feat(claim-check): flag persistence requests by INTENT, not reply phrasing - #294
Merged
Conversation
…asing
Matching the reply was whack-a-mole. Across three live tests, the same request
produced three different lies, each needing a new pattern:
"I have ingested the 10-point instruction set." -> patterns v1
"The instruction set is active." -> patterns v2
"I've logged this in your persistent preferences." -> patterns v3
"Memorized." -> still missed
The model has unlimited ways to say "I saved it". The REQUEST side is small and
stable: "remember this for every future session" is a persistence ask however
the reply is worded. So if the user asked to persist and no persistence skill
ran, the turn is flagged regardless of phrasing — which cannot be evaded by
rewording, as the reply patterns provably could.
Verified against every phrasing the model actually produced, including the bare
"Understood." and "Got it! 🧡" that no reply pattern could ever catch.
Two suppressions keep it honest:
* a persistence skill really ran -> silent (that is the point of building it)
* the reply already admits it can't -> silent (never correct an honest answer)
False-positive discipline: "always use metric" is NOT treated as cross-session —
it usually means "in this conversation". Only explicit markers (future session,
from now on, always remember, standing rule) count. Ordinary asks — weather,
"remember to buy milk", summarise — stay silent.
Wired into both chat paths. The reply patterns remain as a second net for when
the model volunteers a claim unprompted.
Co-Authored-By: Claude Opus 4.8 <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
2 participants
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.
Matching the reply was whack-a-mole. Across three live tests the same request produced three different lies, each needing a new pattern:
The model has unlimited ways to say "I saved it". I was losing.
The fix: check the request, not the reply
The request side is small and stable. "Remember this for every future session" is a persistence ask however the model answers. So: if the user asked to persist and nothing persisted, the turn is flagged regardless of wording. That cannot be evaded by rephrasing — which the reply patterns provably could.
Verified against every phrasing the model actually produced, including bare "Understood." and "Got it! 🧡" that no reply pattern could ever catch.
Two suppressions keep it honest
standing_rules)False-positive discipline
"always use metric"is not treated as cross-session — it usually means "in this conversation". Only explicit markers count: future session, from now on, always remember, standing rule. Ordinary asks — weather, "remember to buy milk", "summarise this" — stay silent.Wired into both chat paths. The reply patterns stay as a second net for when the model volunteers a claim unprompted.
79 claim-check tests · 192 chat/claim/standing-rules tests pass.