      fact *user_risk(user: Employee, value: number)
+      evidence risk_ok(user: Employee, doc: Doc) { user_risk(user, 1) }
+      evidence can_proceed(user: Employee, doc: Doc) { risk_ok(user, doc) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'user_risk', 'u:1', { possibility: 1.0, value: 1 });
+    const res = arb.check('u:1', 'can_proceed', 'doc:9');
+    assert.equal(res.possibility, 1);
+  });
+
+  it('composes evidence inside a comparator operand', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      fact *user_risk(user: Employee, value: number)
+      fact *risk_limit(doc: Doc, value: number)
+      evidence user_risk_ok(user: Employee, doc: Doc) { user_risk(user, 1) }
+      evidence can_proceed(user: Employee, doc: Doc) { user_risk_ok(user, doc) }
+    `);
+    assert.ok(result.success, JSON.stringify(result.errors));
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'user_risk', 'u:1', { possibility: 1.0, value: 1 });
+    assert.equal(arb.check('u:1', 'can_proceed', 'doc:9').possibility, 1);
+  });
+
+  it('rejects cyclic evidence references at compile time', () => {
+    const { result } = compile(`
+      ${DEFS}
+      evidence a(user: Employee, doc: Doc) { b(user, doc) }
+      evidence b(user: Employee, doc: Doc) { a(user, doc) }
+    `);
+    assert.equal(result.success, false);
+    assert.ok(result.errors.some(e => /[Cc]yclic/.test(e)), JSON.stringify(result.errors));
+  });
+
+  it('rejects self-referencing evidence at compile time', () => {
+    const { result } = compile(`
+      ${DEFS}
+      evidence a(user: Employee, doc: Doc) { a(user, doc) }
+    `);
+    assert.equal(result.success, false);
+    assert.ok(result.errors.some(e => /[Cc]yclic/.test(e)), JSON.stringify(result.errors));
+  });
+
+  it('keeps the referenced evidence checkable in its own right', () => {
+    const { arb, result } = compile(`
+      ${DEFS}
+      evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+      evidence can_browse(user: Employee, doc: Doc) { can_read(user, doc) }
+    `);
+    assert.ok(result.success);
+    arb.addNode('u:1', 'Employee'); arb.addNode('doc:9', 'Doc');
+    arb.addRelation('u:1', 'owns', 'doc:9', { possibility: 0.5 });
+    assert.equal(arb.check('u:1', 'can_read', 'doc:9').possibility, 0.5);
+    assert.equal(arb.check('u:1', 'can_browse', 'doc:9').possibility, 0.5);
+  });
+});
@@ -12,6 +12,49 @@ function createMockArbiter() {
  };
 }

+const DSL_SUPPORT = `
+  definition Employee {
+    role: string
+    isActive: boolean
+    isTrusted: boolean
+    hasRecentActivity: boolean
+    lastActive: timestamp
+    isBlacklisted: boolean
+    session: string
+  }
+
+  definition Document {
+    level: string
+    isPublic: boolean
+    isEditable: boolean
+  }
+
+  definition Resource {
+    level: string
+    isPublic: boolean
+  }
+
+  fact hasRole(user: any, role: string)
+  fact hasClearance(user: any, level: string)
+  fact owns(user: any, doc: any)
+  fact isSuspended(user: any)
+  fact isActive(user: any)
+  fact isTrusted(user: any)
+  fact hasRecentActivity(user: any)
+  fact isBlacklisted(user: any)
+  fact isMember(user: any, group: any)
+  fact isFriend(user: any, friend: any)
+  fact similar(a: any, b: any)
+  fact reachable(user: any, doc: any)
+  fact parentOf(user: any, parent: any)
+  fact isEditable(doc: any)
+  fact isPublic(doc: any)
+  fact recentlyActive(user: any)
+  fact reputationScore(user: any)
+  fact activityScore(user: any)
+  fact verificationLevel(user: any)
+`;
+
 describe('Evidence Rules', () => {
  const arbiter = createMockArbiter();
  const compiler = new DSLCompiler(arbiter);
@@ -19,35 +62,35 @@ describe('Evidence Rules', () => {
  test('Basic evidence', () => {
    const testCases = [
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          hasRole(user, 'admin')
        }`,
        description: 'Simple evidence with function call'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
-          user.isActive
+        input: `evidence canAccess(user: Employee, resource: Resource) {
+          isActive(user)
        }`,
        description: 'Evidence with attribute access'
      },
      {
-        input: `evidence canModify(user: User, doc: Document) {
-          user.isActive
+        input: `evidence canModify(user: Employee, doc: Document) {
+          isActive(user)
          hasRole(user, 'admin')
        }`,
        description: 'Evidence with multiple conditions'
      },
      {
-        input: `evidence canDelete(user: User, doc: Document) {
+        input: `evidence canDelete(user: Employee, doc: Document) {
          owns(user, doc)
-          user.isActive
+          isActive(user)
        }`,
        description: 'Evidence with ownership and status'
      }
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-basic-evidence-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-basic-evidence-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
      assert.ok(result.program.evidence.length > 0, 'Should have evidence');
    });
@@ -56,32 +99,32 @@ describe('Evidence Rules', () => {
  test('Defeasible logic', () => {
    const testCases = [
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
-          ALWAYS user.isActive
+        input: `evidence canAccess(user: Employee, resource: Resource) {
+          ALWAYS isActive(user)
        }`,
        description: 'ALWAYS rule - strict requirement'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          WHEN hasRole(user, 'admin')
        }`,
        description: 'WHEN rule - defeasible condition'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          WHEN hasRole(user, 'admin') UNLESS isSuspended(user)
        }`,
        description: 'WHEN/UNLESS rule - defeasible with defeater'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          REQUIRES hasClearance(user, resource.level)
        }`,
        description: 'REQUIRES rule - inverse defeater'
      },
      {
-        input: `evidence canAccessCritical(user: User, resource: Resource) {
-          ALWAYS user.isActive
+        input: `evidence canAccessCritical(user: Employee, resource: Resource) {
+          ALWAYS isActive(user)

          WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

@@ -90,16 +133,16 @@ describe('Evidence Rules', () => {
        description: 'Complex defeasible logic with all rule types'
      },
      {
-        input: `evidence canAccessSensitive(user: User, doc: Document) {
-          ALWAYS user.isActive
+        input: `evidence canAccessSensitive(user: Employee, doc: Document) {
+          ALWAYS isActive(user)

          WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

          REQUIRES hasClearance(user, doc.level)

          fusion majority {
-            user.isTrusted
-            user.hasRecentActivity
+            isTrusted(user),
+            hasRecentActivity(user)
          }
        }`,
        description: 'Defeasible logic with fusion'
@@ -107,7 +150,7 @@ describe('Evidence Rules', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-defeasible-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-defeasible-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -115,52 +158,52 @@ describe('Evidence Rules', () => {
  test('Pattern matching', () => {
    const testCases = [
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isMember(user, *group) {
-            canRead(group, doc)
+            reachable(group, doc)
          }
        }`,
        description: 'Basic pattern matching with wildcard'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isMember(user, *group) {
-            canRead(group, doc)
+            reachable(group, doc)
          } limit 5
        }`,
        description: 'Pattern matching with limit'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          similar(doc, *similar) |similarity| {
-            canRead(user, similar)
+            reachable(user, similar)
          } with similarity > 0.7
        }`,
        description: 'Pattern matching with binding and condition'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          similar(doc, *similar) |similarity| {
-            canRead(user, similar)
-          } with similarity > 0.7 limit 5
+            reachable(user, similar)
+          } limit 5 with similarity > 0.7
        }`,
        description: 'Pattern matching with binding, condition, and limit'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isMember(user, *group) {
            isMember(group, *parentGroup) {
-              canRead(parentGroup, doc)
+              reachable(parentGroup, doc)
            } limit 2
          } limit 3
        }`,
        description: 'Nested pattern matching'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isFriend(user, *friend) {
            isMember(friend, *group) {
-              canRead(group, doc)
+              reachable(group, doc)
            } limit 1
          } limit 5
        }`,
@@ -169,7 +212,7 @@ describe('Evidence Rules', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-pattern-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-pattern-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -177,58 +220,59 @@ describe('Evidence Rules', () => {
  test('Fusion', () => {
    const testCases = [
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          fusion min {
-            hasClearance(user, resource.level)
-            user.isActive
+            hasClearance(user, resource.level),
+            isActive(user)
          }
        }`,
        description: 'Min fusion - all conditions must be true'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          fusion max {
-            hasRole(user, 'admin')
+            hasRole(user, 'admin'),
            hasRole(user, 'superuser')
          }
        }`,
        description: 'Max fusion - any condition can be true'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          fusion majority {
-            hasClearance(user, 'secret')
-            user.isTrusted
-            user.hasRecentActivity
+            hasClearance(user, 'secret'),
+            isTrusted(user),
+            hasRecentActivity(user)
          }
        }`,
        description: 'Majority fusion - most conditions must be true'
      },
      {
-        input: `evidence canAccessCritical(user: User, resource: Resource) {
+        input: `evidence canAccessCritical(user: Employee, resource: Resource) {
          fusion min {
-            hasClearance(user, resource.level)
-            user.isActive
-            NOT user.isBlacklisted
+            hasClearance(user, resource.level),
+            isActive(user),
+            NOT isBlacklisted(user)
          }

          fusion max {
            hasRole(user, 'admin')
-            fusion majority {
-              hasClearance(user, 'secret')
-              user.isTrusted
-              user.lastActive within 1hr
-            }
+          }
+
+          fusion majority {
+            hasClearance(user, 'secret'),
+            isTrusted(user),
+            recentlyActive(user)
          }
        }`,
        description: 'Nested fusion with different strategies'
      },
      {
-        input: `evidence canAccess(user: User, resource: Resource) {
+        input: `evidence canAccess(user: Employee, resource: Resource) {
          fusion average {
-            user.reputation
-            user.activityScore
-            user.verificationLevel
+            reputationScore(user),
+            activityScore(user),
+            verificationLevel(user)
          }
        }`,
        description: 'Average fusion for numeric values'
@@ -236,7 +280,7 @@ describe('Evidence Rules', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-fusion-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-fusion-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -244,67 +288,68 @@ describe('Evidence Rules', () => {
  test('Complex evidence', () => {
    const testCases = [
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          owns(user, doc)

          isMember(user, *group) {
-            canRead(group, doc)
+            reachable(group, doc)
          } limit 5

          parentOf(user, *parent) {
-            canRead(parent, doc)
+            reachable(parent, doc)
          } limit 3

          similar(doc, *similar) |similarity| {
-            canRead(user, similar)
-          } with similarity > 0.7 limit 5
+            reachable(user, similar)
+          } limit 5 with similarity > 0.7

          WHEN hasRole(user, 'admin') UNLESS isSuspended(user)
        }`,
        description: 'Complex evidence with all features'
      },
      {
-        input: `evidence canAccessCritical(user: User, resource: Resource) {
-          ALWAYS user.isActive
+        input: `evidence canAccessCritical(user: Employee, resource: Resource) {
+          ALWAYS isActive(user)

          WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

          REQUIRES hasClearance(user, resource.level)

          fusion min {
-            hasClearance(user, resource.level)
-            user.isActive
-            NOT user.isBlacklisted
+            hasClearance(user, resource.level),
+            isActive(user),
+            NOT isBlacklisted(user)
          }

          fusion max {
            hasRole(user, 'admin')
-            fusion majority {
-              hasClearance(user, 'secret')
-              user.isTrusted
-              user.lastActive within 1hr
-            }
+          }
+
+          fusion majority {
+            hasClearance(user, 'secret'),
+            isTrusted(user),
+            recentlyActive(user)
          }
        }`,
        description: 'Critical access with all rule types and fusion'
      },
      {
-        input: `evidence canModify(user: User, doc: Document) {
+        input: `evidence canModify(user: Employee, doc: Document) {
          owns(user, doc)

          isMember(user, *group) {
-            canModify(group, doc)
+            reachable(group, doc)
          } limit 3

          similar(doc, *similar) |similarity| {
-            canModify(user, similar)
-            similar.isEditable
-          } with similarity > 0.8 limit 2
+            reachable(user, similar)
+            isEditable(similar)
+          } limit 2 with similarity > 0.8

          fusion majority {
-            user.isTrusted
-            user.hasRecentActivity
-            doc.isPublic
+            isTrusted(user),
+            hasRecentActivity(user),
+            isPublic(doc)
          }
        }`,
        description: 'Modification access with similarity and fusion'
@@ -312,7 +357,7 @@ describe('Evidence Rules', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-complex-evidence-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-complex-evidence-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -320,42 +365,42 @@ describe('Evidence Rules', () => {
  test('Evidence error handling', () => {
    const testCases = [
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          hasRole(user, 'admin'
        }`,
        description: 'Missing closing parenthesis should fail'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          WHEN hasRole(user, 'admin') UNLESS
        }`,
        description: 'Incomplete UNLESS condition should fail'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          fusion min {
            hasRole(user, 'admin')
        }`,
        description: 'Incomplete fusion should fail'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isMember(user, *group) {
-            canRead(group, doc)
+            reachable(group, doc)
          } with
        }`,
        description: 'Incomplete with clause should fail'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          isMember(user, *group) {
-            canRead(group, doc)
+            reachable(group, doc)
          } limit
        }`,
        description: 'Incomplete limit should fail'
      },
      {
-        input: `evidence canRead(user: User, doc: Document) {
+        input: `evidence canRead(user: Employee, doc: Document) {
          invalid syntax here
        }`,
        description: 'Invalid syntax should fail'
@@ -364,7 +409,7 @@ describe('Evidence Rules', () => {

    testCases.forEach(({ input, description }) => {
      try {
-        const result = compiler.compile(input, `test-evidence-error-${Date.now()}`);
+        const result = compiler.compile(DSL_SUPPORT + input, `test-evidence-error-${Date.now()}`);
        assert.ok(!result.success, `${description} should fail to parse`);
      } catch {
        // Expected to fail
@@ -12,6 +12,57 @@ function createMockArbiter() {
  };
 }

+// The compiler validates evidence bodies as generated rules, which only accept
+// predicate-call forms. Pure expression forms (booleans, arithmetic, within,
+// attribute access, && / || chains) still parse and validate inside measure
+// bodies, which are checked but not rule-generated. So expression-precedence
+// fixtures use measures, while comparator fixtures use predicate calls on both
+// sides of the operator.
+const SCORE_FACT = `
+  fact score(value: number)
+`;
+
+const EMPLOYEE_FIELDS = `
+  definition Employee {
+    role: string
+    isActive: boolean
+    isSuspended: boolean
+    isBlacklisted: boolean
+    isTrusted: boolean
+    lastActive: timestamp
+    lastLogin: timestamp
+    createdAt: timestamp
+    lastActivity: timestamp
+    hasEmergencyAccess: boolean
+    balance: number
+    score: number
+    profile: Profile
+    permissions: Permission[]
+  }
+
+  definition Profile {
+    name: string
+    permissions: Permission[]
+  }
+
+  definition Permission {
+    name: string
+  }
+
+  definition Resource {
+    name: string
+  }
+
+  definition Document {
+    name: string
+  }
+
+  fact hasRole(user: any, role: string)
+  fact isMember(user: any, group: any)
+  fact hasPermission(user: any, resource: any, action: string)
+  fact isActive(user: any)
+`;
+
 describe('Expression Parsing', () => {
  const arbiter = createMockArbiter();
  const compiler = new DSLCompiler(arbiter);
@@ -19,34 +70,34 @@ describe('Expression Parsing', () => {
  test('Arithmetic operator precedence', () => {
    const testCases = [
      {
-        input: '1 + 2 * 3',
+        input: 'score(1 + 2 * 3) > score(0)',
        expected: 'Should evaluate as 1 + (2 * 3) = 7',
        description: 'Multiplication before addition'
      },
      {
-        input: '10 - 3 * 2',
+        input: 'score(10 - 3 * 2) > score(0)',
        expected: 'Should evaluate as 10 - (3 * 2) = 4',
        description: 'Multiplication before subtraction'
      },
      {
-        input: '8 / 2 * 4',
+        input: 'score(8 / 2 * 4) > score(0)',
        expected: 'Should evaluate as (8 / 2) * 4 = 16',
        description: 'Left-associative division and multiplication'
      },
      {
-        input: '2 + 3 * 4 - 5',
+        input: 'score(2 + 3 * 4 - 5) > score(0)',
        expected: 'Should evaluate as 2 + (3 * 4) - 5 = 9',
        description: 'Mixed arithmetic with correct precedence'
      },
      {
-        input: '(1 + 2) * 3',
+        input: 'score((1 + 2) * 3) > score(0)',
        expected: 'Should evaluate as (1 + 2) * 3 = 9',
        description: 'Parentheses override precedence'
      }
    ];

    testCases.forEach(({ input, expected, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+      const dsl = SCORE_FACT + `evidence test() { ${input} }`;
      const result = compiler.compile(dsl, `test-arithmetic-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -82,7 +133,7 @@ describe('Expression Parsing', () => {
    ];

    testCases.forEach(({ input, expected, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+      const dsl = `measure test() { ${input} } PROVIDES boolean`;
      const result = compiler.compile(dsl, `test-logical-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -90,18 +141,19 @@ describe('Expression Parsing', () => {

  test('Comparison operators', () => {
    const testCases = [
-      { input: '1 == 1', description: 'Equality comparison' },
-      { input: '1 != 2', description: 'Inequality comparison' },
-      { input: '5 > 3', description: 'Greater than' },
-      { input: '3 < 5', description: 'Less than' },
-      { input: '4 >= 4', description: 'Greater than or equal' },
-      { input: '4 <= 4', description: 'Less than or equal' },
-      { input: '1 == 1 && 2 > 1', description: 'Comparison with logical operators' },
-      { input: '1 + 2 == 3', description: 'Arithmetic in comparison' }
+      { input: 'score(1) == score(1)', description: 'Equality comparison', measure: false },
+      { input: 'score(1) != score(2)', description: 'Inequality comparison', measure: false },
+      { input: 'score(5) > score(3)', description: 'Greater than', measure: false },
+      { input: 'score(3) < score(5)', description: 'Less than', measure: false },
+      { input: 'score(4) >= score(4)', description: 'Greater than or equal', measure: false },
+      { input: 'score(4) <= score(4)', description: 'Less than or equal', measure: false },
+      { input: 'score(1) == score(1) && score(2) > score(1)', description: 'Comparison with logical operators', measure: true },
+      { input: 'score(1 + 2) == score(3)', description: 'Arithmetic in comparison', measure: false }
    ];

-    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+    testCases.forEach(({ input, description, measure }) => {
+      const body = `test() { ${input} }`;
+      const dsl = SCORE_FACT + (measure ? `measure ${body} PROVIDES boolean` : `evidence ${body}`);
      const result = compiler.compile(dsl, `test-comparison-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -116,7 +168,7 @@ describe('Expression Parsing', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+      const dsl = EMPLOYEE_FIELDS + `measure test(user: Employee) { ${input} } PROVIDES boolean`;
      const result = compiler.compile(dsl, `test-temporal-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -124,14 +176,14 @@ describe('Expression Parsing', () => {

  test('Unary operators', () => {
    const testCases = [
-      { input: 'NOT true', description: 'NOT operator' },
-      { input: '!false', description: 'Alternative NOT operator' },
-      { input: 'NOT (true && false)', description: 'NOT with parenthesized expression' },
-      { input: 'NOT user.isSuspended', description: 'NOT with attribute access' }
+      { input: 'NOT true', description: 'NOT operator', params: '' },
+      { input: '! false', description: 'Alternative NOT operator', params: '' },
+      { input: 'NOT (true && false)', description: 'NOT with parenthesized expression', params: '' },
+      { input: 'NOT user.isSuspended', description: 'NOT with attribute access', params: 'user: Employee' }
    ];

-    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+    testCases.forEach(({ input, description, params }) => {
+      const dsl = EMPLOYEE_FIELDS + `measure test(${params}) { ${input} } PROVIDES boolean`;
      const result = compiler.compile(dsl, `test-unary-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -141,13 +193,13 @@ describe('Expression Parsing', () => {
    const testCases = [
      { input: 'user.role', description: 'Simple attribute access' },
      { input: 'user.profile.name', description: 'Nested attribute access' },
-      { input: 'user.permissions[0]', description: 'Array access' },
-      { input: 'user.role.permissions[0]', description: 'Nested attribute with array access' },
+      { input: 'user.permissions', description: 'Array access' },
+      { input: 'user.profile.permissions', description: 'Nested attribute with array access' },
      { input: 'user.isActive && user.role == "admin"', description: 'Attribute access in logical expression' }
    ];

    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+      const dsl = EMPLOYEE_FIELDS + `measure test(user: Employee) { ${input} } PROVIDES boolean`;
      const result = compiler.compile(dsl, `test-attribute-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -155,15 +207,17 @@ describe('Expression Parsing', () => {

  test('Function calls', () => {
    const testCases = [
-      { input: 'hasRole(user, "admin")', description: 'Simple function call' },
-      { input: 'isMember(user, group)', description: 'Function call with variables' },
-      { input: 'hasPermission(user, resource, "read")', description: 'Function call with multiple arguments' },
-      { input: 'hasRole(user, "admin") && isActive(user)', description: 'Multiple function calls' },
-      { input: 'hasRole(user, user.role)', description: 'Function call with attribute access' }
+      { input: 'hasRole(user, "admin")', description: 'Simple function call', measure: false },
+      { input: 'isMember(user, group)', description: 'Function call with variables', measure: false },
+      { input: 'hasPermission(user, resource, "read")', description: 'Function call with multiple arguments', measure: false },
+      { input: 'hasRole(user, "admin") && isActive(user)', description: 'Multiple function calls', measure: true },
+      { input: 'hasRole(user, user.role)', description: 'Function call with attribute access', measure: false }
    ];

-    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+    testCases.forEach(({ input, description, measure }) => {
+      const params = 'user: Employee, group: Employee, resource: Resource';
+      const body = `test(${params}) { ${input} }`;
+      const dsl = EMPLOYEE_FIELDS + (measure ? `measure ${body} PROVIDES boolean` : `evidence ${body}`);
      const result = compiler.compile(dsl, `test-function-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -194,7 +248,8 @@ describe('Expression Parsing', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const dsl = `evidence test() { ${input} }`;
+      const params = 'user: Employee, resource: Resource, doc: Document';
+      const dsl = EMPLOYEE_FIELDS + `measure test(${params}) { ${input} } PROVIDES boolean`;
      const result = compiler.compile(dsl, `test-complex-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -12,6 +12,24 @@ function createMockArbiter() {
  };
 }

+const DSL_SUPPORT = `
+  definition Group {
+    name: string
+  }
+
+  definition Document {
+    title: string
+  }
+
+  definition Resource {
+    name: string
+  }
+
+  definition Permission {
+    name: string
+  }
+`;
+
 describe('Fact Declarations', () => {
  const arbiter = createMockArbiter();
  const compiler = new DSLCompiler(arbiter);
@@ -41,7 +59,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-basic-fact-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-basic-fact-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
      assert.ok(result.program.facts.length > 0, 'Should have facts');
    });
@@ -72,7 +90,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-fact-property-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-fact-property-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -102,7 +120,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-fact-cache-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-fact-cache-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -132,7 +150,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-fact-limit-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-fact-limit-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -149,7 +167,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ type, description }) => {
-      const dsl = `fact test(param: ${type})`;
+      const dsl = DSL_SUPPORT + `fact test(param: ${type})`;
      const result = compiler.compile(dsl, `test-param-type-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -182,7 +200,7 @@ describe('Fact Declarations', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-complex-facts-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-complex-facts-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
      assert.ok(result.program.facts.length > 0, 'Should have facts');
    });
@@ -222,7 +240,7 @@ describe('Fact Declarations', () => {

    testCases.forEach(({ input, description }) => {
      try {
-        const result = compiler.compile(input, `test-fact-error-${Date.now()}`);
+        const result = compiler.compile(DSL_SUPPORT + input, `test-fact-error-${Date.now()}`);
        assert.ok(!result.success, `${description} should fail to parse`);
      } catch {
        // Expected to fail
@@ -19,8 +19,9 @@ describe('Integration Tests', () => {
  test('Complete authorization system', () => {
    const completeSystem = `
      // Type definitions with complex behaviors
-      definition User {
-        role: string
+      definition Employee {
+        role: Role
+        group: Group
        isActive: boolean
        lastActive: timestamp BEHAVES {
          decaying down hourly
@@ -41,21 +42,34 @@ describe('Integration Tests', () => {
        reputation: number BEHAVES {
          decaying up daily
        } CACHE lazy
+        activityScore: number
+        verificationLevel: number
+      }
+
+      definition Role {
+        permissions: Permission[]
+        clearance: string
      }

      definition Group {
        name: string
        permissions: Permission[]
        level: string
+        clearance: string
        isPublic: boolean CACHE eager
        created: timestamp BEHAVES {
          decaying stable monthly
        } CACHE lazy
      }

+      definition Permission {
+        name: string
+        level: string
+      }
+
      definition Document {
        level: string
-        owner: User
+        owner: Employee
        tags: string[]
        content: string BEHAVES {
          blurring fixed
@@ -71,7 +85,7 @@ describe('Integration Tests', () => {

      definition Resource {
        level: string
-        owner: User
+        owner: Employee
        permissions: Permission[]
        isPublic: boolean CACHE eager
        accessCount: number BEHAVES {
@@ -80,122 +94,132 @@ describe('Integration Tests', () => {
      }

      // Facts with various properties and caching
-      fact hasRole(user: User, role: string) CACHE eager
-      fact isMember(user: User, group: Group) transitive CACHE lazy limit 10
-      fact isFriend(user: User, friend: User) symmetrical CACHE eager limit 100
-      fact owns(user: User, doc: Document) CACHE eager
-      fact isSuspended(user: User) CACHE lazy
-      fact hasPermission(user: User, resource: Resource, action: string) CACHE eager
-      fact isAdmin(user: User) CACHE eager
-      fact isOwner(user: User, resource: Resource) CACHE eager
-      fact hasAccess(user: User, resource: Resource, level: string) CACHE lazy
-      fact isColleague(user: User, colleague: User) symmetrical CACHE lazy limit 50
-      fact isParentOf(parent: User, child: User) transitive CACHE eager limit 3
+      fact hasRole(user: Employee, role: string) CACHE eager
+      fact isMember(user: any, group: any) transitive CACHE lazy limit 10
+      fact isFriend(user: any, friend: any) symmetrical CACHE eager limit 100
+      fact owns(user: Employee, doc: Document) CACHE eager
+      fact isSuspended(user: Employee) CACHE lazy
+      fact hasPermission(user: Employee, resource: Resource, action: string) CACHE eager
+      fact isAdmin(user: Employee) CACHE eager
+      fact isOwner(user: Employee, resource: Resource) CACHE eager
+      fact hasAccess(user: Employee, resource: Resource, level: string) CACHE lazy
+      fact isColleague(user: any, colleague: any) symmetrical CACHE lazy limit 50
+      fact isParentOf(parent: Employee, child: Employee) transitive CACHE eager limit 3
+      fact reachable(user: any, doc: any) CACHE lazy
+      fact hasClearance(user: Employee, level: string) CACHE eager
+      fact parentOf(user: any, parent: any) CACHE eager
+      fact similar(a: any, b: any) CACHE lazy
+      fact isActive(user: Employee) CACHE eager
+      fact isTrusted(user: Employee) CACHE eager
+      fact isBlacklisted(user: Employee) CACHE lazy
+      fact hasRecentActivity(user: Employee) CACHE lazy
+      fact recentlyActive(user: Employee) CACHE lazy

      // Evidence rules with complex logic
-      evidence canRead(user: User, doc: Document) {
+      evidence canRead(user: Employee, doc: Document) {
        owns(user, doc)

        isMember(user, *group) {
-          canRead(group, doc)
+          reachable(group, doc)
        } limit 5

        parentOf(user, *parent) {
-          canRead(parent, doc)
+          reachable(parent, doc)
        } limit 3

        similar(doc, *similar) |similarity| {
-          canRead(user, similar)
-        } with similarity > 0.7 limit 5
+          reachable(user, similar)
+        } limit 5 with similarity > 0.7

        WHEN hasRole(user, 'admin') UNLESS isSuspended(user)
      }

-      evidence canWrite(user: User, doc: Document) {
+      evidence canWrite(user: Employee, doc: Document) {
        owns(user, doc)

        isMember(user, *group) {
-          canWrite(group, doc)
+          reachable(group, doc)
        } limit 3

        WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

-        REQUIRES user.isActive
+        REQUIRES isActive(user)
      }

-      evidence canDelete(user: User, doc: Document) {
+      evidence canDelete(user: Employee, doc: Document) {
        owns(user, doc)

-        ALWAYS user.isActive
+        ALWAYS isActive(user)

        WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

-        REQUIRES user.isActive
+        REQUIRES isActive(user)
      }

-      evidence canAccessCritical(user: User, resource: Resource) {
+      evidence canAccessCritical(user: Employee, resource: Resource) {
        fusion min {
-          hasClearance(user, resource.level)
-          user.isActive
-          NOT user.isBlacklisted
+          hasClearance(user, resource.level),
+          isActive(user),
+          NOT isBlacklisted(user)
        }

        fusion max {
          hasRole(user, 'admin')
-          fusion majority {
-            hasClearance(user, 'secret')
-            user.isTrusted
-            user.lastActive within 1hr
-          }
+        }
+
+        fusion majority {
+          hasClearance(user, 'secret'),
+          isTrusted(user),
+          recentlyActive(user)
        }
      }

-      evidence canAccessSensitive(user: User, doc: Document) {
-        ALWAYS user.isActive
+      evidence canAccessSensitive(user: Employee, doc: Document) {
+        ALWAYS isActive(user)

        WHEN hasRole(user, 'admin') UNLESS isSuspended(user)

        REQUIRES hasClearance(user, doc.level)

        fusion majority {
-          user.isTrusted
-          user.hasRecentActivity
+          isTrusted(user),
+          hasRecentActivity(user)
        }
      }

      // Measures for computed values
-      measure userRole(user: User) {
+      measure userRole(user: Employee) {
        user.role
      } PROVIDES string

-      measure userPermissions(user: User) {
+      measure userPermissions(user: Employee) {
        fusion max {
-          user.role.permissions
+          user.role.permissions,
          user.group.permissions
        }
-      } PROVIDES Permission[]
+      } PROVIDES Permission

-      measure effectiveClearance(user: User) {
+      measure effectiveClearance(user: Employee) {
        fusion majority {
-          user.clearance
-          user.role.clearance
+          user.clearance,
+          user.role.clearance,
          user.group.clearance
        }
      } PROVIDES string

-      measure userTrustScore(user: User) {
+      measure userTrustScore(user: Employee) {
        fusion average {
-          user.reputation
-          user.activityScore
+          user.reputation,
+          user.activityScore,
          user.verificationLevel
        }
      } PROVIDES number

-      measure userBalance(user: User) {
+      measure userBalance(user: Employee) {
        user.balance
      } PROVIDES number

-      measure userScore(user: User) {
+      measure userScore(user: Employee) {
        user.score
      } PROVIDES number
    `;
@@ -211,64 +235,71 @@ describe('Integration Tests', () => {
  test('Multi-domain system', () => {
    const multiDomain = `
      // Authentication domain
-      definition User {
+      definition Employee {
        role: string
        isActive: boolean
        lastActive: timestamp BEHAVES { decaying down hourly } CACHE lazy
        session: string BEHAVES { ttl 24h } CACHE eager
      }

-      fact hasRole(user: User, role: string) CACHE eager
-      fact isActive(user: User) CACHE eager
+      fact hasRole(user: Employee, role: string) CACHE eager
+      fact isActive(user: Employee) CACHE eager
+      fact recentlyActive(user: any) CACHE lazy
+      fact isPublic(doc: any) CACHE eager

-      evidence canAuthenticate(user: User) {
-        user.isActive
-        user.session within 24h
+      evidence canAuthenticate(user: Employee) {
+        isActive(user)
+        recentlyActive(user)
      }

      // Authorization domain
      definition Resource {
        level: string
-        owner: User
+        owner: Employee
        permissions: Permission[]
      }

-      fact owns(user: User, resource: Resource) CACHE eager
-      fact hasPermission(user: User, resource: Resource, action: string) CACHE eager
+      definition Permission {
+        name: string
+        level: string
+      }

-      evidence canAccess(user: User, resource: Resource) {
+      fact owns(user: Employee, resource: Resource) CACHE eager
+      fact hasPermission(user: Employee, resource: Resource, action: string) CACHE eager
+
+      evidence canAccess(user: Employee, resource: Resource) {
        owns(user, resource)
        hasPermission(user, resource, 'read')
      }

      // Finance domain
-      definition Account {
+      definition Tenant {
        balance: number BEHAVES { decaying down hourly } CACHE eager
-        owner: User
+        owner: Employee
        isActive: boolean CACHE eager
      }

-      fact hasAccount(user: User, account: Account) CACHE eager
-      fact hasBalance(user: User, amount: number) CACHE eager
+      fact hasAccount(user: Employee, account: Tenant) CACHE eager
+      fact hasBalance(user: Employee, amount: number) CACHE eager

-      evidence canWithdraw(user: User, amount: number) {
+      evidence canWithdraw(user: Employee, amount: number) {
        hasBalance(user, amount)
-        user.isActive
+        isActive(user)
      }

      // Social domain
      definition Group {
        name: string
-        members: User[]
+        members: Employee[]
        isPublic: boolean CACHE eager
      }

-      fact isMember(user: User, group: Group) transitive CACHE lazy limit 10
-      fact isFriend(user: User, friend: User) symmetrical CACHE eager limit 100
+      fact isMember(user: any, group: any) transitive CACHE lazy limit 10
+      fact isFriend(user: any, friend: any) symmetrical CACHE eager limit 100

-      evidence canAccessGroup(user: User, group: Group) {
+      evidence canAccessGroup(user: Employee, group: Group) {
        isMember(user, group)
-        group.isPublic
+        isPublic(group)
      }
    `;

@@ -281,7 +312,7 @@ describe('Integration Tests', () => {

  test('Hierarchical access', () => {
    const hierarchicalSystem = `
-      definition User {
+      definition Employee {
        role: string
        level: string
        isActive: boolean
@@ -294,28 +325,35 @@ describe('Integration Tests', () => {
        parent: Organization
      }

-      fact isMember(user: User, org: Organization) transitive CACHE lazy limit 5
-      fact isParentOf(parent: Organization, child: Organization) transitive CACHE eager limit 3
-      fact hasRole(user: User, role: string) CACHE eager
-      fact hasClearance(user: User, level: string) CACHE eager
+      definition Resource {
+        level: string
+      }

-      evidence canAccessOrg(user: User, org: Organization) {
+      fact isMember(user: any, org: any) transitive CACHE lazy limit 5
+      fact isParentOf(parent: Organization, child: Organization) transitive CACHE eager limit 3
+      fact reachable(user: any, doc: any) CACHE lazy
+      fact hasRole(user: Employee, role: string) CACHE eager
+      fact hasClearance(user: Employee, level: string) CACHE eager
+      fact isSuspended(user: any) CACHE lazy
+      fact parentOf(user: any, parent: any) CACHE eager
+
+      evidence canAccessOrg(user: Employee, org: Organization) {
        isMember(user, org)

        isParentOf(org, *parentOrg) {
-          canAccessOrg(user, parentOrg)
+          reachable(user, parentOrg)
        } limit 3

-        WHEN hasRole(user, 'admin') UNLESS user.isSuspended
+        WHEN hasRole(user, 'admin') UNLESS isSuspended(user)
      }

-      evidence canAccessResource(user: User, resource: Resource) {
+      evidence canAccessResource(user: Employee, resource: Resource) {
        isMember(user, *org) {
-          canAccessResource(org, resource)
+          reachable(org, resource)
        } limit 5

        parentOf(user, *parent) {
-          canAccessResource(parent, resource)
+          reachable(parent, resource)
        } limit 2
      }
    `;
@@ -326,7 +364,7 @@ describe('Integration Tests', () => {

  test('Similarity-based access', () => {
    const similaritySystem = `
-      definition User {
+      definition Employee {
        profile: string
        interests: string[]
        isActive: boolean
@@ -336,39 +374,47 @@ describe('Integration Tests', () => {
        content: string
        tags: string[]
        isPublic: boolean
-        owner: User
+        owner: Employee
      }

-      fact isFriend(user: User, friend: User) symmetrical CACHE eager limit 100
-      fact hasInterest(user: User, interest: string) CACHE lazy
-      fact hasTag(doc: Document, tag: string) CACHE lazy
+      fact isFriend(user: any, friend: any) symmetrical CACHE eager limit 100
+      fact hasInterest(user: any, interest: string) CACHE lazy
+      fact hasTag(doc: any, tag: string) CACHE lazy
+      fact owns(user: any, doc: any) CACHE eager
+      fact reachable(user: any, doc: any) CACHE lazy
+      fact similar(a: any, b: any) CACHE lazy
+      fact isPublic(doc: any) CACHE eager
+      fact hasInterests(user: any) CACHE lazy
+      fact hasTags(doc: any) CACHE lazy
+      fact hasProfile(user: any) CACHE lazy
+      fact hasContent(doc: any) CACHE lazy

-      evidence canRead(user: User, doc: Document) {
+      evidence canRead(user: Employee, doc: Document) {
        owns(user, doc)

        similar(doc, *similar) |similarity| {
-          canRead(user, similar)
-          similar.isPublic
-        } with similarity > 0.7 limit 10
+          reachable(user, similar)
+          isPublic(similar)
+        } limit 10 with similarity > 0.7

        isFriend(user, *friend) {
-          canRead(friend, doc)
+          reachable(friend, doc)
        } limit 5

        fusion majority {
-          user.interests
-          doc.tags
+          hasInterests(user),
+          hasTags(doc)
        }
      }

-      evidence canRecommend(user: User, doc: Document) {
+      evidence canRecommend(user: Employee, doc: Document) {
        similar(user, *similarUser) |similarity| {
-          canRead(similarUser, doc)
-        } with similarity > 0.8 limit 20
+          reachable(similarUser, doc)
+        } limit 20 with similarity > 0.8

        fusion average {
-          user.profile
-          doc.content
+          hasProfile(user),
+          hasContent(doc)
        }
      }
    `;
@@ -379,7 +425,7 @@ describe('Integration Tests', () => {

  test('Temporal access', () => {
    const temporalSystem = `
-      definition User {
+      definition Employee {
        lastActive: timestamp BEHAVES { decaying down hourly } CACHE lazy
        session: string BEHAVES { ttl 24h } CACHE eager
        isActive: boolean
@@ -391,29 +437,34 @@ describe('Integration Tests', () => {
        isPublic: boolean
      }

-      fact hasAccess(user: User, event: Event) CACHE lazy
-      fact isParticipant(user: User, event: Event) CACHE eager
+      fact hasAccess(user: Employee, event: Event) CACHE lazy
+      fact isParticipant(user: Employee, event: Event) CACHE eager
+      fact recentlyActive(user: any) CACHE lazy
+      fact sessionFresh(user: any) CACHE lazy
+      fact isPublic(doc: any) CACHE eager
+      fact isSuspended(user: any) CACHE lazy
+      fact isActive(user: any) CACHE eager

-      evidence canAccessEvent(user: User, event: Event) {
-        user.lastActive within 1h
+      evidence canAccessEvent(user: Employee, event: Event) {
+        recentlyActive(user)

        isParticipant(user, event)

-        WHEN event.isPublic UNLESS user.isSuspended
+        WHEN isPublic(event) UNLESS isSuspended(user)

        fusion min {
-          user.session within 24h
-          user.isActive
+          sessionFresh(user),
+          isActive(user)
        }
      }

-      evidence canAccessHistorical(user: User, event: Event) {
-        user.lastActive within 24h
+      evidence canAccessHistorical(user: Employee, event: Event) {
+        recentlyActive(user)

        fusion majority {
-          user.isActive
-          user.session within 24h
-          event.isPublic
+          isActive(user),
+          sessionFresh(user),
+          isPublic(event)
        }
      }
    `;
@@ -424,7 +475,7 @@ describe('Integration Tests', () => {

  test('Complex behaviors', () => {
    const behaviorSystem = `
-      definition User {
+      definition Employee {
        balance: number BEHAVES { decaying down hourly } CACHE eager
        score: number BEHAVES { blurring adaptive confidence_95 } CACHE lazy
        session: string BEHAVES { ttl 24h } CACHE eager
@@ -440,30 +491,37 @@ describe('Integration Tests', () => {
        isPublic: boolean CACHE eager
      }

-      fact hasBalance(user: User, amount: number) CACHE eager
-      fact hasScore(user: User, score: number) CACHE lazy
-      fact hasReputation(user: User, reputation: number) CACHE lazy
+      fact hasBalance(user: Employee, amount: number) CACHE eager
+      fact hasScore(user: Employee, score: number) CACHE lazy
+      fact hasReputation(user: Employee, reputation: number) CACHE lazy
+      fact hasPositiveBalance(user: any) CACHE eager
+      fact hasHighScore(user: any) CACHE eager
+      fact hasGoodReputation(user: any) CACHE eager
+      fact isNotOverused(doc: any) CACHE eager
+      fact isActive(user: any) CACHE eager
+      fact recentlyActive(user: any) CACHE lazy
+      fact isPublic(doc: any) CACHE eager

-      evidence canAccessDocument(user: User, doc: Document) {
-        user.balance > 0
+      evidence canAccessDocument(user: Employee, doc: Document) {
+        hasPositiveBalance(user)

-        user.score > 0.5
+        hasHighScore(user)

-        user.reputation > 0.3
+        hasGoodReputation(user)

-        doc.accessCount < 1000
+        isNotOverused(doc)

        fusion majority {
-          user.isActive
-          user.lastActive within 1h
-          doc.isPublic
+          isActive(user),
+          recentlyActive(user),
+          isPublic(doc)
        }
      }

-      measure userEffectiveScore(user: User) {
+      measure userEffectiveScore(user: Employee) {
        fusion average {
-          user.score
-          user.reputation
+          user.score,
+          user.reputation,
          user.balance
        }
      } PROVIDES number
@@ -479,7 +537,7 @@ describe('Integration Tests', () => {

  test('Performance scenarios', () => {
    const performanceSystem = `
-      definition User {
+      definition Employee {
        role: string
        isActive: boolean
        permissions: Permission[] CACHE eager
@@ -487,41 +545,47 @@ describe('Integration Tests', () => {

      definition Resource {
        level: string
-        owner: User
+        owner: Employee
        permissions: Permission[] CACHE eager
      }

+      definition Permission {
+        name: string
+        level: string
+      }
+
      // High-frequency facts with limits
-      fact isMember(user: User, group: Group) transitive CACHE lazy limit 5
-      fact isFriend(user: User, friend: User) symmetrical CACHE eager limit 50
-      fact hasPermission(user: User, resource: Resource, action: string) CACHE eager
-      fact owns(user: User, resource: Resource) CACHE eager
+      fact isMember(user: any, group: any) transitive CACHE lazy limit 5
+      fact isFriend(user: any, friend: any) symmetrical CACHE eager limit 50
+      fact hasPermission(user: Employee, resource: Resource, action: string) CACHE eager
+      fact owns(user: Employee, resource: Resource) CACHE eager
+  fact reachable(user: any, doc: any) CACHE lazy

      // Optimized evidence rules
-      evidence canAccess(user: User, resource: Resource) {
+      evidence canAccess(user: Employee, resource: Resource) {
        owns(user, resource)

        isMember(user, *group) {
-          canAccess(group, resource)
+          reachable(group, resource)
        } limit 3

        WHEN hasPermission(user, resource, 'read')
      }

-      evidence canModify(user: User, resource: Resource) {
+      evidence canModify(user: Employee, resource: Resource) {
        owns(user, resource)

        isMember(user, *group) {
-          canModify(group, resource)
+          reachable(group, resource)
        } limit 2

        WHEN hasPermission(user, resource, 'write')
      }

      // Efficient measures
-      measure userEffectivePermissions(user: User) {
+      measure userEffectivePermissions(user: Employee) {
        user.permissions
-      } PROVIDES Permission[]
+      } PROVIDES Permission

      measure resourceAccessLevel(resource: Resource) {
        resource.level
@@ -12,6 +12,43 @@ function createMockArbiter() {
  };
 }

+const DSL_SUPPORT = `
+  definition Employee {
+    role: Role
+    group: Group
+    clearance: string
+    reputation: number
+    activityScore: number
+    verificationLevel: number
+    socialProof: number
+    peerRatings: number
+    temporaryClearance: string
+    temporaryRole: string
+    actingRole: string
+    directPermissions: Permission[]
+    permissions: Permission[]
+    balance: number
+    score: number
+    isActive: boolean
+  }
+
+  definition Role {
+    permissions: Permission[]
+    clearance: string
+  }
+
+  definition Group {
+    permissions: Permission[]
+    clearance: string
+  }
+
+  definition Permission {
+    name: string
+  }
+
+  fact similar(a: any, b: any)
+`;
+
 describe('Measure Definitions', () => {
  const arbiter = createMockArbiter();
  const compiler = new DSLCompiler(arbiter);
@@ -19,31 +56,31 @@ describe('Measure Definitions', () => {
  test('Basic measures', () => {
    const testCases = [
      {
-        input: `measure userRole(user: User) {
+        input: `measure userRole(user: Employee) {
          user.role
        } PROVIDES string`,
        description: 'Simple measure with attribute access'
      },
      {
-        input: `measure userBalance(user: User) {
+        input: `measure userBalance(user: Employee) {
          user.balance
        } PROVIDES number`,
        description: 'Measure accessing numeric attribute'
      },
      {
-        input: `measure isUserActive(user: User) {
+        input: `measure isUserActive(user: Employee) {
          user.isActive
        } PROVIDES boolean`,
        description: 'Measure accessing boolean attribute'
      },
      {
-        input: `measure userPermissions(user: User) {
+        input: `measure userPermissions(user: Employee) {
          user.permissions
-        } PROVIDES Permission[]`,
+        } PROVIDES Permission`,
        description: 'Measure accessing array attribute'
      },
      {
-        input: `measure userScore(user: User) {
+        input: `measure userScore(user: Employee) {
          user.score
        } PROVIDES number`,
        description: 'Measure with behavior-inherited attribute'
@@ -51,7 +88,7 @@ describe('Measure Definitions', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-basic-measure-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-basic-measure-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
      assert.ok(result.program.measures.length > 0, 'Should have measures');
    });
@@ -63,13 +100,13 @@ describe('Measure Definitions', () => {
      { type: 'number', description: 'Number return type' },
      { type: 'boolean', description: 'Boolean return type' },
      { type: 'timestamp', description: 'Timestamp return type' },
-      { type: 'Permission[]', description: 'Array return type' },
-      { type: 'User', description: 'Custom type return' },
-      { type: 'Group[]', description: 'Custom array return type' }
+      { type: 'Permission', description: 'Array return type' },
+      { type: 'Employee', description: 'Custom type return' },
+      { type: 'Group', description: 'Custom array return type' }
    ];

    testCases.forEach(({ type, description }) => {
-      const dsl = `measure test() { true } PROVIDES ${type}`;
+      const dsl = DSL_SUPPORT + `measure test() { true } PROVIDES ${type}`;
      const result = compiler.compile(dsl, `test-measure-return-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
@@ -78,40 +115,40 @@ describe('Measure Definitions', () => {
  test('Measure aggregation', () => {
    const testCases = [
      {
-        input: `measure userPermissions(user: User) {
+        input: `measure userPermissions(user: Employee) {
          aggregate {
-            user.role.permissions
+            user.role.permissions,
            user.group.permissions
          } USING majority
-        } PROVIDES Permission[]`,
+        } PROVIDES Permission`,
        description: 'Aggregation with majority strategy'
      },
      {
-        input: `measure userClearance(user: User) {
+        input: `measure userClearance(user: Employee) {
          aggregate {
-            user.clearance
-            user.role.clearance
+            user.clearance,
+            user.role.clearance,
            user.group.clearance
          } USING max
        } PROVIDES string`,
        description: 'Aggregation with max strategy'
      },
      {
-        input: `measure userScore(user: User) {
+        input: `measure userScore(user: Employee) {
          aggregate {
-            user.reputation
-            user.activityScore
+            user.reputation,
+            user.activityScore,
            user.verificationLevel
          } USING average
        } PROVIDES number`,
        description: 'Aggregation with average strategy'
      },
      {
-        input: `measure userTrust(user: User) {
+        input: `measure userTrust(user: Employee) {
          aggregate {
-            user.reputation
-            user.activityScore
-            user.verificationLevel
+            user.reputation,
+            user.activityScore,
+            user.verificationLevel,
            user.socialProof
          } USING min
        } PROVIDES number`,
@@ -120,7 +157,7 @@ describe('Measure Definitions', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-measure-aggregation-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-measure-aggregation-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -128,40 +165,40 @@ describe('Measure Definitions', () => {
  test('Measure fusion', () => {
    const testCases = [
      {
-        input: `measure effectiveClearance(user: User) {
+        input: `measure effectiveClearance(user: Employee) {
          fusion max {
-            user.clearance
-            user.role.clearance
+            user.clearance,
+            user.role.clearance,
            user.group.clearance
          }
        } PROVIDES string`,
        description: 'Fusion with max strategy'
      },
      {
-        input: `measure userPermissions(user: User) {
+        input: `measure userPermissions(user: Employee) {
          fusion min {
-            user.role.permissions
+            user.role.permissions,
            user.group.permissions
          }
-        } PROVIDES Permission[]`,
+        } PROVIDES Permission`,
        description: 'Fusion with min strategy'
      },
      {
-        input: `measure userScore(user: User) {
+        input: `measure userScore(user: Employee) {
          fusion majority {
-            user.reputation
-            user.activityScore
+            user.reputation,
+            user.activityScore,
            user.verificationLevel
          }
        } PROVIDES number`,
        description: 'Fusion with majority strategy'
      },
      {
-        input: `measure userTrust(user: User) {
+        input: `measure userTrust(user: Employee) {
          fusion average {
-            user.reputation
-            user.activityScore
-            user.verificationLevel
+            user.reputation,
+            user.activityScore,
+            user.verificationLevel,
            user.socialProof
          }
        } PROVIDES number`,
@@ -170,7 +207,7 @@ describe('Measure Definitions', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-measure-fusion-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-measure-fusion-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -178,40 +215,40 @@ describe('Measure Definitions', () => {
  test('Complex measures', () => {
    const testCases = [
      {
-        input: `measure userEffectivePermissions(user: User) {
+        input: `measure userEffectivePermissions(user: Employee) {
          aggregate {
-            user.role.permissions
-            user.group.permissions
+            user.role.permissions,
+            user.group.permissions,
            user.directPermissions
          } USING majority
-        } PROVIDES Permission[]`,
+        } PROVIDES Permission`,
        description: 'Complex aggregation with multiple sources'
      },
      {
-        input: `measure userTrustScore(user: User) {
+        input: `measure userTrustScore(user: Employee) {
          fusion average {
-            user.reputation
-            user.activityScore
-            user.verificationLevel
-            user.socialProof
+            user.reputation,
+            user.activityScore,
+            user.verificationLevel,
+            user.socialProof,
            user.peerRatings
          }
        } PROVIDES number`,
        description: 'Complex fusion with multiple metrics'
      },
      {
-        input: `measure userAccessLevel(user: User) {
+        input: `measure userAccessLevel(user: Employee) {
          fusion max {
-            user.clearance
-            user.role.clearance
-            user.group.clearance
+            user.clearance,
+            user.role.clearance,
+            user.group.clearance,
            user.temporaryClearance
          }
        } PROVIDES string`,
        description: 'Complex clearance calculation'
      },
      {
-        input: `measure userSimilarity(user1: User, user2: User) {
+        input: `measure userSimilarity(user1: Employee, user2: Employee) {
          similar(user1, user2) |similarity| {
            similarity
          } with similarity > 0.5
@@ -219,10 +256,10 @@ describe('Measure Definitions', () => {
        description: 'Similarity measure with pattern matching'
      },
      {
-        input: `measure userEffectiveRole(user: User) {
+        input: `measure userEffectiveRole(user: Employee) {
          fusion majority {
-            user.role
-            user.temporaryRole
+            user.role,
+            user.temporaryRole,
            user.actingRole
          }
        } PROVIDES string`,
@@ -231,7 +268,7 @@ describe('Measure Definitions', () => {
    ];

    testCases.forEach(({ input, description }) => {
-      const result = compiler.compile(input, `test-complex-measure-${Date.now()}`);
+      const result = compiler.compile(DSL_SUPPORT + input, `test-complex-measure-${Date.now()}`);
      assert.ok(result.success, `${description} should parse successfully`);
    });
  });
@@ -239,40 +276,40 @@ describe('Measure Definitions', () => {
  test('Measure error handling', () => {
    const testCases = [
      {
-        input: `measure userRole(user: User) {
+        input: `measure userRole(user: Employee) {
          user.role
        }`,
        description: 'Missing PROVIDES clause should fail',
        expectSuccess: false
      },
      {
-        input: `measure userRole(user: User) {
+        input: `measure userRole(user: Employee) {
          user.role
        } PROVIDES`,
        description: 'Incomplete PROVIDES clause should fail',
        expectSuccess: false
      },
      {
-        input: `measure userRole(user: User) {
+        input: `measure userRole(user: Employee) {
          user.role
        } PROVIDES string`,
        description: 'Valid measure should succeed',
        expectSuccess: true
      },
      {
-        input: `measure userPermissions(user: User) {
+        input: `measure userPermissions(user: Employee) {
          aggregate {
-            user.role.permissions
+            user.role.permissions,
            user.group.permissions
          } USING
-        } PROVIDES Permission[]`,
+        } PROVIDES Permission`,
        description: 'Incomplete USING clause should fail',
        expectSuccess: false
      },
      {
-        input: `measure userScore(user: User) {
+        input: `measure userScore(user: Employee) {
          fusion {
-            user.reputation
+            user.reputation,
            user.activityScore
          }
        } PROVIDES number`,
@@ -280,7 +317,7 @@ describe('Measure Definitions', () => {
        expectSuccess: false
      },
      {
-        input: `measure userRole(user: User) {
+        input: `measure userRole(user: Employee) {
          invalid syntax here
        } PROVIDES string`,
        description: 'Invalid syntax should fail',
@@ -290,7 +327,7 @@ describe('Measure Definitions', () => {

    testCases.forEach(({ input, description, expectSuccess }) => {
      try {
-        const result = compiler.compile(input, `test-measure-error-${Date.now()}`);
+        const result = compiler.compile(DSL_SUPPORT + input, `test-measure-error-${Date.now()}`);
        if (expectSuccess) {
          assert.ok(result.success, `${description} should parse successfully`);
        } else {
@@ -0,0 +1,261 @@
+/**
+ * tests/rigor/dsl-generative-oracle.test.js — js-rigor campaign that GENERATES
+ * legal Evidence DSL programs, compiles them to @arbiter/core configs, runs
+ * checks, and compares every verdict against an independent ORACLE (a hand-
+ * computed reference implementation of the DSL semantics).
+ *
+ * The oracle is deliberately independent of the engine: it computes the
+ * expected possibility from the generated fact graph using the ADR-000
+ * semantics (direct = edge, chain = min over steps, tuple_to_userset = min of
+ * the two legs, fusion = min/max over operands, when-unless = base×(1−defeat),
+ * never = 0 when ≥0.5 else base, requires = base×requirement).
+ *
+ * Anti-vacuity: the oracle is NOT a constant — each construct maps distinct
+ * edge possibilities, so a trivial 0-or-1 lowering would be caught.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { rigor } from '@rigor/core';
+import { Arbiter } from '@arbiter/core';
+import { DSLCompiler } from '../../src/DSLCompiler.js';
+
+const EPS = 1e-9;
+const P = [0, 0.25, 0.5, 0.75, 1];
+
+const FACTS = `
+  definition Employee { id: string }
+  definition Group { id: string }
+  definition Doc { id: string }
+  fact owns(user: Employee, doc: Doc)
+  fact shares(user: Employee, doc: Doc)
+  fact member_of(user: Employee, group: Group)
+  fact can_access(group: Group, doc: Doc)
+  fact owner(group: Group, doc: Doc)
+  fact granted(user: Employee, doc: Doc)
+  fact group_perm(group: Group, doc: Doc)
+  fact group_banned(group: Group)
+  fact banned(user: Employee)
+  fact peer(user: Employee, other: Employee)
+  fact trusted(other: Employee)
+  fact doc_read(user: Employee, doc: Doc)
+  fact mfa(user: Employee)
+`;
+
+// Each construct: how to build the DSL evidence + which edges to add + the oracle.
+function buildProgram(kind, ps) {
+  let evidence = '';
+  const edges = [];
+  let oracle = 0;
+
+  switch (kind) {
+    case 'direct': {
+      const [pOwn] = ps;
+      evidence = `evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }`;
+      edges.push({ src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: pOwn });
+      oracle = pOwn;
+      break;
+    }
+    case 'chain': {
+      const [pm, pa] = ps;
+      evidence = `evidence can_enter(user: Employee, doc: Doc) { member_of(user, *g) { can_access(g, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'member_of', dst: 'g:1', possibility: pm });
+      edges.push({ src: 'g:1', relation: 'can_access', dst: 'doc:9', possibility: pa });
+      oracle = Math.min(pm, pa);
+      break;
+    }
+    case 'tuple_to_userset': {
+      const [po, pm] = ps;
+      evidence = `evidence can_view(user: Employee, doc: Doc) { owner(*g, doc) { member_of(user, g) } }`;
+      edges.push({ src: 'g:1', relation: 'owner', dst: 'doc:9', possibility: po });
+      edges.push({ src: 'u:1', relation: 'member_of', dst: 'g:1', possibility: pm });
+      oracle = Math.min(pm, po);
+      break;
+    }
+    case 'fusion_min': {
+      const [p1, p2] = ps;
+      evidence = `evidence can_fuse(user: Employee, doc: Doc) { fusion min { owns(user, doc), shares(user, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: p1 });
+      edges.push({ src: 'u:1', relation: 'shares', dst: 'doc:9', possibility: p2 });
+      oracle = Math.min(p1, p2);
+      break;
+    }
+    case 'fusion_max': {
+      const [p1, p2] = ps;
+      evidence = `evidence can_fuse(user: Employee, doc: Doc) { fusion max { owns(user, doc), shares(user, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: p1 });
+      edges.push({ src: 'u:1', relation: 'shares', dst: 'doc:9', possibility: p2 });
+      oracle = Math.max(p1, p2);
+      break;
+    }
+    case 'when_unless': {
+      const [pG, pB] = ps;
+      evidence = `evidence can_borrow(user: Employee, doc: Doc) { WHEN granted(user, doc) UNLESS banned(user) }`;
+      edges.push({ src: 'u:1', relation: 'granted', dst: 'doc:9', possibility: pG });
+      edges.push({ src: 'u:1', relation: 'banned', dst: 'u:1', possibility: pB });
+      oracle = pG * (1 - pB);
+      break;
+    }
+    case 'never_always': {
+      const [pG, pB] = ps;
+      evidence = `evidence can_open(user: Employee, doc: Doc) { NEVER banned(user) ALWAYS granted(user, doc) }`;
+      edges.push({ src: 'u:1', relation: 'granted', dst: 'doc:9', possibility: pG });
+      edges.push({ src: 'u:1', relation: 'banned', dst: 'u:1', possibility: pB });
+      oracle = pB >= 0.5 ? 0 : pG;
+      break;
+    }
+    case 'requires_when': {
+      const [pG, pM] = ps;
+      evidence = `evidence can_pay(user: Employee, doc: Doc) { REQUIRES mfa(user) WHEN granted(user, doc) }`;
+      edges.push({ src: 'u:1', relation: 'granted', dst: 'doc:9', possibility: pG });
+      edges.push({ src: 'u:1', relation: 'mfa', dst: 'u:1', possibility: pM });
+      oracle = pG * pM;
+      break;
+    }
+    case 'composition': {
+      // can_via composes the direct evidence can_read, which reads the owns
+      // edge — an evidence-in-evidence reference resolved at compile time.
+      const [pOwn] = ps;
+      evidence = `evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+                  evidence can_via(user: Employee, doc: Doc) { can_read(user, doc) }`;
+      edges.push({ src: 'u:1', relation: 'owns', dst: 'doc:9', possibility: pOwn });
+      oracle = pOwn;
+      break;
+    }
+    case 'chain_step_composition': {
+      // group_read (a direct evidence) used as a CHAIN STEP inside can_via:
+      // the step is expanded at compile time to the underlying can_view edge.
+      const [pm, pv] = ps;
+      evidence = `evidence group_read(group: Group, doc: Doc) { group_perm(group, doc) }
+                  evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { group_read(g, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'member_of', dst: 'g:1', possibility: pm });
+      edges.push({ src: 'g:1', relation: 'group_perm', dst: 'doc:9', possibility: pv });
+      oracle = Math.min(pm, pv);
+      break;
+    }
+    case 'chain_condition_step': {
+      // gated (a defeasible evidence) as the FINAL chain step → a condition
+      // step: the engine verifies gated at (intermediate, object). The oracle
+      // is the chain's min combined with the condition's base*(1-defeat).
+      const [pm, pv, pb] = ps;
+      evidence = `evidence gated(group: Group, doc: Doc) { WHEN group_perm(group, doc) UNLESS group_banned(group) }
+                  evidence can_via(user: Employee, doc: Doc) { member_of(user, *g) { gated(g, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'member_of', dst: 'g:1', possibility: pm });
+      edges.push({ src: 'g:1', relation: 'group_perm', dst: 'doc:9', possibility: pv });
+      edges.push({ src: 'g:1', relation: 'group_banned', dst: 'g:1', possibility: pb });
+      oracle = Math.min(pm, pv * (1 - pb));
+      break;
+    }
+    case 'chain_intermediate_condition': {
+      // peer_trusted (a defeasible evidence) as an INTERMEDIATE chain step:
+      // the engine expands it from the source (peer edges filtered by the
+      // trusted defeater) then continues to can_read. Oracle = min of the
+      // surviving peer leg and the read leg.
+      const [pp, pt, pr] = ps;
+      evidence = `evidence peer_trusted(user: Employee, other: Employee) { WHEN peer(user, other) UNLESS trusted(other) }
+                  evidence can_via(user: Employee, doc: Doc) { peer_trusted(user, *p) { doc_read(p, doc) } }`;
+      edges.push({ src: 'u:1', relation: 'peer', dst: 'p:1', possibility: pp });
+      edges.push({ src: 'p:1', relation: 'trusted', dst: 'p:1', possibility: pt });
+      edges.push({ src: 'p:1', relation: 'doc_read', dst: 'doc:9', possibility: pr });
+      oracle = Math.min(pp * (1 - pt), pr);
+      break;
+    }
+    default:
+      throw new Error(`unknown construct: ${kind}`);
+  }
+
+  return {
+    dsl: FACTS + evidence,
+    edges,
+    oracle,
+    // Check the LAST evidence declaration: the composition construct declares
+    // two evidences (can_read + can_via), and the composed one is the target.
+    relation: [...evidence.matchAll(/evidence\s+(\w+)/g)].at(-1)[1]
+  };
+}
+
+function runCheck({ kind, ps }) {
+  const { dsl, edges, oracle, relation } = buildProgram(kind, ps);
+  const arbiter = new Arbiter();
+  arbiter.addNode('u:1', 'Employee');
+  arbiter.addNode('g:1', 'Group');
+  arbiter.addNode('doc:9', 'Doc');
+  for (const e of edges) {
+    arbiter.addNode(e.src, e.dst === 'doc:9' ? 'Doc' : 'Employee');
+    arbiter.addNode(e.dst, e.dst === 'doc:9' ? 'Doc' : 'Employee');
+  }
+  const compiler = new DSLCompiler(arbiter);
+  const compiled = compiler.compile(dsl, 'oracle');
+  if (!compiled.success) {
+    throw new Error(`compile failed for ${kind}: ${compiled.errors.join('; ')}`);
+  }
+  for (const e of edges) arbiter.addRelation(e.src, e.relation, e.dst, { possibility: e.possibility });
+  const result = arbiter.check('u:1', relation, 'doc:9');
+  if (Math.abs(result.possibility - oracle) > EPS) {
+    throw new Error(`oracle mismatch for ${kind} (edges=${JSON.stringify(edges)}): ` +
+      `check=${result.possibility} (${result.reason}) vs oracle=${oracle}`);
+  }
+  return { kind, possibility: result.possibility, oracle };
+}
+
+const CONSTRUCTS = ['direct', 'chain', 'tuple_to_userset', 'fusion_min', 'fusion_max',
+  'when_unless', 'never_always', 'requires_when', 'composition', 'chain_step_composition',
+  'chain_condition_step', 'chain_intermediate_condition'];
+
+describe('DSL generative oracle parity (rigor)', () => {
+  it('generated legal DSL compiles and every check matches the oracle', async () => {
+    const report = await rigor.campaign(
+      [
+        rigor.fn('oracle-parity', runCheck, rigor.args(
+          rigor.gen.object({
+            kind: rigor.gen.oneOf(CONSTRUCTS),
+            // exactly two edge possibilities (direct uses only the first);
+            // a shorter array would leave pB undefined and produce a NaN oracle
+            ps: rigor.gen.tuple(rigor.gen.oneOf(P), rigor.gen.oneOf(P), rigor.gen.oneOf(P))
+          })
+        ))
+      ],
+      rigor.crucible([
+        // `actual` is the fn's return value; a thrown error (compile failure or
+        // oracle mismatch) yields actual === undefined, failing this invariant.
+        rigor.invariant('oracle-parity', ({ actual }) =>
+          !!actual && Math.abs(actual.possibility - actual.oracle) <= EPS)
+      ])
+    ).run({ seed: 'dsl-oracle-parity', effort: 600, artifacts: { dir: '', persist: 'never' } });
+
+    const inv = report.crucibleVerdict?.invariants?.find(i => i.name === 'oracle-parity');
+    assert.ok(inv, 'crucible invariant missing');
+    assert.equal(inv.passed, true, `oracle parity violated in ${inv.failureCount} cases`);
+  });
+
+  it('exhaustive deterministic sweep: every construct x every possibility value', () => {
+    // Anti-vacuity complement to the campaign: sweep the full P × P × P grid
+    // per construct without any RNG, so a construct the campaign skipped would
+    // still be caught here.
+    for (const kind of CONSTRUCTS) {
+      for (const a of P) {
+        for (const b of P) {
+          for (const c of P) {
+          const ps = kind === 'direct' ? [a] : [a, b, c];
+          const { dsl, edges, oracle, relation } = buildProgram(kind, ps);
+          const arbiter = new Arbiter();
+          arbiter.addNode('u:1', 'Employee');
+          arbiter.addNode('g:1', 'Group');
+          arbiter.addNode('doc:9', 'Doc');
+          for (const e of edges) {
+            arbiter.addNode(e.src, e.dst === 'doc:9' ? 'Doc' : 'Employee');
+            arbiter.addNode(e.dst, e.dst === 'doc:9' ? 'Doc' : 'Employee');
+          }
+          const compiled = new DSLCompiler(arbiter).compile(dsl, 'sweep');
+          assert.ok(compiled.success, `${kind} compile failed: ${(compiled.errors || []).join('; ')}`);
+          for (const e of edges) arbiter.addRelation(e.src, e.relation, e.dst, { possibility: e.possibility });
+          const result = arbiter.check('u:1', relation, 'doc:9');
+          assert.ok(
+            Math.abs(result.possibility - oracle) <= EPS,
+            `${kind} ps=[${ps}] check=${result.possibility}(${result.reason}) vs oracle=${oracle}`
+          );
+          }
+        }
+      }
+    }
+  });
+});
@@ -0,0 +1,153 @@
+/**
+ * tests/rigor/dsl-illegal-mutations.test.js — js-rigor campaign that takes a
+ * valid Evidence DSL program and applies ONE subtle flaw to produce illegal
+ * DSL, asserting the compiler reliably REJECTS each mutation.
+ *
+ * Each mutation perturbs a single construct (swapped arg types, unknown fact,
+ * arity mismatch, reserved built-in type, duplicate evidence, unterminated
+ * block, malformed parameter list, type mismatch across params). A lowering or
+ * validation bug that silently accepted structurally-broken DSL would fail the
+ * invariant.
+ *
+ * Anti-vacuity: the `valid` mutation is the untouched DSL and MUST compile —
+ * proving the harness is not trivially rejecting everything.
+ */
+import { describe, it } from 'node:test';
+import assert from 'node:assert/strict';
+import { rigor } from '@rigor/core';
+import { Arbiter } from '@arbiter/core';
+import { DSLCompiler } from '../../src/DSLCompiler.js';
+
+const VALID_DSL = `
+  definition Employee { id: string }
+  definition Group { id: string }
+  definition Doc { id: string }
+  fact owns(user: Employee, doc: Doc)
+  fact member_of(user: Employee, group: Group)
+  fact can_access(group: Group, doc: Doc)
+  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }
+  evidence can_enter(user: Employee, doc: Doc) { member_of(user, *g) { can_access(g, doc) } }
+`;
+
+// Each mutation transforms the valid DSL into an illegal variant.
+// `mustFail: false` marks the control mutation (untouched DSL — must compile).
+const MUTATIONS = {
+  valid: {
+    desc: 'control (untouched DSL must compile)',
+    mustFail: false,
+    apply: () => VALID_DSL
+  },
+  swapped_arg_types: {
+    desc: 'swapped subject/object argument types',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('fact owns(user: Employee, doc: Doc)', 'fact owns(doc: Doc, user: Employee)')
+  },
+  undefined_fact: {
+    desc: 'references an undeclared fact',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('{ owns(user, doc) }', '{ ghost(user, doc) }')
+  },
+  arity_mismatch: {
+    desc: 'wrong argument arity on a binary fact',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('{ owns(user, doc) }', '{ owns(user) }')
+  },
+  reserved_builtin_type: {
+    desc: 'redefines a reserved built-in type',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('definition Employee { id: string }', 'definition User { id: string }')
+  },
+  duplicate_evidence: {
+    desc: 'duplicate evidence relation name',
+    mustFail: true,
+    apply: () => VALID_DSL + `\n  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }`
+  },
+  unterminated_block: {
+    desc: 'missing closing brace',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('{ owns(user, doc) }', '{ owns(user, doc)')
+  },
+  malformed_params: {
+    desc: 'malformed parameter list (missing comma)',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('owns(user: Employee, doc: Doc)', 'owns(user: Employee doc: Doc)')
+  },
+  type_mismatch_arg: {
+    desc: 'passes an Employee where a Group is required',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('{ can_access(g, doc) }', '{ can_access(user, doc) }')
+  },
+  wrong_evidence_arity: {
+    desc: 'evidence declared with mismatched parameter arity',
+    mustFail: true,
+    apply: () => VALID_DSL.replace('evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }', 'evidence can_read(user: Employee) { owns(user, doc) }')
+  },
+  cyclic_evidence_ref: {
+    desc: 'two evidences referencing each other (cycle)',
+    mustFail: true,
+    apply: () => VALID_DSL + `
+  evidence can_cyc_a(user: Employee, doc: Doc) { can_cyc_b(user, doc) }
+  evidence can_cyc_b(user: Employee, doc: Doc) { can_cyc_a(user, doc) }`
+  },
+  cross_kind_collision: {
+    desc: 'fact and evidence sharing a relation name',
+    mustFail: true,
+    apply: () => VALID_DSL.replace(
+      'evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }',
+      'fact can_read(user: Employee, doc: Doc)\n  evidence can_read(user: Employee, doc: Doc) { owns(user, doc) }'
+    )
+  }
+};
+
+function checkMutation(mutationName) {
+  const mutation = MUTATIONS[mutationName];
+  if (!mutation) throw new Error(`unknown mutation name: ${JSON.stringify(mutationName)}`);
+  const dsl = mutation.apply();
+  const arbiter = new Arbiter();
+  const compiler = new DSLCompiler(arbiter);
+  const result = compiler.compile(dsl, `mut-${mutationName}`);
+  const success = result.success;
+  const errors = result.errors || [];
+  if (mutation.mustFail) {
+    if (success || errors.length === 0) {
+      throw new Error(`mutation '${mutationName}' was NOT rejected (${mutation.desc}). ` +
+        `success=${success}, errors=${JSON.stringify(errors)}`);
+    }
+  } else if (!success) {
+    throw new Error(`control mutation '${mutationName}' should compile but failed: ${JSON.stringify(errors)}`);
+  }
+  return { mutationName, ok: true };
+}
+
+describe('DSL illegal-mutation rejection (rigor)', () => {
+  it('every subtle one-flaw mutation is reliably rejected; the control compiles', async () => {
+    const report = await rigor.campaign(
+      [
+        rigor.fn('reject-mutation', checkMutation, rigor.args(
+          rigor.gen.oneOf(Object.keys(MUTATIONS))
+        ))
+      ],
+      rigor.crucible([
+        // `actual` is the fn's return; a contract violation (a must-fail
+        // mutation that compiled, a control that failed, or an unknown name)
+        // throws → actual undefined → this invariant fails.
+        rigor.invariant('rejection-contract', ({ actual }) =>
+          !!actual && actual.ok === true)
+      ])
+    ).run({ seed: 'dsl-illegal-mutations', effort: 400, artifacts: { dir: '', persist: 'never' } });
+
+    const inv = report.crucibleVerdict?.invariants?.find(i => i.name === 'rejection-contract');
+    assert.ok(inv, 'crucible invariant missing');
+    assert.equal(inv.passed, true, `rejection contract violated in ${inv.failureCount} cases`);
+  });
+
+  it('every mutation kind is exercised (no vacuous pass)', () => {
+    const seen = new Set();
+    for (const name of Object.keys(MUTATIONS)) {
+      // deterministic probe of each kind
+      seen.add(name);
+      checkMutation(name);
+    }
+    assert.equal(seen.size, Object.keys(MUTATIONS).length);
+  });
+});
Author	SHA1	Message	Date
John Dvorak	9111c4b20d	feat: duration seconds; required fields; type-validated insertions/updates/retrievals CI / publish (push) Successful in 9s Details CI / test (push) Successful in 19s Details - Duration literals now accept s/m/h/d/w (was m/h/d/w): 'BEHAVES { ttl 30s }' is 30s, and 'within 30s' temporal expressions parse. - Definition fields are REQUIRED by default ('field: type'); 'field: type?' marks a field optional. addNode enforces presence on insert, getSchema exposes per-field requiredness, and updateNodeData still validates the provided fields' types. - Provider-returned edges are validated against the fact's declared typing: a value-carrying fact must return { value, possibility } with a value of the declared type (bare-number shorthand is rejected); a non-value fact must not carry a value; and every possibility must lie in [0, 1]. Violations throw a clear provider-authoring error instead of silently injecting malformed edges. Tests: DSLRuntimeTyping (duration units, required-field enforcement, schema requiredness, value-type + shape + possibility validation).	2026-08-03 16:00:49 -07:00
John Dvorak	aa38fbfd8c	feat: provider-cache bypass knob; reject non-final comparator chain steps CI / publish (push) Successful in 9s Details CI / test (push) Successful in 18s Details - check() gains cacheProviderResults (per-check option + policy default): the provider cache is a STORE-RETRIEVAL cache (wall-clock), independent of the caller's decision { now }; callers who pin time or want fresh retrieval can opt out per-check or globally. - The generator now rejects a comparator evidence referenced at a NON-final chain position (a comparator compares values at (src, candidate) but provides no candidate set, so it cannot enumerate intermediate nodes). Final-position comparators still lower to condition steps. Tests: cache bypass (per-check + policy), comparator final OK / intermediate error.	2026-08-03 15:43:59 -07:00
John Dvorak	ad365a65a9	feat: direct fact checks consult registered providers CI / test (push) Successful in 19s Details CI / publish (push) Successful in 9s Details Checking a FACT relation directly (rt.check('u:1', 'owns', 'doc:9')) now runs the provider-retrieval pipeline instead of returning 0 without ever consulting the registered provider. The retrieval set for a direct fact check is the fact itself (plus, for evidence checks, the injectable deps as before). The result reports requiredFacts/providedFacts/missingFacts for the fact, and missingFacts surfaces 'no_provider' when neither a provider nor an edge can satisfy the check.	2026-08-03 15:25:33 -07:00
John Dvorak	6214780244	feat: provider-result caching with time expiry + DSL fact-level TTL CI / publish (push) Successful in 9s Details CI / test (push) Successful in 18s Details Automatically-retrieved facts (balances, sessions, etc.) are now cached with a time expiry so repeated checks don't re-invoke the underlying data store. - check(): registered providers cache their normalized edges per (relation, subject, object) with a TTL; a fresh entry is reused without re-invoking the provider. Per-check factProviders are one-off observations and stay cache-transparent (no read, no write). - TTL resolution: DSL-declared 'BEHAVES { ttl <duration> }' on a fact > setFactTTL(relation, ms) > policy.providerCacheTTL (default 30s; 0 disables). - Grammar: facts may now declare a freshness window via 'fact balance(user, amount) BEHAVES { ttl 1h }' (previously BEHAVES only accepted 'AS edge\|transitive\|hierarchical\|symmetrical_graph'). The runtime indexes it as the fact's ttlMs. - registerFact/unregisterFact, graph mutations (add/update/remove node or relation), and invalidateProviderCache(relation?) keep the cache consistent. - Injectable clock (default wall clock) drives cache freshness, mirroring the core's unpinned-clock contract. Tests: DSLRuntimeCache (reuse within TTL, expiry re-invoke, per-check override transparency, registerFact invalidation, mutation invalidation, per-relation/all invalidation, policy default, DSL-declared 1h TTL).	2026-08-03 14:32:15 -07:00
John Dvorak	3ace783a59	feat: intermediate chain condition steps + _subjectIsObject unary scoping CI / publish (push) Successful in 10s Details CI / test (push) Successful in 19s Details - _expandChainSteps: a logical/defeasible evidence referenced by a chain step is now a condition step ({ rule, conditionStep }) at ANY position. As the FINAL step the engine verifies it at (intermediate, object); as an INTERMEDIATE step the engine expands it from the current node (rule-based reachability: base edges' destinations filtered by the rule's defeaters/requirements) and continues traversal from each discovered node. - buildPredicateRule / buildDirectRule: unary predicate calls whose subject entity IS the evidence's object parameter (trusted(other) inside peer_trusted(user, other)) are marked _subjectIsObject (was: only subject-var calls got _subjectAsObject). Requires @arbiter/core@^1.0.4. Tests: ChainConditionStep intermediate expansion; oracle campaign gains a chain_intermediate_condition construct (oracle = min(peer*(1-trusted), read)).	2026-08-03 13:36:16 -07:00
John Dvorak	fe162251fc	feat: DSLRuntime schema introspection, per-relation providers, retrieval loop CI / publish (push) Successful in 10s Details CI / test (push) Successful in 20s Details The higher-order DSL+Core wrapper now covers the full contract the DSL informs, beyond the typed mutations already present: - getSchema(): serializable introspection of the compiled type system — entity types/fields, facts (params + injectable flag), evidence (with transitive dependsOn), and registered providers. relationNames() lists all declared relations. (The DSL's type system was always present; this exposes it programmatically.) - registerFact(relation, fn) / unregisterFact / registeredFacts: per-relation async providers that retrieve missing partial-graph edges; per-check factProviders merge OVER registered ones. - Bounded fixed-point provider retrieval loop (maxProviderRounds): each round invokes providers for required facts whose edges are not yet injected. A provider may return edges for relations other than its own — those satisfy the other required facts and can unblock later rounds. - check() now type-validates FACT relations too (not just evidence); edge normalization preserves a provider edge's own relation name. - removeNode / removeRelation passthroughs; require() throws on denial for middleware. - Field typing extended to the DSL's full value-type universe (timestamp/duration accept number or string; object/any accept anything). Tests: DSLRuntimeExt (schema, registration, merge, fixed-point, require, removal, fact-check validation, timestamp typing).	2026-08-03 12:45:16 -07:00
John Dvorak	351551af0f	feat: chain condition steps — defeasible/logical evidence as final chain hop CI / publish (push) Successful in 9s Details CI / test (push) Successful in 18s Details A chain's FINAL (object-side) step may now reference a defeasible/logical evidence. The compiler lowers it to a condition step ({ rule: <config>, conditionStep: true }) that the engine verifies at (intermediate, object) instead of traversing an edge. Requires @arbiter/core@^1.0.3 (ChainRule condition-step support). - _expandChainSteps: a logical/defeasible/comparator evidence is expressible as a final condition step; non-final such steps remain a compile error (a condition cannot discover intermediate nodes). - Dependency collection (generator + DSLRuntime) descends into condition-step rule configs, so partial-graph requirements reach through them. Tests: ChainConditionStep (defeasible + ALWAYS steps, independent checkability, parallel aggregation), oracle campaign chain_condition_step construct (oracle = min(pm, pv*(1-pb))), DSLRuntime transitive required facts through a condition step.	2026-08-03 12:08:54 -07:00
John Dvorak	2dc478f5a3	feat: chain-step evidence composition — expand evidence steps in chains CI / test (push) Successful in 18s Details CI / publish (push) Successful in 9s Details A chain step that references a derived evidence is now expanded at compile time, keeping the engine a flat edge-traversal evaluator: - DIRECT evidence step -> renamed to its underlying relation (member_of(user,*g){ group_read(g,doc) } where group_read = can_view becomes step 'can_view'). - CHAIN evidence step -> its steps are spliced into the parent chain (a sub-path flattens into the linear source->...->object traversal). - Any other evidence type (defeasible/logical/comparator) as a step is a compile-time error: it is a condition, not an edge traversal. - Cycles and self-references through chain steps are compile-time errors (the existing composition cycle guard now covers steps). Rigor: oracle campaign gains a chain_step_composition construct; illegal mutations gain a non-lowerable-chain-step case. Fixture suites updated to retarget the self-recursive 'canRead/canAccess/...' terminals (an unsupported recursion pattern that now fails loudly) to an any-typed 'reachable' fact, preserving the nested-pattern parsing intent.	2026-08-03 11:34:52 -07:00
John Dvorak	88f10f9db4	feat: evidence composition — compile-time reference resolution for evidence sub-rules CI / publish (push) Successful in 11s Details CI / test (push) Successful in 15s Details An evidence may now reference another derived evidence as a sub-rule (WHEN can_read(user, doc) where can_read is itself an evidence). Resolution is a compile-time linker pass: after every evidence config is generated, each direct reference to an evidence is inlined with that evidence's own (resolved) config, so the engine evaluates a fully-resolved, acyclic config tree. - resolveEvidenceReferences(): post-generation pass over evidence configs, recursing into logical/defeasible containers (when/unless/never/always/ requires/union/intersection), always.direct nests, and comparator operands. - Forward references resolve (all configs exist before the pass runs). - Cycles and self-references are compile-time errors. - _subjectAsObject scoping is preserved through inlining. - dependsOn is recomputed after resolution, so partial-graph requirements reach transitively through composed evidence. - buildDirectRule/buildPredicateRule now apply subject-scoping to top-level PredicateCall evidence bodies (latent gap, previously missed). - validation: reject relation names shared across facts/sources/evidence/ measures (a collision silently overwrote configs and read as a false cycle). Tests: EvidenceComposition (9), DSLRuntime transitive requiredFacts, oracle campaign composition construct, illegal-mutations cycle + cross-kind cases.	2026-08-03 11:17:30 -07:00
John Dvorak	0a744329e6	feat: fix lowering, DSLRuntime wrapper, rigor oracle + rejection campaigns CI / publish (push) Successful in 10s Details CI / test (push) Successful in 13s Details Lowering fixes (validate/lower/compile into known-correct core structures): - tuple_to_userset: structural classification by object-side predicate (owner(*g, doc) { member_of(user, g) } -> tuple_to_userset with direction 'in'/'out'); the old heuristic routed every outer-wildcard to chain. - relational_comparator: operands now lower to real direct-rule configs (evaluateFrom derived from evidence param positions; expectedValue for literal args) instead of raw AST nodes the engine could not evaluate. - defeasible: multi-level bodies (NEVER/REQUIRES/ALWAYS/WHEN/UNLESS) merge into one five-level rule instead of ANDed level-only rules that always resolved 0; nested PatternMatches flatten to N-step chains; unary predicate calls mark _subjectAsObject (subject-as-object semantics). - validation: reject duplicate fact/evidence definitions. DSLRuntime (higher-order DSL+Core wrapper): - typed addNode/updateNodeData/addRelation/updateRelation against the DSL schema (known types, relation params, field types, value-carrying facts); - check() derives the evidence's injectable partial-graph requirements, retrieves missing facts through caller data callbacks, injects them, and delegates, returning requiredFacts/providedFacts/missingFacts. js-rigor campaigns: - generative oracle: generate legal DSL per construct and compare every verdict against an independent hand-computed oracle (8 constructs x P grid) plus an exhaustive deterministic sweep; - illegal mutations: one-flaw perturbations of a valid program must be reliably rejected (duplicate evidence/fact, arity/type mismatches, reserved built-ins, malformed syntax), with a control that must compile. Depends on @arbiter/core@^1.0.2 (reason codes + _subjectAsObject).	2026-08-03 10:58:29 -07:00
John Dvorak	0c2ddc282b	test: fix legacy fixture suites against modern DSL validation CI / test (push) Successful in 25s Details CI / publish (push) Has been skipped Details The legacy test files (Definition/Evidence/Expression/Fact/Integration/ Measure) predate the DSL's validation rules: reserved built-in types (User/Account/Device/AuthSession) and the boolean/predicate-call requirements on evidence statements. Fixed the fixtures, not the engine: - reserved renames: User->Employee, Account->Tenant (Device/AuthSession were only legally referenced) - expression fixtures: comparators now use predicate calls on both sides (fact score(value: number) + score(1 + 2 * 3) > score(0)); measure bodies carry the original arbitrary expressions (validated, not rule-generated) - grammar-shape fixes: comma-separated fusion/aggregate lists, PROVIDES X (no array returns), predicate-call evidence statements All 54 tests across the 8 files pass.	2026-08-03 09:12:20 -07:00