feat(prompts): treat orchestrator-authored digests as hypotheses, not evidence - #32
Open
stephanbrez wants to merge 1 commit into
Open
Conversation
… evidence A real mission showed digest paraphrase manufacturing absolutes (only, every, never, all) that subagents never verified; three wrong claims propagated onto the deliverable before the accuracy lane caught them. Runtime validation cannot catch free-text drift by construction, so the rule lives in the prompt: file:line-backed claims, no unverified absolutes, a standing verify-before-use header, source-wins on disagreement, and immediate fold-in of worker corrections. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a rule to the orchestrator prompt's Memory Folder section governing research/facts artifacts the orchestrator writes from subagent investigation reports.
In a real end-to-end Zenith mission (63 contract assertions, 18 tasks, ~7h wall clock), the orchestrator wrote a citation-backed source digest from parallel investigation subagents, and workers built the deliverable from it. Four of its factual claims were wrong, and three propagated onto the delivered artifact before the accuracy validation lane caught them. Every error was a manufactured absolute (only, every, never, all) that no subagent had verified — digest paraphrase sharpens hedged findings into confident claims, and downstream agents treat the digest as authoritative because it is citation-shaped.
No validation lane catches this by construction: content lanes judge whether a point is explained, and accuracy lanes audit the deliverable against source, not the intermediate digest. Since the runtime cannot structurally validate free text, the fix belongs in the prompt.
The added rule requires: (a) each digest claim backed by
file:lineor command output, (b) no absolutes a subagent did not explicitly verify — prefer the source's own wording over paraphrase, (c) a standing verify-before-use warning at the head of the file, source-wins on any disagreement, and immediate fold-in of worker corrections. In the mission above, workers did correct the digest twice unprompted where the project's AGENTS.md carried (c)-style guidance — that mechanism works; what was missing was the rule against manufacturing absolutes in the first place.Verification
uv run pytestinzenith/: 240 passed, 7 skipped (suite as of this branch point).🤖 Generated with Claude Code