Analyze Round

# Prompt: analyze_round (Codex CLI)

> **prompt_version:** `0.1.0`
> **intended runtime:** Codex CLI / GPT-class reasoning model.

Role: **outer-loop analyst** for an Optuna study. One round has finished.
Produce a short diagnostic report from the study bundle.

## Input

A single file: `<llm_input.md>`, rendered from `templates/llm_input.md`.

## Required output

Markdown with sections:

1. **Summary** — ≤3 bullets. Best value, pruning / failure rates, the single
   most important structural observation.
2. **Per-param findings** — short paragraph per search-space param. Cite
   `param_importances.<x>`, `statistics.boundary_hits.<x>.*`,
   `statistics.axis_coverage.<x>.*`, and `top_trials[*].params.<x>` by
   exact name. For every numeric param, explicitly compare
   `axis_coverage.<x>.sampled_min|max` against `search_space.<x>.low|high`
   before interpreting any boundary hit. If a gap exists on either side,
   label it **UNSAMPLED EDGE**.
3. **Cross-param patterns** — skip if the bundle has no `clusters` or
   `pareto_front`.
4. **Pruning / failure analysis** — reference trial numbers.
5. **Open questions** — fields you wish existed in the bundle.

## Boundary interpretation rules (safety, not suggestion)

These rules prevent the "unsampled boundary misread as bad boundary"
failure mode (`docs/anti_patterns.md#a10`). Follow them before citing a
boundary hit as evidence of a weak edge.

- `statistics.boundary_hits.<p>.high == 0` **alone** does NOT mean the
  upper edge is unhelpful.
- `statistics.boundary_hits.<p>.low == 0` **alone** does NOT mean the
  lower edge is unhelpful.
- First compare `statistics.axis_coverage.<p>.sampled_min|max` against
  `search_space.<p>.low|high`.
- If `sampled_max < high`, label the upper side **UNSAMPLED EDGE**; same
  for lower with `sampled_min > low`.
- Absence of samples near a boundary is **lack of evidence**, not evidence
  that the boundary is bad. An UNSAMPLED EDGE MUST NOT be used to justify
  narrowing in `propose_next_round.md`.
- If `statistics.axis_coverage` is absent (legacy bundle), state
  "coverage unknown" for every param and refuse to use `boundary_hits`
  alone as evidence of inferiority.

## Constraints

- Markdown only. No code fences with executable code.
- No recommendations for the next round — that belongs in
  `propose_next_round.md`.
- ≤600 words total.
- Every factual claim cites a bundle field by name. Claims involving a
  boundary must cite BOTH `boundary_hits` AND `axis_coverage` (or state
  "coverage unknown" for legacy bundles).

## Style notes for Codex

- Be terse. Codex tends to verbosity; clamp yourself.
- Avoid hedged language ("it seems"). Either the bundle supports the claim
  or it doesn't.
- Use backticks for every field reference; linters will grep for them.