Fix delegate_editor: writes-vs-claim verification guard (raffle-app fabricated-completion incident) by allocsys · Pull Request #154 · allocsys/madmcp

export extractMechanicalClaims so editor_delegate.js can reuse it without duplicating the claim-extraction logic

allocsys

committed

2cad16b

Import claim-verification helpers from agent_delegate.js for the new writes-vs-claim guard

allocsys

committed

01813f9

Add EDITOR_VERIFICATION_PROMPT + writes-vs-claim guard builder (fix for the fabricated-completion-report failure mode)

allocsys

committed

4f0a01c

Add pendingVerification state + restore from checkpoint (single-fire guard, same pattern as agent_delegate.js)

allocsys

committed

1179207

Restore pendingVerification from checkpoint on resume

allocsys

committed

33e769f

Persist pendingVerification via saveState so it survives a resume

allocsys

committed

a954e6a

Track fallbackModelUsed across the run so it can be surfaced on the final result, not just buried in the transcript

allocsys

committed

1637fb8

Restore and persist fallbackModelUsed across resumes

allocsys

committed

060662a

Capture fallbackModelUsed whenever a step is served by a Gemini fallback model

allocsys

committed

de99f05

Wire in the writes-vs-claim verification pass before trusting a draft final answer, and surface fallbackModelUsed on the persisted/returned result

allocsys

committed

eb0a1e5

Surface fallbackModelUsed on the already-done checkpoint short-circuit too, for a polling caller

allocsys

committed

5cd1dcd

Update guardrail #8 test for the new writes-vs-claim verification pass -- a successful draft answer now gets one extra verification round-trip (tools re-enabled) before being trusted, same pattern …

allocsys

committed

15310ae

Account for the new writes-vs-claim verification round in the fresh-synchronous-run checkpoint test

allocsys

committed

1051871

Add direct coverage for the writes-vs-claim verification guard: catches a fabricated completion report (the raffle-app incident's exact failure shape), lets the model self-correct by actually writi…

allocsys

committed

4cc6df5

Make verification prompt tool-access-aware; add looksLikeCompletionClaim heuristic

allocsys

committed

1d4e814

Cover the final-step/stuck-loop case the writes-vs-claim guard used to skip entirely

allocsys

committed

19a5e4a

Update single-fire test: a repeated zero-write completion claim is now flagged, not silently accepted

allocsys

committed

85d84d7

Navigation Menu

Fix delegate_editor: writes-vs-claim verification guard (raffle-app fabricated-completion incident) - #154

Fix delegate_editor: writes-vs-claim verification guard (raffle-app fabricated-completion incident)#154
allocsys merged 17 commits into
mainfrom
fix/editor-writes-vs-claim-guard

Commits on Sep 7, 2026

export extractMechanicalClaims so editor_delegate.js can reuse it without duplicating the claim-extraction logic

Import claim-verification helpers from agent_delegate.js for the new writes-vs-claim guard

Add EDITOR_VERIFICATION_PROMPT + writes-vs-claim guard builder (fix for the fabricated-completion-report failure mode)

Add pendingVerification state + restore from checkpoint (single-fire guard, same pattern as agent_delegate.js)

Restore pendingVerification from checkpoint on resume

Persist pendingVerification via saveState so it survives a resume

Track fallbackModelUsed across the run so it can be surfaced on the final result, not just buried in the transcript

Restore and persist fallbackModelUsed across resumes

Capture fallbackModelUsed whenever a step is served by a Gemini fallback model

Wire in the writes-vs-claim verification pass before trusting a draft final answer, and surface fallbackModelUsed on the persisted/returned result

Surface fallbackModelUsed on the already-done checkpoint short-circuit too, for a polling caller

Update guardrail #8 test for the new writes-vs-claim verification pass -- a successful draft answer now gets one extra verification round-trip (tools re-enabled) before being trusted, same pattern …

Account for the new writes-vs-claim verification round in the fresh-synchronous-run checkpoint test

Add direct coverage for the writes-vs-claim verification guard: catches a fabricated completion report (the raffle-app incident's exact failure shape), lets the model self-correct by actually writi…

Make verification prompt tool-access-aware; add looksLikeCompletionClaim heuristic

Cover the final-step/stuck-loop case the writes-vs-claim guard used to skip entirely

Update single-fire test: a repeated zero-write completion claim is now flagged, not silently accepted