Skill 寫法圖鑑

recursive-decision-ledger

affaan-m | Agent 與工作流 | 原文 | MIT

重複 rollout 時維護決策帳本,遞迴信心不能自動核准實盤或 deploy

中文翻譯

---
name: recursive-decision-ledger
description: 重複做 rollout(「Prime Gauss」風格的遞迴提示),同時維護一份只能追加的決策帳本,記錄各次試驗、標記、一致性檢查和晉升關卡,讓遞迴產生的信心絕對不會自動核准實盤交易、deploy 或破壞性操作。使用者要求重複 rollout、標記式決策流程、高維度搜尋、隨機最佳化、探索局部最佳解、ensemble 比較,或需要看得到證據軌跡的遞迴推理時使用。
license: MIT
metadata:
  origin: ECC
tools: Read, Write, Edit, Bash, Grep, Glob
---

# 遞迴決策帳本

使用者想靠重複 rollout 或「Prime Gauss」風格的遞迴提示,逼出更深的運算時,使用這個 skill。
保留有用的部分:重複試驗、先前的記憶、新資訊,和明確的標記。
拿掉不安全的部分:假裝這個迴圈能證明確定性。
Remove the unsafe part: pretending the loop proves certainty.

## 帳本契約

每次 rollout 都要記錄:

- rollout id 和時間戳;
- 先前被接受的贏家和先前的觀察名單;
- 吸收進來的新資訊;
- 搜尋空間大小;
- 用到的模型家族或啟發式方法;
- 試驗次數和有效試驗次數;
- 前幾名候選;
- 決策標記;
- 與先前帳本的一致性標記;
- 晉升關卡的結果。

append-only 的帳本優先用 JSONL,給人看的摘要用 Markdown。

## Rollout 迴圈

1. 載入先前的帳本。
2. 在時間步零擷取新資訊。
3. 跑有界的搜尋。
4. 標記每個候選:accept、watch、reject、decay watch,或 needs replay。
5. 把贏家與先前的贏家、最近一次標記過的 rollout 做比較。
6. 當漂移、尾部風險、資料過期或 replay 失敗讓先前的標記失效時,
   就把候選降級。
7. 先追加產出物,再做摘要。

## 一致性標記

加上一個精簡的一致性標記:

```text
Ensemble matches prior winner: true
Recursive matches prior winner: false
Latest rollout match: true
Live promotion allowed: false
Reason: replay and freshness gates not satisfied
```

## 晉升規則

交易、資金配置、production deploy、migration 或破壞性
操作,遞迴產生的信心不等於核准。

預設用 paper、dry-run、唯讀、預覽或分階段模式,除非使用者
明確核准實際執行,而且 repo/服務的關卡也支援。

只有在以下情況才晉升:

- 候選在選定的指標上勝過先前被接受的贏家;
- 正確性和 replay 檢查通過;
- 風險上限明確;
- 證據是持久的;
- 需要時,使用者已核准實際執行的步驟。

## 摘要形式

先講決策,不要講戲劇性的過程:

```text
第 15 次 rollout 完成。先前的贏家仍然成立,但優勢衰退了 17%。
狀態:watch,不上實盤。下一個關卡:20 筆 replay 成交,且新的
orderbook 時效低於門檻。
```

英文原文

---
name: recursive-decision-ledger
description: Run repeated rollouts ("Prime Gauss" style recursive prompting) while keeping an append-only decision ledger of trials, marks, coherence checks, and promotion gates, so recursive confidence never auto-approves live trading, deploy, or destructive actions. Use when the user asks for repeated rollouts, marked decision processes, high-dimensional search, stochastic optimization, local-optima exploration, ensemble comparison, or recursive reasoning with a visible evidence trail.
license: MIT
metadata:
  origin: ECC
tools: Read, Write, Edit, Bash, Grep, Glob
---

# Recursive Decision Ledger

Use this skill when the user is trying to force deeper computation through
repeated rollouts or "Prime Gauss" style recursive prompting. Preserve the useful
part: repeated trials, prior memory, fresh information, and explicit marks.
Remove the unsafe part: pretending the loop proves certainty.

## Ledger Contract

Every rollout should record:

- rollout id and timestamp;
- prior accepted winner and prior watchlist;
- fresh information ingested;
- search space size;
- model families or heuristics used;
- trial count and effective trial count;
- top candidates;
- decision marks;
- coherence marks against the prior ledger;
- promotion gate result.

Prefer JSONL for append-only ledgers and Markdown for human summaries.

## Rollout Loop

1. Load the prior ledger.
2. Capture new information at time-step zero.
3. Run the bounded search.
4. Mark each candidate: accept, watch, reject, decay watch, or needs replay.
5. Compare winners against prior winners and latest marked rollout.
6. Downgrade candidates when drift, tail risk, stale data, or failed replay
   invalidates the previous mark.
7. Append artifacts before summarizing.

## Coherence Mark

Include a compact coherence mark:

```text
Ensemble matches prior winner: true
Recursive matches prior winner: false
Latest rollout match: true
Live promotion allowed: false
Reason: replay and freshness gates not satisfied
```

## Promotion Rules

For trading, capital allocation, production deploys, migrations, or destructive
ops, recursive confidence is not approval.

Default to paper, dry-run, read-only, preview, or staged mode unless the user
explicitly approves the live action and the repo/service gate supports it.

Promote only when:

- the candidate beats the prior accepted winner on the chosen metric;
- correctness and replay checks pass;
- risk limits are explicit;
- the evidence is durable;
- the user has approved the live step when needed.

## Summary Shape

Lead with the decision, not the drama:

```text
Rollout 15 complete. The prior winner still holds, but edge deteriorated 17%.
Status: watch, not live. Next gate: 20 replay fills with fresh orderbook age
below threshold.
```