dsh-aris-panel 0.2.7 → 0.3.1

package/skills/skills-codex/shared-references/reviewer-routing.md

@@ -5,18 +5,18 @@
 All reviewer-heavy Codex base skills use the same default contract:
 
 - executor: current Codex main agent
-- reviewer: second Codex reviewer, model `gpt-5.6-sol` (GPT-5.6-Sol)
-- reasoning effort: **two tiers** — **deep-audit** skills use `max` (`proof-checker`, `kill-argument` core threads, `research-review`,
+- reviewer: second Codex reviewer, model `gpt-6-astra` (GPT-6-Astra)
+- reasoning effort: **two tiers** (since 2026-07-10; `ultra`/`max` need codex-cli ≥ 0.144.1) —
+  **deep-audit** skills use `max` (`proof-checker`, `kill-argument` core threads, `research-review`,
   `experiment-audit`, `paper-claim-audit`, `result-to-claim`, `meta-apply`); **every other**
   reviewer call uses `xhigh` (multi-round loops and per-item fan-outs stay `xhigh` — a
-  follow-up `send_input` cannot change model/effort, and per-item `max` multiplies cost).
-  `gpt-5.6-sol` also supports `ultra`, but deep audits stay at `max` unless an explicit ultra request is made (`max`/`ultra` need codex-cli ≥ 0.144.1)
+  follow-up `send_input` cannot change model/effort, and per-item `ultra` multiplies cost)
 - round 1: `spawn_agent`
 - follow-up rounds: `send_input`
 
 This is the base default for `skills/skills-codex/`. No ARIS `— effort:` level or unrelated parameter changes the tier (ARIS `— effort: max` ≠ `reasoning_effort: max` — pipeline workload vs reviewer reasoning are different axes).
 
-**Capability fallback (first spawn of each tier only):** if `spawn_agent` errors explicitly on the effort enum (older codex-cli — applies only to the deep tier's `max`; `xhigh` predates 0.144.1), retry `reasoning_effort: xhigh`; if it errors explicitly on the model being unknown/unavailable to this account, retry `model: gpt-5.5` + `xhigh` (skip redundant intermediate steps). NEVER downgrade on timeout / rate-limit / auth / transport / server / context-length errors (risk of double-running). Never run a verdict-bearing review below `xhigh`; if no allowed pair works, report `REVIEW_UNAVAILABLE` — never substitute the executor's own judgment.
+**Capability fallback (first spawn of each tier only):** if `spawn_agent` errors explicitly on the effort enum (older codex-cli — applies only to the deep tier's `ultra`; `xhigh` predates 0.144.1), retry `reasoning_effort: xhigh`; if it errors explicitly on the model being unknown/unavailable to this account, retry `model: gpt-5.6-sol` + `xhigh`, then `model: gpt-5.5` + `xhigh`. NEVER downgrade on timeout / rate-limit / auth / transport / server / context-length errors (risk of double-running). Never run a verdict-bearing review below `xhigh`; if no allowed pair works, report `REVIEW_UNAVAILABLE` — never substitute the executor's own judgment.
 
 > ⚠️ **Same-family by default — provisional, never accepted.** The executor here
 > is Codex (GPT family) and the reviewer is a fresh Codex agent from the same
@@ -40,8 +40,8 @@ Single-round review:
 
 ```text
 spawn_agent:
-  model: gpt-5.6-sol
-  reasoning_effort: xhigh   # deep-audit skills: max (see contract above)
+  model: gpt-6-astra
+  reasoning_effort: xhigh   # deep-audit skills: ultra (see tier table above)
   message: |
     [role + task]
     Read the listed files directly.
@@ -51,8 +51,8 @@ Multi-round review:
 
 ```text
 spawn_agent:
-  model: gpt-5.6-sol
-  reasoning_effort: xhigh   # deep-audit skills: max (see contract above)
+  model: gpt-6-astra
+  reasoning_effort: xhigh   # deep-audit skills: ultra (see tier table above)
   message: |
     [initial review prompt]
 ```
@@ -70,7 +70,7 @@ send_input:
 
 When the user explicitly passes `--reviewer: oracle-pro`, switch only the reviewer route:
 
-- default reviewer remains Codex at the call's declared tier (deep-audit: max / regular: xhigh) if no reviewer is specified
+- default reviewer remains Codex at the call's declared tier (deep-audit: ultra / regular: xhigh) if no reviewer is specified
 - `oracle-pro` is optional, not the base default
 
 Routing rule:

package/skills/skills-codex/slides-polish/SKILL.md

@@ -36,7 +36,7 @@ manually) — do not run `/slides-polish` for that.
 
 ## Constants
 
-- **REVIEWER_MODEL = `gpt-5.6-sol`** — Codex MCP model for per-page review. xhigh reasoning is non-negotiable (see `../shared-references/effort-contract.md`). If the account has no `gpt-5.6-sol` access, follow the capability fallback chain in `../shared-references/reviewer-routing.md` (`gpt-5.5`+`xhigh`; `gpt-5.4` only as an explicit user override).
+- **REVIEWER_MODEL = `gpt-6-astra`** — Codex MCP model for per-page review. xhigh reasoning is non-negotiable (see `../shared-references/effort-contract.md`). If the account has no `gpt-6-astra` access, follow the capability fallback chain in `../shared-references/reviewer-routing.md` (`gpt-5.5`+`xhigh`; `gpt-5.4` only as an explicit user override).
 - **REVIEWER_REASONING = `xhigh`** — Hard invariant; the effort knob does **not** change this.
 - **CONTEXT_POLICY = `fresh`** — Each per-page review uses a **fresh** Codex reviewer call (`spawn_agent`, never `send_input`). See `../shared-references/reviewer-independence.md`. This prevents the reviewer from anchoring on prior fixes.
 - **REFERENCE_VISUAL** — Path to a PDF the user wants the polished deck to **align with** in visual weight (typography proportion, color discipline, callout density). Required input. If polishing PPTX only, the **Beamer compile of the same talk** is the ideal reference. If no reference exists yet, ask the user; do not silently default to "Why-RF" or any preset.
@@ -191,7 +191,7 @@ Schema notes:
 
 ```
 spawn_agent:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   reasoning_effort: xhigh
   message: |
     Triage pass. For each of N slides in <pptx-pdf-path>, compared against
@@ -225,7 +225,7 @@ recompile or save, move to next slide.
 
 ```
 spawn_agent:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   reasoning_effort: xhigh
   message: |
     SLIDE K review. Compare PPTX page K against reference page K.

package/skills/skills-codex/specification-writing/SKILL.md

@@ -13,7 +13,7 @@ Adapted from `/paper-write` for patent specifications. The specification support
 
 ## Constants
 
-- `REVIEWER_MODEL = gpt-5.6-sol` — External reviewer for specification quality
+- `REVIEWER_MODEL = gpt-6-astra` — External reviewer for specification quality
 - `JURISDICTION = "auto"` — Inherit from pipeline or detect from args; `CN`, `US`, `EP`, `ALL`
 - `OUTPUT_FORMAT = "markdown"` — Markdown drafts; converted to filing format by `/jurisdiction-format`
 - `OUTPUT_DIR = "patent/"` — Base output directory
@@ -156,7 +156,7 @@ Call `REVIEWER_MODEL` via a dedicated Codex reviewer agent at xhigh reasoning:
 
 ```text
 spawn_agent:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   reasoning_effort: xhigh
   message: |
     You are a patent examiner reviewing a specification for completeness.

package/skills/slides-polish/SKILL.md

@@ -48,7 +48,7 @@ manually) — do not run `/slides-polish` for that.
 
 ## Constants
 
-- **REVIEWER_MODEL = `gpt-5.6-sol`** — Codex MCP model for per-page review. xhigh reasoning is non-negotiable (see `../shared-references/effort-contract.md`). If the account has no `gpt-5.6-sol` access, follow the capability fallback chain in `../shared-references/reviewer-routing.md` (`gpt-5.5`+`xhigh`; `gpt-5.4` only as an explicit user override).
+- **REVIEWER_MODEL = `gpt-6-astra`** — Codex MCP model for per-page review. xhigh reasoning is non-negotiable (see `../shared-references/effort-contract.md`). If the account has no `gpt-6-astra` access, follow the capability fallback chain in `../shared-references/reviewer-routing.md` (`gpt-5.5`+`xhigh`; `gpt-5.4` only as an explicit user override).
 - **REVIEWER_REASONING = `xhigh`** — Hard invariant; the effort knob does **not** change this.
 - **CONTEXT_POLICY = `fresh`** — Each per-page review uses a **fresh** Codex thread (`mcp__codex__codex`, never `codex-reply`). See `../shared-references/reviewer-independence.md`. This prevents the reviewer from anchoring on prior fixes.
 - **REFERENCE_VISUAL** — Path to a PDF the user wants the polished deck to **align with** in visual weight (typography proportion, color discipline, callout density). Required input. If polishing PPTX only, the **Beamer compile of the same talk** is the ideal reference. If no reference exists yet, ask the user; do not silently default to "Why-RF" or any preset.
@@ -203,7 +203,7 @@ Schema notes:
 
 ```
 mcp__codex__codex:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   config: {"model_reasoning_effort": "xhigh"}
   sandbox: read-only
   prompt: |
@@ -238,7 +238,7 @@ recompile or save, move to next slide.
 
 ```
 mcp__codex__codex:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   config: {"model_reasoning_effort": "xhigh"}
   sandbox: read-only
   prompt: |

package/skills/specification-writing/SKILL.md

@@ -13,7 +13,7 @@ Adapted from `/paper-write` for patent specifications. The specification support
 
 ## Constants
 
-- `REVIEWER_MODEL = gpt-5.6-sol` — External reviewer for specification quality
+- `REVIEWER_MODEL = gpt-6-astra` — External reviewer for specification quality
 - `JURISDICTION = "auto"` — Inherit from pipeline or detect from args; `CN`, `US`, `EP`, `ALL`
 - `OUTPUT_FORMAT = "markdown"` — Markdown drafts; converted to filing format by `/jurisdiction-format`
 - `OUTPUT_DIR = "patent/"` — Base output directory
@@ -156,7 +156,7 @@ Call `REVIEWER_MODEL` via `mcp__codex__codex` with xhigh reasoning:
 
 ```
 mcp__codex__codex:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   config: {"model_reasoning_effort": "xhigh"}
   prompt: |
     You are a patent examiner reviewing a specification for completeness.

package/skills/training-check/SKILL.md

@@ -23,7 +23,7 @@ Periodically read WandB metrics during training to catch problems early. Do not
 
 - WANDB_ENTITY and WANDB_PROJECT: read from CLAUDE.md or passed as argument (format: `entity/project/run_id`)
 - CHECK_INTERVAL: starts at 10 minutes, then gradually increases if consistently healthy: 10 min → 20 min → 30 min → 60 min (cap)
-- REVIEWER_MODEL = `gpt-5.6-sol` — used via Codex MCP for ambiguous cases only
+- REVIEWER_MODEL = `gpt-6-astra` — used via Codex MCP for ambiguous cases only
 
 ## When to Use
 
@@ -73,7 +73,7 @@ Only escalate to Codex when the signal is ambiguous. For clearly good or clearly
 
 ```
 mcp__codex__codex:
-  model: gpt-5.6-sol
+  model: gpt-6-astra
   config: {"model_reasoning_effort": "xhigh"}
   prompt: |
     TRAINING HEALTH CHECK — need your judgment on ambiguous metrics.

package/skills/verify-manager/SKILL.md

@@ -0,0 +1,88 @@
+---
+name: verify-manager
+description: "ARIS 想法验证一条龙：在主线目录旁的 `_verify-` 同名平行目录里，从想法→可证伪假设与预注册判据→实验方案与任务→运行（含基线复现）→对照基线分析结果→提出下一轮新思路，一次调用走完全程，主线不受影响，结论成立才交接转正。用户说“验证一下这个想法”“我有个个人的想法需要验证”“做/进行一个简单的验证实验”“先小规模试一下”“跑个 pilot 看看行不行”“verify this idea”“quick sanity check”，或主线工作中冒出需要单独试的子想法时使用；也用于续做、列出、查看、转正、归档已有验证。用户提到全量、主线、正式实验、基线实验（跑基线、全量基线）时不用本 skill，那是主线正式实验：走 /experiment-plan、/experiment-bridge、/run-experiment。查新用 /novelty-check；核对论文数字、引用、证明用对应 audit 或 /proof-checker。"
+argument-hint: "[要验证的想法 | 继续 <slug> | list | show <slug> | promote <slug> | archive <slug>]"
+allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, Skill
+---
+
+# 验证空间（`_verify-`）
+
+一个想法从提出到出结论，再到提出下一轮新思路，全部在这个 skill 里完成。每一步的产物都写进与主线同名、带 `_verify-` 前缀的平行目录，主线目录始终只读。
+
+## 阶段
+
+| # | 阶段 | 产物 | 细则 |
+|---|---|---|---|
+| 1 | 定义假设 | `_verify-idea-stage/<slug>/HYPOTHESIS.md` | `phases/1-specify.md` |
+| 2 | 制定方案 | `_verify-refine-logs/<slug>/PILOT_PLAN.md` | `phases/2-plan.md` |
+| 3 | 拆任务 | `_verify-refine-logs/<slug>/PILOT_TASKS.md` | `phases/3-tasks.md` |
+| 4 | 运行实验 | `_verify-<实验目录>/<slug>/`，并勾选任务 | `phases/4-run.md` |
+| 5 | 对照基线分析 | `_verify-idea-stage/<slug>/RESULT.md` | `phases/5-analyze.md` |
+| 6 | 提出新思路 | `_verify-idea-stage/<slug>/NEXT_IDEAS.md` | `phases/6-next.md` |
+
+细则文件位于本 skill 的 base directory 下；进入某个阶段时才读对应的文件。
+
+## 硬规则
+
+1. **主线只读。** 不写 `idea-stage/`、`refine-logs/`、`paper/`、`review-stage/`、`research-wiki/`、实验目录，以及项目根下的 ARIS 产物；要用主线代码时，只读引用或复制进镜像目录。
+2. **先定判据再动手。** `HYPOTHESIS.md` 里的判据在任何实验或计算之前写定，之后不改；确需修改时，追加一条"判据变更"并写明原因。
+3. **基线先过关。** 基线复现检查未通过，不跑处理组，结论只能是 inconclusive。
+4. **结论只来自证据。** `RESULT.md` 里的每个数字都要能追到日志或输出文件；不事后放宽判据，不挑种子。
+5. **转正不碰主线。** `promote` 只写交接说明；进主线须经用户确认，交给主线 skill，用它自己的流程和审稿门完成。
+6. **预算闸。** 需要远程 GPU、单次运行预计超过 30 分钟、或要下载大文件时，先把 `PILOT_PLAN.md` 给用户看，等确认后再跑。
+
+## 辅助脚本
+
+`scripts/verify_space.py`（相对本 skill 的 base directory；只依赖 Python 3 标准库）负责建目录、渲染模板、记录状态、判断当前阶段、检测主线是否被改动，输出均为 JSON。下文 `VS` 指 `python "<base directory>/scripts/verify_space.py"`，没有 `python` 时用 `python3`。各命令都接受 `--root <项目根>`，默认当前目录。
+
+| 命令 | 作用 |
+|---|---|
+| `VS new --title "<标题>" [--slug <英文短名>] [--source "<出处>"] [--parent <slug>]` | 新建验证并渲染全部模板；`--parent` 记录派生关系 |
+| `VS check <slug>` | 各产物的待办数、闸门、任务进度，以及当前阶段 `phase` 和下一步 `next` |
+| `VS dir <slug> <实验目录名>` | 创建或获取实验镜像目录，如传 `pilot` 得到 `_verify-pilot/<slug>/` |
+| `VS conclude <slug> --verdict supported\|refuted\|inconclusive --summary "<一句话>"` | 结项：检查各产物完整、任务处理完、闸门通过，并报告主线是否被改动 |
+| `VS list [--all]` / `VS show <slug>` | 列出验证 / 查看单个验证 |
+| `VS promote <slug> --to <主线 skill>` / `VS archive <slug> --reason "<原因>"` | 转正交接 / 归档 |
+
+## 流程
+
+**新验证**：先定位项目根（主线 `idea-stage/`、`refine-logs/` 所在的那一层）。`VS new` 报 "no ARIS mainline found" 且上下文看不出是哪个项目时，问用户一句。然后按阶段 1→6 推进；每个阶段做完，跑 `VS check <slug>`，确认 `phase` 已前进再进入下一阶段。
+
+**续做**："继续 xxx""接着做那个验证"等：`VS check <slug>` 给出断点所在的 `phase` 和 `next`，读对应的细则接着做，已完成的产物不重做。
+
+**停顿点**：默认一路做到结项，只在三处停下——阶段 1 有【待澄清】时，附建议默认值一次性提问；触发预算闸时；结项后请用户选下一步。用户说过"直接做""不用问我"时，阶段 1 采用建议默认值，并把它们记入"解读与假设"。
+
+**汇报**：结项后用几行说明：假设、基线与处理组的关键数字、结论、主线是否未受影响、推荐的下一轮思路。然后给出选项：转正（supported 时）、开下一轮（`VS new --parent <slug>`）、归档。**不要自行转正或开下一轮**，除非用户事先要求连续迭代。
+
+## 管理已有验证
+
+| 用户意图 | 做法 |
+|---|---|
+| 有哪些验证、进展如何 | `VS list`（`--all` 含已归档） |
+| 看某个验证 | `VS show <slug>`，再读其中的 `RESULT.md`、`NEXT_IDEAS.md` |
+| 转正 | `VS promote <slug> --to <主线 skill>` |
+| 放弃、归档 | `VS archive <slug> --reason "<原因>"` |
+
+**转正**要求结论为 supported；否则需加 `--force --reason`。`--to` 按项目状态选：
+
+- 还没有主线想法：`idea-discovery`；
+- 已有主线方向，要把验证结果并入方案：`research-refine`；同时需要实验计划时用 `research-refine-pipeline`；
+- 验证的是实验设计本身：`experiment-plan`。
+
+命令会生成 `PROMOTION.md` 骨架。补全其中的交接摘要和建议的主线命令后展示给用户，用户确认后才运行主线 skill。
+
+## 结论的可信度
+
+结论由执行者自评，`STATUS.json` 记为 `assurance: self-assessed`，适合快速筛选，不等于独立审稿。用户需要独立判断时，对 `RESULT.md` 跑 `/research-review`；转正后，主线 skill 也会按自己的审稿门重新评估。
+
+## 边界
+
+| 用户要的是 | 走 |
+|---|---|
+| 隔离地试一个想法，从假设一直做到结论和下一轮思路 | **本 skill** |
+| 有没有人做过（查新） | `/novelty-check` |
+| 核对论文数字、引用、证明 | `/paper-claim-audit`、`/citation-audit`、`/proof-checker` |
+| 全量、主线、正式实验、基线实验（跑基线、全量基线） | 主线：`/experiment-plan` 定计划 → `/experiment-bridge` 实现并部署 → `/run-experiment` 运行 |
+| 已有完整实验计划，要部署执行 | `/experiment-bridge` |
+| 从宽泛方向找 idea | `/idea-discovery` |
+| 结合你的意见、参考论文修订 ARIS 初稿 | `/human-in-loop-refine`（其中需验证的改动会再调用本 skill） |

package/skills/verify-manager/phases/1-specify.md

@@ -0,0 +1,17 @@
+# 阶段 1：定义假设
+
+产物：`_verify-idea-stage/<slug>/HYPOTHESIS.md`
+
+1. **起名**：slug 用 2-5 个英文词的 kebab-case 概括想法，如 `median-robust`、`lora-rank-sweep`；想不出合适的英文名，就不传 `--slug`，由 `VS new` 生成时间戳 slug。
+2. **提取要素**：从用户原话或前文对话里，逐字取出研究问题、想法本身、对照对象、关心的指标、数据或场景、约束（算力、只用标准库、不联网等）。不要让用户重述。
+3. **填写**：用户没说清的地方，按领域常规做最合理的解读，写进"解读与假设"。只有答案会改变判据、基线或预算时，才在该处标【待澄清】，最多 3 处。其余一律用合理默认值，不提问。
+4. **自检**（全部满足才进入阶段 2）：
+   - 假设可证伪：写明条件、对照、指标和数字阈值；
+   - 只有一个主指标，其余都标为次要；
+   - 基线写清是什么、为什么是公平对照；
+   - 三种结论的判据互斥，覆盖所有结果；"基线复现检查未通过"必须归入 inconclusive；
+   - 假设只描述现象，不写实现细节；实现放到阶段 2；
+   - 范围写明不覆盖什么。
+5. **【待澄清】**：一次性列出所有问题，每个附 2-3 个选项和推荐默认值。用户回答后，把对应标记替换成答案；用户说"直接做"，就采用推荐默认值。
+6. **可选查新**：用户关心"有没有人做过"时，跑 `/novelty-check`，把结论摘要写进"解读与假设"，不单独存文件。
+7. `VS check <slug>`：确认 `phase` 已推进到 `plan`。

package/skills/verify-manager/phases/2-plan.md

@@ -0,0 +1,14 @@
+# 阶段 2：制定方案
+
+产物：`_verify-refine-logs/<slug>/PILOT_PLAN.md`
+
+1. **实验上下文**：逐项填写；拿不准的先标【待澄清】，再自己查清——读主线代码、日志、文档，或做快速检索。查不清且影响判据时才问用户，问法同阶段 1。进入第 7 步前，标记必须全部清除。
+2. **设计**：用能区分 supported 和 refuted 的最小规模。随机实验至少 2 个种子；能配对的，基线和处理组用同一组种子。纯推导、文献核对、玩具仿真也算验证，按实际情况写。
+3. **基线**：
+   - 优先在镜像目录里，用同样的数据、预算和评测协议复现基线，并写下"复现参考值 ± 容差"，作为阶段 4 的基线检查点；
+   - 只能引用主线已有结果时，写清出处文件，并在 G1 里说明双方是否对等。
+4. **关键决策**：每个会影响结论的选择写一行，含理由和考虑过的备选，供日后复盘。
+5. **预算**：估算耗时、算力、下载量，对照硬规则 6；超出时，先把方案给用户确认，再拆任务。
+6. **研究诚信闸门**：逐条核对 G1–G6。满足的写成 `- [x] Gn`；确实不适用的写成 `- [-] Gn`，并注明原因。有任何一条还是 `- [ ]`，或闸门整节被删掉，都不能进入阶段 3。
+7. 写"复现命令"（预计会跑的命令；执行后更新为实际命令）和"不做什么"。
+8. `VS check <slug>`：确认 `phase` 已推进到 `tasks`。

package/skills/verify-manager/phases/3-tasks.md

@@ -0,0 +1,19 @@
+# 阶段 3：拆任务
+
+产物：`_verify-refine-logs/<slug>/PILOT_TASKS.md`
+
+1. **格式**（每条任务一行，`VS check` 靠它统计进度）：
+   `- [ ] T001 [P] [BL] <动作>：<产出路径>`
+   - 状态：`[ ]` 待办、`[x]` 完成、`[-]` 跳过（行末写原因）、`[!]` 失败（行末写原因）；
+   - 编号 T001 起，按执行顺序递增；
+   - `[P]` 可选，只加在可并行的任务上：输出不同、互不依赖，例如不同种子的运行；
+   - 阶段标签必填，取 `[SU]` 准备、`[BL]` 基线、`[TR]` 处理组、`[AN]` 分析之一；
+   - 产出路径必须位于 `_verify-*` 镜像目录下，它是任务完成的凭证。
+2. **分阶段**：
+   - `[SU]`：建镜像目录、数据加载、评测代码、记录环境；
+   - `[BL]`（阻塞）：按同一协议跑基线，并做复现检查；检查不通过，不进入 `[TR]`；
+   - `[TR]`：在相同协议下跑处理组，每个种子或设定一条任务；
+   - `[AN]`：汇总数字、计算统计量、按判据判定。
+3. **检查点**：每个阶段末尾写一行"检查点：<验收标准>"。基线检查点必须写出复现参考值和容差。
+4. **粒度**：一条任务对应一次可验证的产出；预计超过 30 分钟的任务要拆开。
+5. 删掉模板里的【TODO】占位行；`VS check <slug>` 确认任务数大于 0，且 `phase` 已推进到 `run`。

package/skills/verify-manager/phases/4-run.md

@@ -0,0 +1,12 @@
+# 阶段 4：运行实验
+
+产物：镜像目录中的代码、日志和产出，以及 `PILOT_TASKS.md` 的勾选状态
+
+1. **预检**：`VS check <slug>`，确认 `phase` 是 `run`，三份文档没有待办，闸门全部处理完。
+2. **镜像目录**：`VS dir <slug> <实验目录名>` 取得路径。项目用 `pilot/` 就传 `pilot`；代码和产出分别放在 `src/`、`results/` 时，各传一次；都没有时传 `pilot`。代码和产出全部放进镜像目录；需要的主线代码只读引用，或复制进来。
+3. **执行**：按编号和依赖顺序执行，`[P]` 任务可以并行。每完成一条，立即把 `- [ ]` 改为 `- [x]`，并在行末追加实际产出路径。命令、种子、环境和输出写入 `run.log` 等固定文件名的日志。
+4. **失败**：先排错，最多 2 次；仍失败就标 `[!]` 并写明原因。`[SU]` 或 `[BL]` 失败时停下来汇报；其他阶段的任务标 `[!]` 后可以继续。
+5. **基线检查点**：`[BL]` 全部完成后，把结果和"复现参考值 ± 容差"对照。未通过时：把相关 `[BL]` 任务标 `[!]` 并写原因，所有 `[TR]` 任务标 `[-]`（原因写"基线未过检查点"），然后直接进入阶段 5，结论只能是 inconclusive。
+6. **偏离**：执行中改了方案的，在 `PILOT_PLAN.md` 的"偏离记录"里写明改了什么、为什么，不悄悄改。判据不改；确需修改时按硬规则 2 追加"判据变更"。
+7. **重算力**：需要远程 GPU 时用 `/run-experiment`；任务超过 10 个时用 `/experiment-queue`。两者的工作目录都指向镜像目录，不写主线目录。
+8. `VS check <slug>`：确认非 `[AN]` 任务都已处理（`[x]`、`[-]` 或 `[!]`），`phase` 已推进到 `analyze`。

package/skills/verify-manager/phases/5-analyze.md

@@ -0,0 +1,21 @@
+# 阶段 5：对照基线分析
+
+产物：`_verify-idea-stage/<slug>/RESULT.md`，并完成 `[AN]` 任务
+
+1. **取数**：从日志和结果文件里逐项取数，每个数字都注明出处文件；日志里没有的数字，不写。
+2. **主结果表**：列出基线和处理组在每个种子下的值、均值 ± 标准差、差值或比值；能配对的，计算配对差。
+3. **不确定性**：
+   - 种子数 ≥ 5：报告 bootstrap 95% 置信区间，或配对检验；
+   - 种子数 < 5：只报告均值 ± 标准差，以及各种子的方向是否一致；不报 p 值，并写明样本不足。
+4. **判定**：严格按 `HYPOTHESIS.md` 的预注册判据逐条判断。基线复现检查未通过时，一律判 inconclusive，不看处理组的结果。不事后放宽判据，不只挑有利的种子或设定。
+5. **混杂与有效性**：逐项检查，有问题就写进"局限"，并评估它会不会推翻结论：
+   - 基线与处理组的调参力度是否对等；
+   - 数据划分是否一致；
+   - 是否在测试集上做了选择；
+   - 是否有在日志里找不到的数字；
+   - 代理指标是否偏离了假设要的指标。
+6. **局限**：规模、种子数、玩具设定、代理指标、未覆盖的场景，以及因任务失败或跳过而缺失的证据。
+7. **可选工具**：
+   - 需要复杂统计或作图时可用 `/analyze-results`，产出放到 `_verify-pilot/<slug>/analysis/`，不写主线的 `figures/` 或 `results/`；
+   - 需要独立判断时，对 `RESULT.md` 跑 `/research-review`；用户要求时跑 `/experiment-audit`。
+8. 把 `[AN]` 任务标为 `[x]`；`VS check <slug>`：确认 `phase` 已推进到 `next`。

package/skills/verify-manager/phases/6-next.md

@@ -0,0 +1,19 @@
+# 阶段 6：提出新思路
+
+产物：`_verify-idea-stage/<slug>/NEXT_IDEAS.md`
+
+1. **观察**：从 `RESULT.md` 里摘出值得追的现象，每条都注明出处：
+   - 处理组输掉的设定；
+   - 出乎意料的数字；
+   - 失败或跳过的任务；
+   - 敏感性分析中的拐点；
+   - "局限"里列出的空白。
+2. **候选思路**：2–4 条。每条都是一个新的可证伪假设，写明：依据的是哪条观察、最小验证方案、粗略成本（时间、算力）、预期价值。不要写成泛泛的"进一步研究"。
+3. **排序**：按价值 / 成本排序，并结合结论：
+   - supported：优先考虑扩展适用边界的思路（更难的设定、更大的规模、更强的基线）；
+   - refuted：优先考虑补救或换方向；
+   - inconclusive：优先考虑补上所缺证据的最小测试。
+4. **推荐**：写出推荐的一条，以及理由。
+5. **结项**：`VS conclude <slug> --verdict <...> --summary "<一句话>"`，然后按 SKILL.md 的"汇报"规则向用户汇报。
+6. **来自修订轮的验证**（`STATUS.json` 的 `source` 以 `refine r` 开头，如 `refine r01 C02`）：结项后不单独让用户选转正，改为提示回到 `/human-in-loop-refine` 的该轮（"继续 r01"），由它按结论回流；改变整体方案的新思路也交回修订轮处理。
+7. **新思路不在本验证里执行。** 用户选定一条后，用 `VS new --title "<该思路>" --parent <slug>` 开下一轮。下一轮可以复用父验证的基线结果来省去重跑，但要在新的 `PILOT_PLAN.md` 里引用父验证的日志，并在 G1 里确认两轮的协议一致。

package/skills/verify-manager/scripts/test_verify_space.py

@@ -0,0 +1,202 @@
+#!/usr/bin/env python3
+"""Self-test for verify_space.py. Run: python test_verify_space.py"""
+
+import json
+import shutil
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+SCRIPT = Path(__file__).with_name("verify_space.py")
+failures = []
+
+HYP = "# 假设\n中位数 RMSE ≤ 0.9 × 均值 RMSE（10% 离群点）\n"
+PLAN_OPEN = "# 方案\n基线：样本均值\n- [ ] G1 对等\n- [x] G2 种子\n"
+PLAN_DONE = "# 方案\n基线：样本均值\n- [x] G1 对等\n- [x] G2 种子\n- [-] G3 不适用：无代理指标\n"
+TASKS_RUN = "# 任务\n- [x] T001 [SU] 建目录：_verify-pilot/s/\n- [ ] T002 [BL] 跑基线：_verify-pilot/s/bl.log\n- [ ] T003 [AN] 汇总：RESULT.md\n"
+TASKS_AN = "# 任务\n- [x] T001 [SU] 建目录\n- [x] T002 [BL] 跑基线\n- [ ] T003 [AN] 汇总\n"
+TASKS_DONE = "# 任务\n- [x] T001 [SU] 建目录\n- [x] T002 [BL] 跑基线\n- [x] T003 [P] [TR] 跑处理组\n- [x] T004 [AN] 汇总\n"
+TASKS_BL_FAIL = "# 任务\n- [x] T001 [SU] 建目录\n- [!] T002 [BL] 跑基线：复现偏差超容差\n- [-] T003 [TR] 跑处理组：基线未过检查点\n- [x] T004 [AN] 汇总\n"
+RESULT = "# 结果\nmedian RMSE 0.20 vs mean 5.24（run.log）\n"
+NEXT = "# 下一轮\n1. 对称污染下是否仍成立\n"
+
+
+def run(root, *args):
+    proc = subprocess.run([sys.executable, str(SCRIPT), *args, "--root", str(root)],
+                          capture_output=True, text=True, encoding="utf-8")
+    try:
+        return proc.returncode, json.loads(proc.stdout)
+    except ValueError:
+        return proc.returncode, {"stdout": proc.stdout, "stderr": proc.stderr}
+
+
+def check(label, cond, detail=""):
+    print(("  PASS  " if cond else "  FAIL  ") + label + ("" if cond else f" — {detail}"))
+    if not cond:
+        failures.append(label)
+
+
+def fill(path, text):
+    path.write_text(text, encoding="utf-8")
+
+
+def phase(root, slug):
+    return run(root, "check", slug)[1].get("phase")
+
+
+def main():
+    tmp = Path(tempfile.mkdtemp(prefix="verify-space-test-"))
+    try:
+        root = tmp / "proj"
+        (root / "idea-stage").mkdir(parents=True)
+        (root / "refine-logs").mkdir()
+        mainline = root / "idea-stage" / "IDEA_REPORT.md"
+        fill(mainline, "# mainline report\n")
+        original = mainline.read_bytes()
+
+        code, out = run(root, "new", "--title", "中位数在离群点下更稳", "--slug", "median-robust", "--source", "用户原话")
+        check("new creates a verification in the specify phase",
+              code == 0 and out.get("slug") == "median-robust" and out.get("phase") == "specify", out)
+        idea = root / "_verify-idea-stage" / "median-robust"
+        plan = root / "_verify-refine-logs" / "median-robust"
+        rendered = [idea / "HYPOTHESIS.md", idea / "RESULT.md", idea / "NEXT_IDEAS.md",
+                    plan / "PILOT_PLAN.md", plan / "PILOT_TASKS.md"]
+        check("all five templates render into _verify- mirrors", all(p.is_file() for p in rendered))
+        check("templates have no unrendered placeholders",
+              all("{{" not in p.read_text(encoding="utf-8") for p in rendered))
+        check("no warning when the mainline exists", "warning" not in out, out)
+
+        code, out = run(root, "new", "--title", "again", "--slug", "median-robust")
+        check("slug collision gets a suffix", code == 0 and out.get("slug") == "median-robust-2", out)
+        code, out = run(root, "new", "--title", "bad", "--slug", "Bad Slug")
+        check("invalid slug is refused", code == 1 and out.get("ok") is False, out)
+
+        fill(idea / "HYPOTHESIS.md", HYP + "数据规模【待澄清】\n")
+        check("an open clarification keeps the phase at specify", phase(root, "median-robust") == "specify")
+        fill(idea / "HYPOTHESIS.md", HYP)
+        check("a filled hypothesis moves to plan", phase(root, "median-robust") == "plan")
+        fill(plan / "PILOT_PLAN.md", PLAN_OPEN)
+        check("an unchecked gate keeps the phase at plan", phase(root, "median-robust") == "plan")
+        fill(plan / "PILOT_PLAN.md", "# 方案\n没有闸门\n")
+        check("a plan with its gates deleted stays at plan", phase(root, "median-robust") == "plan")
+        fill(plan / "PILOT_PLAN.md", PLAN_DONE)
+        check("resolved gates move to tasks", phase(root, "median-robust") == "tasks")
+        fill(plan / "PILOT_TASKS.md", TASKS_RUN)
+        code, out = run(root, "check", "median-robust")
+        check("pending tasks move to run and name the next task",
+              out.get("phase") == "run" and out["tasks"]["next_task"] == "T002" and out["gates"]["open"] == 0, out)
+
+        code, out = run(root, "conclude", "median-robust", "--verdict", "supported", "--summary", "s")
+        check("conclude refuses pending tasks and unfilled docs",
+              code == 1 and "pending" in out.get("error", "") and "RESULT.md" in out.get("error", ""), out)
+
+        fill(plan / "PILOT_TASKS.md", TASKS_AN)
+        check("only analysis tasks left moves to analyze", phase(root, "median-robust") == "analyze")
+        fill(plan / "PILOT_TASKS.md", TASKS_DONE)
+        check("a template-only RESULT.md stays at analyze", phase(root, "median-robust") == "analyze")
+        fill(idea / "RESULT.md", RESULT)
+        check("a filled RESULT.md moves to next", phase(root, "median-robust") == "next")
+        fill(idea / "NEXT_IDEAS.md", NEXT)
+        check("filled next ideas move to conclude", phase(root, "median-robust") == "conclude")
+
+        code, out = run(root, "dir", "median-robust", "pilot")
+        check("dir creates the experiment mirror",
+              code == 0 and (root / "_verify-pilot" / "median-robust").is_dir()
+              and out.get("path") == "_verify-pilot/median-robust", out)
+        code, out = run(root, "dir", "median-robust", "../escape")
+        check("dir refuses path traversal", code == 1, out)
+        code, out = run(root, "dir", "median-robust", "_verify-pilot")
+        check("dir refuses a mirror as the mainline name", code == 1, out)
+        fill(root / "_verify-pilot" / "median-robust" / "run.py", "print('ok')\n")
+
+        code, out = run(root, "conclude", "median-robust", "--verdict", "supported", "--summary", "中位数更稳")
+        check("conclude records the verdict", code == 0 and out.get("verdict") == "supported", out)
+        check("an untouched mainline is reported as untouched", out.get("mainline_touched") is False, out)
+        check("a concluded verification waits for the user's decision", phase(root, "median-robust") == "decide")
+
+        code, out = run(root, "new", "--title", "对称污染", "--slug", "sym-contam", "--parent", "median-robust")
+        check("a child verification records its parent", code == 0 and out.get("parent") == "median-robust", out)
+        code, out = run(root, "show", "median-robust")
+        check("the parent lists its child", "sym-contam" in out["verification"].get("children", []), out)
+        code, out = run(root, "new", "--title", "x", "--parent", "no-such")
+        check("an unknown parent is refused", code == 1, out)
+
+        run(root, "new", "--title", "baseline fails", "--slug", "blfail")
+        b_idea, b_plan = root / "_verify-idea-stage" / "blfail", root / "_verify-refine-logs" / "blfail"
+        fill(b_idea / "HYPOTHESIS.md", HYP)
+        fill(b_plan / "PILOT_PLAN.md", PLAN_DONE)
+        fill(b_plan / "PILOT_TASKS.md", TASKS_BL_FAIL)
+        fill(b_idea / "RESULT.md", RESULT)
+        fill(b_idea / "NEXT_IDEAS.md", NEXT)
+        code, out = run(root, "conclude", "blfail", "--verdict", "supported", "--summary", "x")
+        check("a failed baseline task blocks a supported verdict", code == 1 and "baseline" in out.get("error", ""), out)
+        fill(root / "idea-stage" / "NEW_FILE.md", "polluted\n")
+        code, out = run(root, "conclude", "blfail", "--verdict", "inconclusive", "--summary", "基线未复现")
+        check("mainline changes during a verification are reported",
+              code == 0 and out.get("mainline_touched") is True
+              and "idea-stage/NEW_FILE.md" in out.get("mainline_changes", []), out)
+
+        code, out = run(root, "promote", "blfail", "--to", "research-refine")
+        check("promote refuses an inconclusive verification", code == 1, out)
+        code, out = run(root, "promote", "blfail", "--to", "research-refine", "--force")
+        check("a forced promotion needs a reason", code == 1, out)
+        code, out = run(root, "promote", "median-robust", "--to", "research-refine")
+        check("promote writes PROMOTION.md for a supported verification",
+              code == 0 and (idea / "PROMOTION.md").is_file(), out)
+        check("PROMOTION.md has no unrendered placeholders",
+              "{{" not in (idea / "PROMOTION.md").read_text(encoding="utf-8"))
+        check("the mainline report is byte-identical after promotion", mainline.read_bytes() == original)
+
+        run(root, "archive", "blfail", "--reason", "基线未复现")
+        code, out = run(root, "list")
+        items = {v["slug"]: v for v in out.get("verifications", [])}
+        check("list hides archived verifications and reports phases",
+              "blfail" not in items and items.get("median-robust", {}).get("phase") == "closed"
+              and items.get("sym-contam", {}).get("phase") == "specify", out)
+        code, out = run(root, "list", "--all")
+        check("list --all includes archived ones", "blfail" in [v["slug"] for v in out.get("verifications", [])], out)
+
+        written = [p.relative_to(root).as_posix() for p in root.rglob("*") if p.is_file()]
+        stray = [p for p in written
+                 if not p.startswith("_verify-") and p not in ("idea-stage/IDEA_REPORT.md", "idea-stage/NEW_FILE.md")]
+        check("nothing is written outside _verify- mirrors", not stray, stray)
+
+        code, out = run(root, "new", "--title", "real templates", "--slug", "real")
+        docs = {k: root / v for k, v in out["docs"].items()}
+
+        def fill_placeholders(path):
+            lines = path.read_text(encoding="utf-8").splitlines()
+            fill(path, "\n".join("已填写" if "【TODO】" in l else l.replace("- [ ] G", "- [x] G") for l in lines) + "\n")
+
+        seen = []
+        for key in ("hypothesis", "plan", "tasks"):
+            fill_placeholders(docs[key])
+            seen.append(phase(root, "real"))
+        fill(docs["tasks"], docs["tasks"].read_text(encoding="utf-8") + "- [x] T001 [SU] 准备\n- [x] T002 [AN] 汇总\n")
+        seen.append(phase(root, "real"))
+        for key in ("result", "next"):
+            fill_placeholders(docs[key])
+            seen.append(phase(root, "real"))
+        check("real templates clear phase by phase once their placeholders are filled",
+              seen == ["plan", "tasks", "tasks", "analyze", "next", "conclude"], seen)
+
+        aris = tmp / "aris"
+        (aris / "skills").mkdir(parents=True)
+        (aris / "tools").mkdir()
+        fill(aris / "AGENT_GUIDE.md", "x")
+        code, out = run(aris, "new", "--title", "x")
+        check("refuses to run inside the ARIS repository", code == 1 and "ARIS repository" in out.get("error", ""), out)
+
+        bare = tmp / "bare"
+        bare.mkdir()
+        code, out = run(bare, "new", "--title", "x")
+        check("warns when no ARIS mainline exists", code == 0 and "warning" in out, out)
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+    print("OK" if not failures else f"{len(failures)} check(s) failed")
+    return 0 if not failures else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())