Status: the ordinary V9 lesson-release path now requires exact retrieved passage bytes plus semantic claim entailment; production-approved source bundles and calibrated live evaluator evidence remain absent.
Date: 2026-07-27
Gap closed by this increment#
The canonical semantic-grounding envelope already required retrieval,
passage-match, and claim-entailment proof types, and the ordinary
ContentGenerator.generateLesson() path already routed through that evaluator.
The upstream retrieval record, however, bound only a source and passage
reference. A caller could retain those identifiers, substitute the passage text,
and let the V9 adapter mint a new lesson-scoped retrieval record over the
substituted text.
This increment closes that evidence-laundering path:
- canonical textual retrieval records carry the SHA-256 of the exact UTF-8 passage bytes;
- every current V1, V2, V6, V7, and V9 semantic-grounding producer supplies those exact bytes when recording retrieval;
- the shared evaluator recomputes the digest from the bounded passage and rejects reference-preserving text substitution;
- V9 validates the upstream proof digest before any provider call or derived lesson-scoped proof is created; and
- the ordinary lesson path still requires a configured semantic entailment evaluator and complete retrieval, passage-match, and claim-entailment coverage for every extracted factual claim.
The adjacent tracing repair aligns @oshun/tracing with the canonical RUN-001
adapter vocabulary by accepting v7-flagship-revision as V7. This removes the
shared type regression that previously blocked Metis, V9, BFF, and
prompt-boundary typechecks; it does not broaden or relabel adapter authority.
Adversarial proof#
The contract regression test retains the source ID, passage reference, and
original retrieval proof while replacing the passage text. Semantic grounding
fails with lacks a valid exact-passage retrieval proof.
The V9 generator regression performs the same substitution in an otherwise
approved source bundle. generateVerifiedLesson() rejects it before the model
is called, proving the adapter cannot launder the changed bytes into a fresh
proof. Existing tests continue to require all three proof classes and reject
invented citations, uncited factual claims, missing semantic evaluators, and
non-entailing passages.
Verification#
Verification ran sequentially with one Vitest worker on the 15 GiB host.
Contracts, focused
proof types + semantic grounding + publication release: 34/34 passed
Metis verification
focused citation/integration: 14/14 passed
complete package: 24 files, 176/176 passed
Metis LLM client
focused ordinary/verified lesson boundary: 13/13 passed
complete package: 26 files, 1,323/1,323 passed
Downstream exact-grounding consumers
Iris V2 commentary: 7/7 passed
Isis V7 Mawu: 23/23 passed
Clio V6 narration: 7/7 passed
V9 lesson gates: 9/9 passed
V8/V9 prompt-injection boundary: 41/41 passed
V1 BFF autonomy binding: 40/40 passed
Shared tracing
focused GenAI operations: 7/7 passed
TypeScript
contracts, tracing, Metis verification, Metis LLM client, Iris conversation,
Isis 3D pipelines, Clio story, V9 lesson gates, prompt-injection evaluations,
and the BFF autonomy boundary passed.
Targeted ESLint passed with zero errors; the Metis generator retains 38
pre-existing warnings outside the changed lines. Prettier and git diff --check
passed. No UI changed, so Playwright, mobile automation, and the frontend
workflow are not applicable. No provider, database, media, or Unreal execution
is claimed.
Remaining completion boundary#
The compound V9-001 checklist item remains open because the repository does not contain production-approved source bundles. No live provider output, licensed source acquisition, calibrated production entailment report, learner delivery, or full source-to-rendered-experience run was available. The implementation now makes exact retrieval bytes and semantic entailment non-substitutable; it does not convert fixtures or repository sources into production approval evidence.