Experiment Hypothesize Skill (Orbit Wars)

Per-case hypothesis-list builder, run once per family / case. The single deliverable is hypotheses.md. All downstream skills (experiment-plan / experiment-execution / experiment-analysis / experiment) consume this file as input.

When this skill is in charge

• The user typed /experiment-hypothesize (explicit trigger), or
• The user asked, in plain language, for any of:
  • enumerate hypotheses for a given case
  • brainstorm follow-ups around a knob (e.g. dropout)
  • organize an investigation plan before opening a new case
  • draft a hypotheses.md
• The user did not ask to plan / implement / run a single hypothesis. Single-shot requests go to experiment-plan / experiment-execution directly.

If hypotheses.md already exists for the target case, ask whether to rebuild / append / read-only display before proceeding.

Scope boundaries (what this skill does NOT do)

• Does not write iterN_plan.md or any per-hypothesis plan file. That is experiment-plan's job.
• Does not edit code under backend/pipeline/.
• Does not commit or push.
• Does not launch RunPod or run any dev/* command.
• Does not iterate. The skill writes hypotheses.md once and exits.

Skill flow

5 phases. Phase 1 = code reality check, Phase 2 = opt-in web research, Phase 3 = hypothesis hearing, Phase 4 = skip list hearing, Phase 5 = write hypotheses.md.

Phase 0 — Confirm the target case

AskUserQuestion, 1 question:

Q	options
対象 case	imitation/case<既存> / rulebase/case<既存> / reinforce/case<既存> / 新規 case を切る / Other (free-text family/caseN/topic)

Description on the first question: 各質問で「Other」を選ぶと自由記述も可能です。

For a new case, propose the next free number from ls backend/pipeline/<family>/. Resolve {topic} (snake_case 1–3 words) in the same question (Other recommended).

If a hypotheses.md already exists at docs/experiment/<family>/{yyyymmdd}_case{N}_{topic}/hypotheses.md, follow up with one more AskUserQuestion: 再構築 (back up existing and reset) / 追記 (append new hypotheses only) / 読み直すだけ (display current contents and exit).

Phase 1 — Code investigation (read-only)

Read the relevant code so the upcoming hearing is concrete. No user interaction.

Mandatory reads (skip those that don't apply):

• ls backend/pipeline/<family>/ — confirm existing case numbers and the next free slot.
• For an existing case: backend/pipeline/<family>/case<N>/ — main.py, agent.py, primary policy / feature modules.
• ls docs/experiment/<family>/ — surface prior experiment directories. Read 1–2 recent result.md / iter*_result.md files to capture what has already been tried + adopted / rejected.
• backend/src/dataset/selfplay/agents.py — AGENT_REGISTRY (relevant when registering a new case).

Surface a 3–5 line orientation report to the user (in Japanese):

pipeline/<family>/case<N>/ の現状: X / Y / Z。過去 iter で採用済み: A、却下済み: B。次の調査余地: C / D。

Phase 2 — Web research (opt-in)

AskUserQuestion, 1 question:

Q	options
外部 web 調査	不要 — Phase 1 のコード調査で十分 / 必要 — Other で具体的なクエリを指定 / 必要 — トピックを skill 側で提案

If the user picks "不要", skip to Phase 3.

If "必要", run WebSearch / WebFetch for 2–4 queries (time-box: ≤ 5 minutes). Bias toward authoritative sources (Kaggle writeups, arXiv, official docs, well-cited GitHub). Capture findings as 1–3 bullet points per source — these will be folded into Phase 5's References section. If the topic warrants a deeper survey, recommend /research-retrieval and stop.

Phase 3 — Hypothesis hearing

AskUserQuestion runs in at most 2 rounds (max 4 questions per round). Selection format with free-text via "Other".

Round 1: Draft, edit, prioritize

Generate 5–10 hypothesis drafts from Phase 1 (code) + Phase 2 (web) findings, present them, then ask:

Q	options
草案のうち採用するもの (multi-select)	H1: ... / H2: ... / ... / Other (add new hypothesis as free-text)
削除する仮説 (multi-select、任意)	H3 を消す / H5 を消す / ... / なし
優先順位の付け方	skill 提案順 (新規性 + 工数低い順) ⭐推薦 / ユーザが手動で並び替え / 期待効果順 / Other

Round 2: Shared metadata for the selected hypotheses

Q	options
最大 iteration 数 (この hypotheses.md 全体)	仮説リストを使い切るまで ⭐推薦 / 5 / 10 / Other
主要メトリクス (default)	local self-play win rate vs. {opponent} ⭐推薦 / loss curve のみ / behavior 指標 (例: 序盤 neutral 確保率) / Other
既定 episode 数	≥300 ⭐推薦 (memory project_imitation_case1_phase3 由来) / 100 ⚠️ / Other
仮説間の依存関係	なし / あり (Other で「H4 depends on H1」のように記述)

Kaggle publicScore is never offered (project rule, memories project_om_finding / project_case5_validation).

Phase 4 — Skip list hearing (実施しない検証 / 評価)

AskUserQuestion, 1 round, multi-select with Other free-text. Bucketed by category:

Q1. Analysis skips (multi-select; "どれもスキップしない" を選んだ場合は他選択を無視して空とする)

• どれもスキップしない (default) — skip list の分析項目はすべて実施 (default n<300 で結論を出さない のみ自動 ON)
• 300 対戦による評価はしない — 学習中 val_acc / loss curve や 30 ep 挙動確認だけで採否
• replay 分析 (experiment-analysis) は実施しない
• n<300 結果で結論を出さない ⭐default ON
• Other

Q2. Execution skips

• どれもスキップしない (default, Recommended)
• smoke test (1-episode self-play) を skip
• dev/test-bot を skip
• RunPod GPU を使わない (local CPU only)
• auto-recover loop を使わない (失敗時は手動)
• Other

Conflict checks:

• User selects Kaggle publicScore を採否に使う → reject (project rule, cite memories). Project default Kaggle publicScore は引用しない / skill rating は使わない are always ON and never asked here.
• User selects n<300 で結論を出す → warn (cite memory project_imitation_case1_phase3) and confirm before allowing.

No individual confirmation dialog is required for smoke test skip / dev/test-bot skip — record the user's choice as-is.

Per-hypothesis exceptions: free-text input via Other. Examples like "Q1 で 300 対戦 skip だが H4 だけは 300 対戦したい" are recorded in Phase 5's 例外条件 section.

Phase 5 — Write hypotheses.md

Output path (per .claude/rules/docs.md):

docs/experiment/{family}/{yyyymmdd}_case{N}_{topic}/hypotheses.md

• {yyyymmdd} — session injected currentDate (NOT shell date)
• {topic} — snake_case from Phase 0
• For "追記" mode, overwrite the same path (Phase 0 already backed it up).

Template:

# Hypotheses — {family}/case{N} {topic}

> 作成日: {yyyy-mm-dd}
> 最終更新: {yyyy-mm-dd}
> 状態: in_progress
> 最大 iteration: {N | リスト消化まで}
> 主要メトリクス: {例: ローカル self-play win rate vs. baseline_v5}
> 既定 episode 数: {300 | 100 | ...}

## 実施しない検証 / 評価 (skip list)

### 評価
- {例: ローカル self-play 300 対戦を行わない (学習 loss curve のみで採否)}
- {例: Kaggle publicScore は引用しない (project rule)}

### 分析
- {例: replay 分析は実施しない}
- {例: n<300 結果で結論を出さない}

### 実行
- {例: なし | smoke test skip ⚠️ | RunPod を使わない}

### 例外条件
- {例: H4 のみ「採否が inconclusive なら 300 対戦を追加実施」}

## 仮説リスト (priority 順)

- [ ] (P1) H1: {一文 hypothesis} — {期待効果 / mechanism}
- [ ] (P1) H2: {...}
- [ ] (P2, depends on H1) H3: {...}
- [ ] (P3) H4: {...}
- [ ] (deferred) H5: {...}  # 工数大 / 後回し

## Iteration log

(各 iter 完了時に experiment-analysis / experiment が追記)

| iter | 開始 | 仮説# | plan path | run_id | 主要メトリクス | 採否 | result path |
|---|---|---|---|---|---|---|---|

## 参考 (References)  # Phase 2 を実施した場合のみ

- {URL} — {1 行所感}

Write directly via Write, then report 3–5 lines to the user (in Japanese):

• written path
• hypothesis count + priority distribution
• key skip-list entries
• next-step suggestion: /experiment to consume the list, or /experiment-plan for a single iteration

Do not auto-spawn experiment-plan / experiment. Wait for the user to invoke them.

Risk gates this skill enforces

• Never offer Kaggle publicScore as a metric. Skip-list keeps it default ON; reject if the user tries to remove it.
• Warn on n<300. When 100-episode evaluation is selected in Phase 4, surface a one-line warning citing memory project_imitation_case1_phase3.
• Confirm smoke / test-backend skips. One explicit confirmation dialog each before they take effect.
• Do not edit code. Read-only; only hypotheses.md is written.
• Time-box web research to ≤ 5 minutes. Deeper surveys belong to /research-retrieval.
• Don't silently overwrite an existing hypotheses.md. Phase 0 must offer rebuild / append / read-only.

Common shapes

User says…	Skill behavior
/experiment-hypothesize imitation/case1 dropout	Phase 0 skipped (target given). Phase 1 reads case1. Phase 2 likely 不要. Phases 3–5 produce 5–10 hypotheses + skip list and write hypotheses.md.
"case4 の harass 周りで仮説を並べて、Lux S3 の解法も参考にして"	Phase 0 confirms case4. Phase 2 runs web research on Kaggle Lux writeups.
"case1 の hypotheses.md もう古いから作り直して"	Phase 0 picks 再構築, backs up the existing file, resets.
"実験プランを書いて" (vague)	Belongs to experiment-plan. Redirect.

Things to avoid

• Running this skill more than once for the same case to grow hypotheses.md. Rule: 1 case = 1 hypotheses.md.
• Running Phase 2 web research for more than 5 minutes. Deeper surveys belong to another skill.
• Writing iterN_plan.md content into hypotheses.md. Keep file responsibilities separate.
• Leaving the skip list empty. Project defaults (Kaggle publicScore は引用しない, n<300 で結論を出さない) must always appear.
• Overwriting an existing hypotheses.md without user confirmation.
• Including Kaggle publicScore as a success metric on any hypothesis.
• Auto-spawning experiment-plan or experiment after writing the file. Suggest by name only.

Language

• Internal reasoning and thinking should be in English
• All user-facing output, AskUserQuestion labels/descriptions, and the written hypotheses.md body must be in Japanese (per .claude/CLAUDE.md)