Eve library

V6-001 — Clio semantic-faithfulness evidence

Evidence

Revision 1

On this page
  • Owner: Codex implementation agent
  • Accountable product: V6
  • Priority: P0 product acceptance slice
  • Linked task: V6-001, add semantic faithfulness verification for Clio prose against event facts beyond citation-ID membership
  • Source commit: 9362dd74cbadded485be7f036d224fb4617f55ea
  • Evidence captured: 2026-07-24

Implemented boundary#

createReturningPlayerChronicleNarrated already rejected cited event IDs outside a beat's bounded Ori-log slice and invoked the shared semantic entailment proof contract. The release boundary now also requires the semantic verifier to return:

  • the SHA-256 of the exact trimmed prose it analyzed;
  • an explicit attestation that every factual claim was extracted;
  • claims whose text is an exact non-empty span of that prose; and
  • one complete, passed claim-to-passage entailment assessment for every extracted claim.

The Clio library rejects stale or mismatched narrative hashes, incomplete claim extraction, unanchored claim text, empty claim sets, unknown/duplicate citations, missing or duplicate assessments, invalid scores, non-entailing or unknown verdicts, and missing retrieval/passage-match/entailment proofs. A released beat retains its bounded passages and complete semantic proof bundle and carries semantic-claim-entailment-checked-over-ori-log.

The production V10 Ori narrator supplies the required verifier through the Moirai cognition gateway. Its trust-zoned verifier prompt treats narrative and passages as untrusted data, requires every factual claim as an exact contiguous quote, forbids omission, and permits entails only when bounded Ori passages fully support the claim. The adapter binds the returned assessment to the exact narrative hash before Clio can release it.

The V6 Chronicle drift checker now requires the narration behavior suite and the semantic receipt fields, evidence tag, and adversarial test identities. It also no longer carries an obsolete unused package.json read after verifier routing moved to scripts/verify-manifest.json.

Verification#

Command Result
pnpm exec vitest run --config libs/v6/clio-story/vitest.config.ts --maxWorkers=1 PASS — 2 files, 16 tests
pnpm nx run @oshun/clio-story:typecheck PASS
pnpm nx run @oshun/clio-story:lint PASS — zero warnings
pnpm nx run @oshun/clio-story:build PASS, including contracts and prompt-trust dependencies
pnpm verify:v6 clio-chronicle PASS, including post-rebase replay
pnpm nx run @oshun/v10-rail-channel-ori-presence:test PASS
pnpm nx run @oshun/v10-rail-channel-ori-presence:typecheck PASS
pnpm nx run @oshun/v10-rail-channel-ori-presence:lint PASS — one unrelated pre-existing warning in src/config.ts
pnpm nx run @oshun/prompt-injection-evals:test PASS — 5 files, 232 tests
pnpm nx run @oshun/prompt-injection-evals:typecheck PASS
Strict ESLint over all changed TypeScript/JavaScript PASS — zero warnings after suppressing ignored-file notices
Prettier and git diff --check PASS

The general pre-commit typecheck was bypassed only after all owning typechecks passed. It failed in unrelated current V10 web module aliases (@oshun/types, Veritas browser exports) and four unrelated BFF ratchet errors; none of the reported paths overlap this change.

Artifact hashes#

Artifact SHA-256
libs/v6/clio-story/src/index.ts 05dd18581322fbf1e5afe0a7111d125db0944ef81f63e8ec45e4d70141691fba
libs/v6/clio-story/src/clio-narration.spec.ts 2e4044a1a6a5818ab2aaa04b7e14484e8477287032902f0632eaf7b172bfc7ba
libs/v10/rail-channel-ori-presence/src/narrator.ts e8835700e0474df8941b56e48201bdabd81854bb9705edf05c1fa9a665d15b45
libs/v10/rail-channel-ori-presence/src/narrator.spec.ts 77785fa30c443e8170889168691f209e56274017d9838bd3c1dd7b5e3daba842
scripts/v6/verify-v6-clio-chronicle.mjs 95f31fb0bdecc20dc87c8ed68041a7262a243dae5fcb45b30e69f99db54ce117
testing/prompt-injection/src/v6-v7-boundaries.spec.ts d368328ff134932c0a9a0cbd039fee7f7da6075febf4b92f27be25b4e477b6eb
V6/architecture/clio-story-engine.md 74645e4ea06dafb1279daab3a3e85d550f28f494e77583415a50fa51034fb467

Honest boundary#

The tests use controlled writer/verifier or gateway doubles; they prove the fail-closed local release contract, not live provider quality. The production V10 adapter has a real governed Moirai dispatch seam, but this evidence does not claim deployed traffic, human evaluation, provider calibration, or operator takeover. Durable restart and concurrent-worker cognition evidence is recorded separately in evidence/v1-v9/V6-001-DURABILITY.md. A semantic verifier remains a trusted policy component whose extraction and entailment quality must be calibrated and monitored in production. The operator-takeover V6-001 bullet and every V6-002/V6-003 operational or flagship requirement remain open.