Eve library

PROOF-001 evidence pack — non-substitutable proof types

Evidence

Revision 1

On this page

Status: repository-complete / closed. The canonical contract represents fourteen proof meanings separately, and every covered release boundary requires an exact type, subject, policy, evidence link, status, and valid evaluation time. This status does not claim that live retrieval, rights, signer/KMS, C2PA, placement, consent, or outcome-observation services are deployed.

Work-package charter#

  • Task ID: PROOF-001
  • Owner: contracts, publication/run protocol, shared release gates, Isis, BFF generation, and downstream release-adapter maintainers; implementation owner: Elikem Agudogo
  • Accountable product: shared V1–V9 proof vocabulary and every invoked publication, durable-run, generation, and promotion boundary named below
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: Critical release-integrity, provenance, rights, consent, semantic-grounding, signature, and observed-outcome authority risk
  • Audit finding: P0-11 — release evidence could overstate what was proved
  • Dependencies: RUN-001 common run protocol, transactional release lifecycle, shared release gates, Isis canonical generation gates, BFF provider measurements, and downstream product adapters
  • Affected schemas and content slots: canonical ProofRecord, ProofRequirement, requirement verdicts, run gate verdicts, release reports, provider measurements, and adapter bindings. No rendered customer content, media bytes, prompt body, or UI content slot changed.
  • Linked task IDs: PROOF-001, PROOF-002, PROOF-003, RUN-001, REL-001, MEDIA-001, and AUTH-003
  • Rollout strategy: additive canonical proof records plus fail-closed exact type, subject, policy, evidence-link, status, and freshness checks at release boundaries
  • Rollback strategy: revert the owning source commit and its downstream adapter remediations together. Reverting only the contract would leave consumers compiled against a missing proof vocabulary.
  • Evidence-pack location: this file

Closure matrix#

Ledger requirement Repository closure proof
Represent every required proof meaning separately ProofRecordSchema is a strict discriminated union of exactly fourteen records: format_validity, identifier_existence, retrieval, passage_match, claim_entailment, source_approval, rights_license_receipt, consent, human_approval, byte_bound_integrity, authenticated_signature, c2pa_verification, runtime_placement, and observed_outcome. Every record has a type-specific strict payload.
Prevent weaker proofs from satisfying stronger gates evaluateProofRequirements accepts only an exact proof type, subject, and policy with passed status, non-future verification time, and an unexpired receipt. There is no strength hierarchy or implication rule. Promotion re-evaluates the actual proof rather than trusting a mutable prior verdict.
Name the exact passed proof type and evidence in release results Requirement verdicts carry the required type, passed type, matched proof ID, and evidence references. Publication bundles, durable run verdicts, shared release reports, Isis gate results, promotion records, and BFF release measurements expose or validate those records. Blocked and skipped gates do not claim a passed proof type.
Add the required adversarial fixtures Canonical fixtures reject citation-shaped strings, existing-but-irrelevant identifiers, unsupported passages, stale rights, locator hashes, unkeyed digests, structural quality proxies, future-dated proof, and forged promotion verdicts whose underlying proof was weakened.

Canonical proof contract#

libs/contracts/src/common/proof-types.ts owns the canonical vocabulary. Every record binds a proof ID, exact subject, status, evidence references, policy, producer, verification time, optional expiry, and a strict type-specific payload. Passed status must agree with that payload.

  • Citation syntax validation proves only format_validity.
  • Finding an identifier proves only identifier_existence.
  • Retrieval, passage match, and claim entailment are three separate records.
  • Source approval, an active rights/license receipt, consent, and human approval are four distinct authority records.
  • byte_bound_integrity requires an artifact_bytes binding and byte length.
  • authenticated_signature requires a supported algorithm, key ID, signer identity, signature reference, signed-payload digest, verification policy, and valid verification result.
  • C2PA verification separately binds its manifest, signer, trust policy, hard binding, validation report, and validation result.
  • Runtime placement and observed outcome do not imply one another.

Duplicate proof or requirement IDs are rejected. An accepted proof must be for the exact subject and policy, already exist at evaluation time, and not be expired. The evaluator returns a typed reason and evidence links for every satisfied, missing, or rejected requirement.

Current contract drift from the original implementation is additive and admitted: retrieval proof can bind the exact retrieved UTF-8 passage digest, and canonical constructors create source-approval and rights-license records bound to one exact source. The current tests exercise both constructors. The later operation-count run measurement is unrelated to proof substitution and does not change proof authority.

Invoked release boundaries#

Publication and durable runs#

The V1 publication gate accepts caller-supplied proof records and requires exact format, retrieval, passage, entailment, source-approval, rights, and human approval meanings. It never manufactures those records. Its decision and archived release bundle retain exact proof verdicts and evidence links.

RunGateVerdictSchema embeds proof requirements and proof records. A pending gate cannot carry resolved proofs; every resolved gate names its requirement; and a passing gate must satisfy every exact requirement and link accepted evidence. Transactional reruns preserve requirements while rebinding them to a replacement artifact. V1, V8, and V9 adapters emit honest format_validity, observed_outcome, or human_approval records rather than promoting structural state into a stronger meaning.

Shared gates, Isis, and BFF#

Every shared GateDefinition declares one exact requiredProofType. Structural validator/manifest adapters are capped at format_validity; score/evaluation adapters emit only observed_outcome; stronger semantic, rights, signature, C2PA, placement, or consent meanings require a canonical proof producer. Promotion checks the proof again against current time, artifact, type, policy, and linked evidence, and rejects report/proof/verdict tampering. Human promotion signoff produces a named human_approval receipt.

The consent registry emits artifact-scoped consent proof from concrete grants or an explicit no-subject finding. Isis maps each raw gate to one meaning: provenance to c2pa_verification, rights to rights_license_receipt, structural shape/disclosure to format_validity, human triggers to human_approval, and measured safety/watermark/quality/policy checks to observed_outcome. A positive boolean or metric cannot release an artifact without the corresponding proof record.

BFF provider measurements carry an exact subject, evaluation time, and proof set. The runtime validates each proof before Isis evaluation. Production providers currently pass empty proof sets where a real receipt producer is not integrated, so signed-, watermarked-, or score-shaped signals remain blocked instead of being upgraded into C2PA, rights, or outcome proof.

Downstream adapter authority#

Yemaya, Calliope, V3, V8, and V9 adapters preserve only the proof meaning their source actually owns. Metis planning and execution adapters now also inventory and bind V9 role_visibility, tenant_region, and the successful safety record as policy authority. The execution adapter requires the safety record's screened input hash to equal the executed request hash. Cross-tenant, late, unknown-role, unknown-trust-zone, empty-region, and request-mismatch records fail before common-run construction.

The Metis correction was necessary because the current OpenAPI schemas had added authority-bearing fields that the original adapters omitted. The inventory tests failed first; commit 9f82832200aba015187521d47c63d33c657c8894 closed the drift and added the adversarial checks. No inventory baseline was widened.

Adversarial fixture matrix#

Fixture Presented evidence Strong requirement Result
Citation-shaped string Schema/format validation Claim entailment Rejected: format_validity is non-substitutable
Existing irrelevant ID Registry existence Retrieval Rejected: existence is not retrieval
Unsupported passage Failed passage match Claim entailment Rejected: passage matching is distinct and failed
Stale rights Once-active expired license receipt Current rights Rejected as expired
Locator hash Digest of a locator-shaped value Byte-bound integrity Rejected: binding scope must be artifact_bytes
Unkeyed digest Byte checksum without signer/key/signature Authenticated signature Rejected by schema and exact-type policy
Structural quality proxy Format-valid structure Observed outcome Rejected: structure is not an observed outcome
Future-dated receipt Proof verified after evaluation time Any current requirement Rejected as future-dated
Forged promotion verdict Verdict claims signature after proof is swapped to format validation Authenticated signature Rejected by promotion-time proof re-evaluation
Cross-tenant V9 safety record Successful screening for a different tenant V9 execution authority Rejected before run construction
Mismatched screened input Successful screening for different request bytes V9 execution authority Rejected before run construction

Automated verification#

Current admitted refresh#

The refresh started from commit c5b9fccc9f07706a3e137661712d5861ae0e7d43. It then published an Isis test-source correction as 602de4516e8e19af1a7ca10824344c68141e59c3 and the Metis runtime/test correction as 9f82832200aba015187521d47c63d33c657c8894. Unaffected suite results remain admitted because neither remediation changed their source or test files. The final Metis row was rerun from the exact published origin/main commit.

The host was app-dev-server, Ubuntu 24.04.4 LTS, Linux 6.8.0-134-generic x86_64. The toolchain was Node.js v24.15.0, pnpm 10.25.0, TypeScript 6.0.3, and Vitest 4.1.7; pnpm-lock.yaml SHA-256 was 08b055786e9b8a6104d5313601d7eef3205dedcf395066fd9a8264c9859d690f. No customer data, production secret, live registry, signer, provider, or external evidence service was used.

UTC interval Exit Focused command/surface and result
2026-07-29T12:34:49.407Z–2026-07-29T12:34:53.196Z 0 pnpm exec vitest run --config libs/contracts/vitest.config.ts over proof-types.test.ts, publication-release-gate.test.ts, run-model.spec.ts, and transactional-intervention.spec.ts — 64/64
2026-07-29T12:35:00.846Z–2026-07-29T12:35:17.884Z 0 pnpm exec vitest run over shared release-gates.test.ts, consent-registry.test.ts, and portfolio-release-gates.test.ts — 47/47
2026-07-29T12:35:36.554Z–2026-07-29T12:35:42.772Z 0 BFF provider-measurement.test.ts, release-gate.test.ts, and isis-route.test.ts under the BFF Vitest config — 39/39
2026-07-29T12:36:31.858Z–2026-07-29T12:36:33.092Z 0 Yemaya common-run-adapter.test.ts under its package config — 4/4
2026-07-29T12:36:38.843Z–2026-07-29T12:36:43.967Z 0 Calliope V4 narrative-quality package — 15/15
2026-07-29T12:36:52.212Z–2026-07-29T12:36:59.796Z 0 V3 concert-quality package — 14/14
2026-07-29T12:37:05.774Z–2026-07-29T12:37:10.523Z 0 V8 case-gates package — 7/7
2026-07-29T12:37:19.793Z–2026-07-29T12:37:24.556Z 0 V9 lesson-gates package — 9/9
2026-07-29T12:37:31.727Z–2026-07-29T12:37:38.999Z 0 tests/integration/flows/v3-concert-export.integration.test.ts and v5-sidequest-generation.integration.test.ts — 3/3
2026-07-29T12:40:17.820Z–2026-07-29T12:40:28.943Z 0 full Isis package — 411/411
2026-07-29T12:42:48.218Z–2026-07-29T12:43:11.210Z 0 full shared content-quality-judge package — 285/285
2026-07-29T12:43:17.439Z–2026-07-29T12:43:46.804Z 0 full creative-autonomy package — 289/289
2026-07-29T13:05:59.190Z–2026-07-29T13:06:12.730Z 0 published Metis plan/execution adapters: 8/8, then exact five-file ESLint with --no-ignore, then tsc -p libs/metis/api-client/tsconfig.lib.json --noEmit

The admitted current value suites therefore passed 1,195/1,195 assertions across 94 files. Reporter duplication in the shared run was excluded from the 47 unique assertions.

Type, lint, and build gates#

UTC interval Result
2026-07-29T12:39:45.223Z–2026-07-29T12:40:10.984Z Isis library and spec TypeScript configurations passed after the test-source remediation.
2026-07-29T12:41:18.302Z–2026-07-29T12:41:34.618Z Yemaya library and spec configurations passed.
2026-07-29T12:41:43.960Z–2026-07-29T12:42:24.813Z Calliope library, V3 library, V8 package, and V9 package TypeScript configurations passed.
2026-07-29T12:43:57.532Z–2026-07-29T12:44:24.425Z Content-quality-judge and creative-autonomy TypeScript configurations passed.
2026-07-29T12:44:39.707Z–2026-07-29T12:45:05.642Z Contracts library, shared release gates, and Metis library TypeScript configurations passed.
2026-07-29T12:45:19.129Z–2026-07-29T12:45:48.356Z ESLint over the 26 explicitly enumerated PROOF source/test files passed with zero errors; ignored test files were included explicitly.
2026-07-29T12:59:04.073Z–2026-07-29T13:00:15.625Z Uncached targeted metis-api-client:build and its six build dependencies passed.

The publication hook for the Metis remediation also passed all 16 dependency-aware affected typecheck projects, including the BFF's official zero-new-error ratchet with 0 known app-owned backlog. The original PROOF-001 publication record retains its passing contracts and Isis builds and direct Node 22 BFF production bundle as historical source-publication provenance; they are not mislabeled as current reruns.

Refresh-discovered regressions and bounded exclusions#

The current audit preserved red gates and fixed task-owned defects:

  • Isis spec typechecking first found six unused hash imports and one unused test helper across seven tests. The same cleanup consolidated split value/type imports in the admin-view test. Commit 602de4516e8e19af1a7ca10824344c68141e59c3 corrected the test sources; both Isis TypeScript configurations and the full 411/411 suite then passed. No baseline changed.
  • The correct-root full Metis adapter run initially failed two schema inventory tests: planning omitted role_visibility and safety_record; execution omitted those fields and tenant_region. Commit 9f82832200aba015187521d47c63d33c657c8894 inventories, validates, maps, and tests that authority. The published package now passes 8/8, exact lint, library typecheck, and targeted build.
  • Metis's package-wide spec TypeScript configuration still reports unrelated Vitest 4 mock typing and unused-fixture debt outside the two adapter files. It is not represented as passing. The owning library typecheck, exact adapter runtime tests, exact test/source lint, build, and OpenAPI inventory assertions all pass.

No broad workspace sweep or AUD-003's separate 54-config refresh was attempted or claimed.

Artifact identity#

These hashes were computed from the final published source commit 9f82832200aba015187521d47c63d33c657c8894.

Artifact SHA-256
Canonical proof contract 26b83832df6f7e8063e0600e532ec23caefcc7291acf1f451ab8402e78e53d7a
Adversarial proof fixtures bc5cdf328bf4366b193c9a263e8a73812b7ee4c1f578b101ac5bd448be9c71c9
Shared release boundary 1ff1b6707541c8738325f125bf9fce60f4a5b6df473368c3cffe97316c729bcb
Durable run proof boundary dae7ad65c314979b47fc899be6ff74c94c3313c5fd53e2187444d51ad3190890
Isis canonical release model 9f0a93cf9442bbacfa28d3a692dd92e92f615daa53e7fcb1ae7b4a5d403d86d6
BFF release validator 28a57055ad8104d9d16c47bb25769eb78be705ebe7da8f6efe333b4ec58253f0
Metis planning authority adapter a77601b2663db817b68eceb29f584dd0c333d767e2e81e56888639e6c3060b7c
Metis execution authority adapter b59d335db5c5e2bc92fbf356573e5a1a7c066cd2f82d5fb5c8fe1133ce0bf5cf
Yemaya stage-proof adapter 4d3d9bcc82f68bc5af0212d33979b194f10e139b272fa47eeb89761766ee7fa5
Calliope skipped-proof report f653051288a03801bbf37840e70d36b91fc33d9e7488b44570271238c5eff068

Human authority and external deployment boundary#

Repository publication of the contract and this evidence pack requires no fabricated human signoff, and none is claimed. A runtime human_approval proof is valid only when it names a real decision, reviewer, subject, policy, time, and linked evidence. Fixtures exercise that schema and fail-closed behavior; they are not approvals.

PROOF-001 defines and invokes proof identity and non-substitution. It does not fabricate the real-world fact represented by a proof:

  • live retrieval, passage capture, and entailment producers remain governed by PROOF-002;
  • rights/consent registries, byte-bound integrity, authenticated signer/KMS, and C2PA receipt producers remain governed by PROOF-003 and deployment configuration;
  • observed_outcome proves that an identified observation was made under an identified methodology. It does not alone prove causality, semantic correctness, or external population validity;
  • production registries, signer keys, provider credentials, C2PA trust services, and population experiments are external-state inputs and are not claimed here.

Absent required proof, affected release paths fail closed.

Resource-safe verification and UI scope#

The refresh host reported 15 GiB RAM, 31 GiB swap, and 16 GiB free disk. An unrelated checkout's long-running Vitest worker and this checkout's Nx daemon were active. Focused suites ran sequentially. The Metis commit hook's repository-defined affected typecheck expanded to 16 projects at parallelism four; it was supervised, bottomed out at approximately 2.8 GiB available memory, and completed without reaching the 2 GiB stop threshold. No full workspace build/test/lint sweep was started.

This package changes contracts, backend release logic, adapters, tests, and evidence only. It does not change customer/admin/PWA/mobile/Unreal UI, so the frontend skill, Playwright, mobile automation, screenshots, and Unreal verification are inapplicable.

Publication identity#

  • Authoritative source commit: 73ee66c5077cd2ea00c7280e530ed0091072ea1b
  • Initial evidence-pack commit: 80d8a061f1fbd2bbdc4f3f8e858c85cfe61f2868
  • Historical declared closure evidence commit: cfb7cb334b9ebc3ce2ffc03dc2f583d3d21ec686
  • Equivalent rebased closure identity on the current origin/main graph: 20eb659405699e2fffb015ba5e630a201ccee117 (same tree 93d4fc21c327cb862f6c5c9ccdff6849fc0e2775)
  • Current Isis test-source remediation: 602de4516e8e19af1a7ca10824344c68141e59c3
  • Current Metis source/test remediation and admitted verification tip: 9f82832200aba015187521d47c63d33c657c8894
  • Current evidence-refresh commit: recorded by the task-scoped commit that contains this file.