Files
evidence-dsl/tests/rigor/dsl-generative-oracle.test.js
T
John Dvorak 88f10f9db4
CI / publish (push) Successful in 11s
CI / test (push) Successful in 15s
feat: evidence composition — compile-time reference resolution for evidence sub-rules
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

208 lines
8.9 KiB
JavaScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/**
* 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×(1defeat),
* 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 banned(user: Employee)
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;
}
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');
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'];
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.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 ${inv.failureCount} cases`);
});
it('exhaustive deterministic sweep: every construct x every possibility value', () => {
// Anti-vacuity complement to the campaign: sweep the full P × P grid per
// construct without any RNG, so a construct the campaign skipped would
// still be caught here.
for (const kind of CONSTRUCTS) {
for (const a of P) {
for (const b of P) {
const ps = kind === 'direct' ? [a] : [a, b];
const { dsl, edges, oracle, relation } = buildProgram(kind, ps);
const arbiter = new Arbiter();
arbiter.addNode('u:1', 'Employee');
arbiter.addNode('g:1', 'Group');
arbiter.addNode('doc:9', 'Doc');
const compiled = new DSLCompiler(arbiter).compile(dsl, 'sweep');
assert.ok(compiled.success, `${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}`
);
}
}
}
});
});