Uh oh!
There was an error while loading. Please reload this page.
Positive framing, consistent word counts, and stronger number-evidence constraint - #475
Closed
neoneye wants to merge 1 commit into
Closed
Positive framing, consistent word counts, and stronger number-evidence constraint#475neoneye wants to merge 1 commit into
neoneye wants to merge 1 commit into
Conversation
…e constraint
Replace negative prohibition ("Do NOT include 'Controls ... vs.'") with
positive framing. Align all three length targets (field desc ×2, system
prompt) to "2-3 sentences" — PR #473 only updated field descriptions,
leaving system prompt at "2-4" which haiku followed instead.
Strengthen number-evidence constraint: "Never invent percentages, costs,
or timeframes — only cite numbers that appear in the project context."
Placed before the cause-effect directive to prevent models from
fabricating specifics to satisfy cause-effect framing.
Reduce options to "one sentence" and review_lever to "1-2 sentences"
for conciseness, helping gpt-oss-20b stay within 600s plan budget.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>neoneye
commented
Mar 31, 2026
MemberAuthor
Self-improve iteration — analysis 71Verdict: CONDITIONAL Wins:
Regressions:
Need: fix system prompt consistency + rethink the number-evidence constraint wording. |
neoneye
commented
Mar 31, 2026
MemberAuthor
Closing — fabricated numbers 4x worse (HKD 5→20, $ 6→28). Need to rethink number-evidence constraint and fix system prompt consistency. |
neoneye added a commit
that referenced
this pull request
Mar 31, 2026
…gets Replace number-focused constraints with qualitative framing: "Exact numbers will be determined further downstream — here the goal is to articulate the cause-effect relationship clearly." This avoids the permission-floor effect where models derive numbers by arithmetic from real plan figures (4x HKD regression in PR #475). Section 5 prohibition changed from "NO fabricated statistics without evidence" to "NO specific numbers, percentages, or monetary amounts — describe effects qualitatively." All length targets consistent: consequences "2-3 sentences" in both field descriptions and system prompt. review_lever "one sentence (20-40 words)" in both field description and system prompt section 6. Replace negative prohibition ("Do NOT include 'Controls ... vs.'") with positive framing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5 tasks
neoneye added a commit
that referenced
this pull request
Mar 31, 2026
Replace number constraints with: "Use numbers only when the project context provides them directly — do not calculate, derive, or estimate figures." This closes the arithmetic-derivation loophole (HK$470M → HK$141M) from PR #475 while still allowing models to cite real plan numbers verbatim. Replace negative prohibition ("Do NOT include 'Controls ... vs.'") with positive framing. Tighten targets: consequences 2-3 sentences, options one sentence, review_lever one sentence (20-40 words). All sources (field descriptions and system prompt) consistent. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This was referenced Mar 31, 2026
neoneye added a commit
that referenced
this pull request
Apr 1, 2026
Document proven failure modes and fixes from 10 self_improve iterations: - Field-description template lock: expanded with evidence from PR #484 (the fix) and PR #479 (replacement phrase also locks). Strip to content-type + word count, let examples teach. - Negative prohibitions activate banned patterns: confirmed across PRs #458, #460, #475. Use positive framing instead. - Consequence parroting in later calls: llama3.1 copies consequences into review_lever. Anti-parrot needed in subsequent-call prompt. - Field descriptions vs system prompt consistency: haiku follows system prompt, text-completion models follow field descriptions. - Stripping field descriptions too far: weak models need "one sentence" and section pointer as structural anchors. Co-Authored-By: Claude Opus 4.6 (1M context) <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 freeto 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
Supersedes #471 and #473.
Test plan
🤖 Generated with Claude Code