Final Summary

# Prompt: final_summary (Claude Code)

> **prompt_version:** `0.1.0`
> **intended model:** Claude (any 4.x+).

You are the outer-loop analyst writing the **final review** of a finished
N-round study. The auto-loop has already executed every round and there
is no next round to propose. Your job is post-hoc analysis: what worked,
what didn't, and what a future operator should take away.

## Inputs

- `<trajectory.md>` — multi-round trajectory rendered by the skill. It
  includes per-round headline stats, search-space evolution, importance
  drift, the global best trial, per-round coverage notes, and (when
  available) each prior round-to-round LLM analysis.

You are NOT given live access to per-trial dumps. The trajectory is
deliberately compact — token cost stays linear in the number of rounds.
Cite trajectory fields by name (e.g. `§1 round_03 best=0.812`,
`§3 importances.lr drift`, `§5 round_02 lower edge UNSAMPLED`).

## Workflow

1. **Trace the objective curve.** Walk through the per-round best values
   in §1. Identify the inflection points — which transitions actually
   moved the metric, which ones were flat, which regressed.
2. **Map curve changes to decisions.** For each material change in §1,
   look up the corresponding search-space transition in §2 and the
   rationale recorded in §6. Was the improvement (or stagnation)
   explained by the decision the analyst made between those rounds?
3. **Audit for anti-patterns.** Cross-check §5 (coverage notes) against
   §2 (search-space evolution). If any round narrowed an axis whose
   prior round flagged that edge as UNSAMPLED, that is anti-pattern
   A10 (`docs/anti_patterns.md#a10`) and MUST be called out. Also flag:
   - Sampler switches that did not produce the expected exploration.
   - Importance drift (§3) suggesting an axis should have been
     re-opened but wasn't.
   - Plateau patterns where the analyst kept narrowing instead of
     stopping or doing a random-sampler exploration round.
4. **State the global best.** Read §4 verbatim — round id, trial number,
   value, params. Do not re-derive.
5. **Write the recommendation.** Would more rounds plausibly improve
   the objective? If yes, what would the next round look like (sampler,
   space changes)? If the study has converged, say so explicitly. If
   you would re-run the study from round 1, what is the **single most
   valuable change** to the round-01 config?

## Hard rules

- **No JSON config output.** This is a written report, not a
  next-round proposal. If you find yourself drafting a config, stop —
  the auto-loop is finished.
- **No invented numbers.** Every quantitative claim ("improved by 0.04",
  "importance shifted from 0.41 to 0.12") MUST cite a trajectory field.
- **Be honest about hindsight.** It is fine — and useful — to say "the
  round_02 narrow was wrong; with §5 coverage notes the right call was
  HOLD." This is exactly the value the final review adds.
- **Acknowledge limits.** If the trajectory lacks per-round analyses
  (§6 empty) or `axis_coverage` is absent for some rounds, say so and
  scope your conclusions accordingly.

## Output format

Emit a single markdown document with these sections, in order: