Repository navigation
fix(leads): count the ask as grounding, not as a fabrication to reject - #148
Merged
Merged
Conversation
A live campaign rejected every draft it produced, five per tick, for hours. Its ask read "download our free whitepaper at https://threatcrush.com/get-whitepaper"; the model included that URL exactly as instructed, and the guard compared the body against the facts list alone and scored following the instruction as an invented link. The grounding set was wrong, not the drafts. Everything the operator authored is grounded — the intro says who is writing, the ask says what to ask for, and an ask naming a URL is an instruction to include it. Checking output against a subset of the instructions makes correct output unrepresentable. Kept: a link or number appearing nowhere the operator wrote is still refused, which is the case the guard exists for. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
vu1nz Security Review0 finding(s) in PR #? No security issues found. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
A live campaign rejected every draft it produced — five per tick, for hours:
Its
pitch_askreads "download our free whitepaper at https://threatcrush.com/get-whitepaper". The model included that URL exactly as instructed. The guard compared the body againstpitch_factsalone — which held only a 30-years-experience line — and scored following the instruction as an invented link.The grounding set was wrong, not the drafts
Everything the operator authored is grounded: the intro says who is writing, the ask says what to ask for. An ask naming a URL is an instruction to include it. Checking output against a subset of the instructions makes correct output unrepresentable.
Now checks against facts + intro + ask.
Still refuses what it exists to refuse
A link or number appearing nowhere the operator wrote is still rejected. Both cases are tested with the real campaign's values:
threatcrush.com/get-whitepaper(named in the ask)not-ours.test/landing(named nowhere)Plus a regression test pinning the old behaviour as the bug.
Checks
tsc --noEmitclean🤖 Generated with Claude Code