Skip to content

Remove template-lock anchor from review_lever - #484

Merged
neoneye merged 1 commit into
mainfrom
fix/identify-minimal-template-lock
Apr 1, 2026
Merged

Remove template-lock anchor from review_lever#484
neoneye merged 1 commit into
mainfrom
fix/identify-minimal-template-lock

Conversation

@neoneye

Copy link
Copy Markdown
Member

Summary

Two lines changed. Remove "the specific gap the three options leave unaddressed" from:

  1. review_lever Pydantic field description (line 127)
  2. System prompt section 4 preamble (line 244)

This phrase caused 85% haiku template lock ("All three options X, but none address Y"). Proven fix from analysis 76 (PR #482). Everything else unchanged from baseline.

Test plan

  • Verify haiku template lock drops from ~85%
  • Verify no new template lock in qwen3-30b or gpt-oss-20b
  • Verify no content regressions
  • GPT5.4 assessment confirms better than analysis 40

🤖 Generated with Claude Code

Remove "identify the primary trade-off this lever introduces, then
state the specific gap the three options leave unaddressed" from two
locations: the Pydantic field description and the system prompt
section 4 preamble. Replace with minimal word-count guidance.
This phrase caused 85% template lock in haiku reviews ("All three
options X, but none address Y"). Proven fix from analysis 76 (#482).
All other field descriptions, system prompt sections, and prohibitions
unchanged from baseline (PR #358).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@neoneye

Copy link
Copy Markdown
MemberAuthor

Self-improve iteration — analysis 78

Verdict: YES

Template lock fully eliminated:

  • haiku: 24.7% → 0% (0/99 reviews across all 5 plans)
  • No new template lock in any model
  • Zero regressions across all 7 models

Secondary wins:

  • Haiku partial recovery: 2 events → 0
  • Haiku lever count: silo +50% (14→21), gta_game +31% (16→21)
  • Haiku call success: 86.7% → 100%
  • Reviews now domain-specific and structurally diverse

Two lines changed, zero regressions. The minimal approach worked.

@neoneye
neoneye merged commit 9012854 into mainApr 1, 2026
3 checks passed
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>
@neoneye
neoneye deleted the fix/identify-minimal-template-lock branch April 1, 2026 23:06
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

@neoneye