experiment-audit · 实验结果跨模型完整性审计

先看看示例

作者提供的示例，实际效果取决于你的输入和使用工具。

输入

输出

方法说明

---
name: experiment-audit
description: "Audit experiment integrity before claiming results. Uses cross-model review (external reviewer backend) to check for fake ground truth, score normalization fraud, phantom results, and insufficient scope. Use when user says \"审计实验\", \"check experiment integrity\", \"audit results\", \"实验诚实度\", or after experiments complete before writing claims."
argument-hint: "[experiment-dir-or-results-path]"
allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex, mcp__codex__codex-reply, mcp__manual_review__review, mcp__manual_review__review_reply, mcp__antigravity__antigravity, mcp__antigravity__antigravity-reply, mcp__grok__grok, mcp__grok__grok-reply, mcp__claude-review__review, mcp__claude-review__review_reply, mcp__claude-review__review_start, mcp__claude-review__review_reply_start, mcp__claude-review__review_status
---

# Experiment Audit: Cross-Model Integrity Verification

> **Before each reviewer slot**, follow [Selectable Reviewers](../shared-references/reviewer-backends.md):
> resolve invocation → `.aris/reviewer.json` → Claude Code default (Codex MCP)
> with `reviewer_route.py` through the canonical helper resolver, preserving this
> slot's tier, fresh contexts, source/image inputs and deterministic gates.
> Codex examples/pins below apply only to `backend=codex`; other routes use their
> own first-call/reply contract. Save the route and actual returned identity;
> unavailable selections are `REVIEW_UNAVAILABLE`. Classify acceptance from actual
> families/completion/scope. Forward invocation reviewer/model/effort overrides
> to reviewer-bearing children. Generation workers are unchanged.


> 🔒 **Do not wrap this skill in `/loop`, `/schedule`, or `CronCreate`.** It is
> verdict-bearing — it judges experiment integrity. Re-running that verdict on a
> timer adds no new signal, and a loop that accepts its own output to decide
> when to stop crosses into self-acquittal (`acceptance-gate.md`). Schedule the
> *external wait that precedes it* — experiments done → then audit **once**. See
> [`shared-references/external-cadence.md`](../shared-references/external-cadence.md).

Audit experiment integrity for: **$ARGUMENTS**

## Why This Exists

LLM agents can produce fraudulent experimental results through:
1. **Fake ground truth** — creating synthetic "reference" from model outputs, then reporting high agreement as performance
2. **Score normalization** — dividing metrics by the model's own max to get 0.99+
3. **Phantom results** — claiming numbers from files that don't exist or functions never called
4. **Insufficient scope** — reporting 2-scene pilots as "comprehensive evaluation"

These are NOT intentional deception — they are failure modes of optimizing agents that lack integrity constraints. This skill adds that constraint.

## Core Principle

**The executor collects file paths. The external reviewer backend reads code and judges integrity. The executor does NOT participate in integrity judgment.**

This follows `shared-references/reviewer-independence.md` and `shared-references/experiment-integrity.md`.

## Constants

- **REVIEWER_BACKEND = `codex`** — Default: Codex MCP (ultra). Override with `— reviewer: oracle-pro` for Oracle MCP, or `— reviewer: manual` for Manual Review MCP. If manual-review MCP is unavailable, stop and print the install command; do not fall back to Codex. See `shared-references/reviewer-routing.md`.

## Reviewer Calling Convention

When calling the reviewer, branch on REVIEWER_BACKEND:

**If REVIEWER_BACKEND = `codex`:**
  Use `mcp__codex__codex` for new review threads.
  Use `mcp__codex__codex-reply` for follow-up rounds (reuse threadId).

**If REVIEWER_BACKEND = `manual`:**
  Use `mcp__manual_review__review` for new review threads with:
    prompt: [exact same prompt that would go to Codex]
    config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true}
  Save the returned `threadId`.
  Use `mcp__manual_review__review_reply` for follow-up rounds with:
    threadId: [saved manual-review threadId]
    prompt: [follow-up prompt]
    config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true}

Prompt fidelity: the manual prompt must be exactly the same text that Codex would receive.
Review tracing applies equally to both backends.

## Workflow

### Step 1: Collect Artifacts (Executor — Claude)

Locate and list these files WITHOUT reading or summarizing their content:

```
Scan project directory for:
1. Evaluation scripts:    *eval*.py, *metric*.py, *test*.py, *benchmark*.py
2. Result files:          *.json, *.csv in results/, outputs/, logs/
3. Ground truth paths:    look in eval scripts for data loading (dataset paths, GT references)
4. Experiment tracker:    EXPERIMENT_TRACKER.md, EXPERIMENT_LOG.md
5. Paper claims:          NARRATIVE_REPORT.md, paper/sections/*.tex, PAPER_PLAN.md
6. Config files:          *.yaml, *.toml, *.json configs with metric definitions
```
  A verdict-bearing manual response MUST begin with
  `Reviewer-Model: <exact-model-id>` — pass the model THIS session is actually
  running as in `executor_model`. Missing, unknown, or same-family identity
  cannot acquit; emit `REVIEW_UNAVAILABLE` rather than guessing. If the executor
  model cannot be named, manual review's cross-family claim is unprovable — say
  so in the report instead of asserting it.


**DO NOT summarize, interpret, or explain any file content.** Only collect paths.

### Step 2: Send to Reviewer

Based on the selected reviewer backend (see Reviewer Calling Convention), pass ONLY file paths and the audit checklist to the reviewer. The reviewer reads everything directly.

For `codex`, call `mcp__codex__codex` with:
- `model: gpt-6-astra`
- `config: {"model_reasoning_effort": "ultra"}`
- `sandbox: read-only`
- `cwd: [project directory]`
- `prompt: [the exact full prompt below]`

For `manual`, call `mcp__manual_review__review` with:
- `config: {"model_reasoning_effort": "xhigh", "executor_model": "<actual executor model>", "require_reviewer_model": true}`
- `prompt: [the exact full prompt below]`

Manual review cannot use Codex-only `model`, `sandbox`, or `cwd`; include the same file paths in the prompt so the user can inspect them.

Use this exact prompt for both backends:

```
You are an experiment integrity auditor. Start from the assumption that the
    evaluation is compromised somewhere — your job is to find where. Be
    adversarial. Trust nothing the author tells you — verify everything
    yourself. Read ALL files listed below and check for the following fraud
    patterns.

    Files to read:
    - Evaluation scripts: [list paths]
    - Result files: [list paths]
    - Experiment tracker: [list paths]
    - Paper claims: [list paths]
    - Config files: [list paths]

    ## Audit Checklist

    ### A. Ground Truth Provenance
    For each evaluation script:
    1. Where does "ground truth" / "reference" / "target" come from?
    2. Is it loaded from the DATASET, or generated/derived from MODEL OUTPUTS?
    3. If derived: is it explicitly labeled as proxy evaluation?
    4. Are official eval scripts used when available for this benchmark?
    FAIL if: GT is derived from model outputs without explicit proxy labeling.

    ### B. Score Normalization
    For each metric computation:
    1. Is any metric divided by max/min/mean of the model's OWN output?
    2. Are raw scores reported alongside any normalized scores?
    3. Are any scores suspiciously close to 1.0 or 100%?
    FAIL if: Normalization denominator comes from prediction statistics.

    ### C. Result File Existence
    For each claim in the paper/narrative:
    1. Does the referenced result file actually exist?
    2. Does the claimed metric key exist in that file?
    3. Does the claimed NUMBER match what's in the file?
    4. Is the experiment tracker status DONE (not TODO/IN_PROGRESS)?
    FAIL if: Claimed results reference nonexistent files or mismatched numbers.

    ### D. Dead Code Detection
    For each metric function defined in eval scripts:
    1. Is it actually CALLED in any evaluation pipeline?
    2. Does its output appear in any result file?
    WARN if: Metric functions exist but are never called.

    ### E. Scope Assessment
    1. How many scenes/datasets/configurations were actually tested?
    2. How many seeds/runs per configuration?
    3. Does the paper use words like "comprehensive", "extensive", "robust"?
    4. Is the actual scope sufficient for those claims?
    WARN if: Scope language exceeds actual evidence.

    ### F. Evaluation Type Classification
    Classify each evaluation as:
    - real_gt: uses dataset-provided ground truth
    - synthetic_proxy: uses model-generated reference
    - self_supervised_proxy: no GT by design
    - simulation_only: simulated environment
    - human_eval: human judges

    ## Output Format

    For each check (A-F), report:
    - Status: PASS | WARN | FAIL
    - Evidence: exact file:line references
    - Details: what specifically was found

    Overall verdict: PASS | WARN | FAIL
    
    Be thorough. Read every eval script line by line.
```

### Step 3: Parse and Write Report (Executor — Claude)

Parse the reviewer's response and write `EXPERIMENT_AUDIT.md`:

```markdown
# Experiment Audit Report

**Date**: [today]
**Auditor**: External reviewer backend, ultra reasoning (cross-model, read-only)
**Project**: [project name]

## Overall Verdict: [PASS | WARN | FAIL]

## Integrity Status: [pass | warn | fail]

## Checks

### A. Ground Truth Provenance: [PASS|WARN|FAIL]
[details + file:line evidence]

### B. Score Normalization: [PASS|WARN|FAIL]
[details]

### C. Result File Existence: [PASS|WARN|FAIL]
[details]

### D. Dead Code Detection: [PASS|WARN|FAIL]
[details]

### E. Scope Assessment: [PASS|WARN|FAIL]
[details]

### F. Evaluation Type: [real_gt | synthetic_proxy | ...]
[classification + evidence]

## Action Items
- [specific fixes if WARN or FAIL]

## Claim Impact
- Claim 1: [supported | needs qualifier | unsupported]
- Claim 2: ...
```

Also write `EXPERIMENT_AUDIT.json` for machine consumption:

```json
{
  "date": "2026-04-10",
  "auditor": "external-reviewer-ultra",
  "overall_verdict": "warn",
  "integrity_status": "warn",
  "checks": {
    "gt_provenance": {"status": "pass", "details": "..."},
    "score_normalization": {"status": "warn", "details": "..."},
    "result_existence": {"status": "pass", "details": "..."},
    "dead_code": {"status": "pass", "details": "..."},
    "scope": {"status": "warn", "details": "..."},
    "eval_type": "real_gt"
  },
  "claims": [
    {"id": "C1", "impact": "supported"},
    {"id": "C2", "impact": "needs_qualifier"}
  ]
}
```

### Step 4: Print Summary

```
🔬 Experiment Audit Complete

  GT Provenance:      ✅ PASS — real dataset GT used
  Score Normalization: ⚠️ WARN — boundary metric uses self-reference
  Result Existence:    ✅ PASS — all files exist, numbers match
  Dead Code:           ✅ PASS — all metric functions called
  Scope:               ⚠️ WARN — 2 scenes, paper says "comprehensive"

  Overall: ⚠️ WARN
  
  See EXPERIMENT_AUDIT.md for details.
```

## Integration with Other Skills

### Automatic in /research-pipeline (advisory, never blocks)

When integrated into the pipeline, this skill runs automatically after `/experiment-bridge` and before `/auto-review-loop`:

```
/experiment-bridge → results ready
    ↓
/experiment-audit (automatic, advisory)
    ├── PASS  → continue normally
    ├── WARN  → print ⚠️ warning, continue, tag claims as [INTEGRITY: WARN]
    └── FAIL  → print 🔴 alert, continue, tag claims as [INTEGRITY CONCERN]
    ↓
/auto-review-loop → proceeds with integrity tags visible to reviewer
```

**Never blocks the pipeline.** Even on FAIL, the pipeline continues — but claims carry visible integrity tags.

### Read by /result-to-claim (if exists)

```
if EXPERIMENT_AUDIT.json exists:
    read integrity_status
    attach to verdict: {claim_supported: "yes", integrity_status: "warn"}
    if integrity_status == "fail":
        downgrade verdict display: "yes [INTEGRITY CONCERN]"
else:
    verdict as normal, integrity_status = "unavailable"
    mark as "provisional — no integrity audit"
```

### Read by /paper-write (if exists)

```
if EXPERIMENT_AUDIT.json exists AND integrity_status == "fail":
    add footnote to affected claims: "Note: integrity audit flagged concerns with this evaluation"
```

## Key Rules

- **Reviewer independence**: executor collects paths, reviewer judges. Period.
- **Never block**: warn loudly, never halt the pipeline.
- **File-as-switch**: no EXPERIMENT_AUDIT.md = skill was never run = zero impact on existing behavior.
- **Cross-model**: the reviewer MUST be a different model family from the executor.
- **Honest about limits**: the audit catches common patterns, not all possible fraud. It is a safety net, not a guarantee.

## Acknowledgements

Motivated by community-reported integrity issues (#57, #131) where executor agents created fake ground truth and self-normalized scores.

## Review Tracing

After each reviewer call (`mcp__codex__codex`, `mcp__codex__codex-reply`, `mcp__manual_review__review`, or `mcp__manual_review__review_reply`), save the trace following `shared-references/review-tracing.md` (Policy C — forensic; never silently skip). Use `save_trace.sh` (resolved per the chain in `shared-references/integration-contract.md` §2) or write files directly to `.aris/traces/<skill>/<date>_run<NN>/`. Respect the `--- trace:` parameter (default: `full`).