Second Opinion — adversarial honesty reviewer

An independent second reader for a finished DRAFT. It checks whether every claim is backed by its evidence — fabrication, transcription slips, invented IDs, wrong PASS/FAIL, conclusions the data don't support — and hands you a findings report that is itself a DRAFT. A human owns the verdict. Domain-neutral: pharma, frontend, backend, anything.

A companion to agentic-sop-kit (it reads that kit's run artifacts), but it installs and runs independently. Deterministic checks in code + a capped, advisory LLM pass. MIT.

🌐 English · 繁體中文

        FULL  ─ /second-opinion <kit run_dir> ──┐
                                                 ├─▶  deterministic checks (HARD)  ─┐
        DEGRADED ─ /second-opinion <doc> --inputs ┘     +  capped advisory LLM (SOFT) ├─▶  DRAFT report ─▶ ✋ human decides
                                                                                      ┘

What it catches

#	Defect	Layer
#1	wrong PASS/FAIL or spec-limit verdict; aggregate that doesn't recompute	deterministic
#2	a number with no matching source token (fabrication / transcription)	deterministic
#3	an invented identifier / date / name (should be 【待補】)	deterministic
#4	a conclusion the data don't support (overreach)	advisory LLM

It does not review source-code quality / bugs / security (that's /code-review), and it does not audit workflow architecture (that's agentic-workflow-audit).

Two modes

• FULL — point at an agentic-sop-kit run_dir. Uses the run's trace, so #1/#2/#3 are HARD at confidence 1.0.
• DEGRADED — point at any document + supply the source inputs. Provenance is reconstructed best-effort, so findings are SOFT at 0.5; #1/#4 lean on the advisory layer.

Install (independent of the kit)

/plugin marketplace add s0912758806p/agentic-sop-to-work
/plugin install second-opinion-honesty@agentic-sop-to-work
/reload-plugins

Verify: /help lists /second-opinion. Requires Python 3.8+ as python3 (stdlib only).

Use

/second-opinion <run_dir>                       # FULL — a kit run dir
/second-opinion <doc> --inputs <file...>        # DEGRADED — any document
/second-opinion                                 # bundled cross-domain demo

Pairing without coupling

The only link to agentic-sop-kit is a read-only artifact contract (run_manifest.json + artifacts shaped {schema, produced_by, data, trace}). Second Opinion re-implements that contract for reading; it never imports the kit and never writes into the kit's run dir. With no kit present it still runs (DEGRADED, or the bundled demo).

The guarantees live in code, not prose

• The report's verdict is the frozen literal ADVISORY_ONLY and human_owns_verdict is always true — it cannot say "Second Opinion approved this."
• The advisory LLM pass is capped (SECONDOP_MAX_LLM_PASSES, default 1), every finding is clamped to SOFT / confidence ≤ 0.5, dropped unless it cites a verbatim draft span (or NO SOURCE), and suppressed if it re-litigates a slot the deterministic layer settled.

Enforce a human acknowledgement (optional — in code, not prose)

By default the human-owned STOP is a prompt the command presents. To make it a hard gate, a project opts in:

mkdir -p .second-opinion && touch .second-opinion/require_ack

Now the bundled Stop-hook blocks the session from ending until the latest DRAFT run (.agentic-sop-runs/<id>/) has a Second Opinion review and a human acknowledged it:

python3 -m secondop.ack            # a human runs this after reading the report

It enforces THAT a human acknowledged — never WHAT they decided. Anti-loop capped by SECONDOP_MAX_ACK_RETRIES (default 3); a silent no-op in projects that didn't opt in.

Develop

python3 tests/verify.py        # full regression (unit + integration + stdlib-only guard)

---

繁體中文

一個獨立的「第二意見」審查者，針對一份完成的 DRAFT，檢查它的每個主張是否有證據支撐 ——臆造、抄錯、捏造編號、判定錯（PASS/FAIL）、資料不支持的結論——然後給你一份本身也是 DRAFT 的發現報告。核准永遠由人決定。領域中立：pharma、前端、後端都能用。

它是 agentic-sop-kit 的搭配 plugin（讀取該 kit 的 run 產物），但可獨立安裝、獨立執行。 確定性檢查寫在程式裡 + 一層封頂的 advisory LLM。MIT 授權。

它抓什麼

• #1 判定錯（PASS/FAIL、規格上下限）、算不回去的彙總 —— 確定性層
• #2 兜不回輸入的數字（臆造／抄錯）—— 確定性層
• #3 捏造的編號／日期／姓名（應標 【待補】）—— 確定性層
• #4 資料不支持的過度結論 —— advisory LLM 層

不審原始碼品質／bug／資安（那是 /code-review）；不稽核工作流架構（那是 agentic-workflow-audit）。本 skill 只查「這次產出的主張 vs. 證據」誠實度。

兩種模式

• FULL：指向 agentic-sop-kit 的 run_dir，用 trace → #1/#2/#3 為 HARD（信心 1.0）。
• DEGRADED：指向任意文件 + 自備來源輸入，盡力重建來源 → 全部降為 SOFT（0.5）；#1/#4 交給 advisory 層。

安裝（不依賴 kit）

/plugin marketplace add s0912758806p/agentic-sop-to-work
/plugin install second-opinion-honesty@agentic-sop-to-work
/reload-plugins

需 Python 3.8+（python3，純標準庫）。

使用

/second-opinion <run_dir>                       # FULL
/second-opinion <doc> --inputs <file...>        # DEGRADED
/second-opinion                                 # 內建跨領域 demo

鬆耦合

與 kit 的唯一連結是唯讀 artifact 契約（run_manifest.json + {schema,produced_by,data,trace}）。 Second Opinion 只重實作該契約來讀取，不 import kit、不寫進 kit 的 run 目錄；沒裝 kit 也能跑。

保證寫在程式裡，不在散文

• 報告的 verdict 鎖死為 ADVISORY_ONLY、human_owns_verdict 恆為 true —— schema 上無法表達「Second Opinion 核准了」。
• advisory LLM 層封頂（SECONDOP_MAX_LLM_PASSES，預設 1）；每條發現一律夾鉗為 SOFT／信心 ≤ 0.5、未引用 draft 逐字片段（或 NO SOURCE）即丟棄、重提已定案 slot 即抑制。

強制人工確認（選用——寫在程式裡，不在散文）

預設那個「人核准」STOP 只是 command 印出來的提示。要把它變成硬閘門，由專案 opt-in：

mkdir -p .second-opinion && touch .second-opinion/require_ack

之後內建的 Stop-hook 會擋住 session 收工，直到最新的 DRAFT run（.agentic-sop-runs/<id>/） 有 Second Opinion 覆核且有人確認：

python3 -m secondop.ack            # 人看完報告後自己跑這個

它強制「有沒有人確認」，不強制「決定內容」。anti-loop 封頂於 SECONDOP_MAX_ACK_RETRIES （預設 3）；沒 opt-in 的專案完全 no-op。

MIT © s0912758806p