optimizing-skills retro — experiment archive

Archived May 29, 2026

Supporting data for the blog post The validation gate would have rejected the fix that worked. Two things ran: a controlled re-validation of the down-skilling v1.2.0 edit through the optimizing-skills gate, then a gate run validating the patch that became optimizing-skills v0.2.0.

Read in this order

• RESULTS.md — the controlled run: best (v1.1.0) vs candidate (v1.2.0), method, the 0/5-vs-3/5 invention result, and what the run exposed about both skills.
• scores.json — per-run scores with rationale (invention + length-in-range, both arms).
• VALIDATION.md — the gate run that validated the v0.2.0 patch against optimizing-skills' own gate (4 decision scenarios; per-criterion rule 4/4 vs collapsed 3/4).
• proposed-patch-optimizing-skills.md — the two edits, written before validation.

The two arms

• skill-versions/down-skilling-v1.1.0.SKILL.md — best: the four v1.2.0 edits reverse-applied.
• skill-versions/down-skilling-v1.2.0.SKILL.md — candidate: as shipped.
• arms/best/haiku_prompt.md + run_01..05 — the prompt a Sonnet author compiled from v1.1.0, and the five Haiku outputs.
• arms/candidate/haiku_prompt.md + run_01..05 — same, compiled from v1.2.0.

Files are markdown / JSON served raw. Source experiment lives in a private repo; this is the published copy.