Integrity Forensics

---
name: integrity-forensics
description: "Run SHA-pinned deterministic numeric/rules integrity forensics and record policy obligations. Codex mode can flag issues but cannot certify CLEAN or full semantic coverage."
argument-hint: "[paper-dir | pdf | arxiv-id]"
---

<!-- EVO-INQUIRY:BEGIN -->
For research work, read [EvoInquiry](../shared-references/evo-inquiry.md),
or reuse it if already loaded and unchanged in this session. Apply relevant
principles within the user's scope and this skill's existing outputs, review
standards, and recovery contracts. Carry them into nested research tasks and
reviewer briefs. Continue authorized work through its necessary steps; honor
budgets, blocking checkpoints, and stopping conditions.
<!-- EVO-INQUIRY:END -->


# Integrity Forensics — thin launcher (Codex-native: deterministic slice)

Audit target: **$ARGUMENTS**

> Same launcher doctrine as the mainline skill: SHA-pin, upstream eval-gate
> validation per pin, delegate unchanged, no vendoring, no forking, **no
> reviewer knobs**. The one Codex-native difference: upstream's nine auditor
> skills are Claude-Code contracts, so this pack runs upstream's
> **deterministic-only mode** — the numeric forensic core (GRIM / GRIMMER /
> statcheck / delta arithmetic) plus the rules-only reporter with an
> all-`review_unavailable` coverage map. That mode is honestly scoped by
> upstream: it can raise HARD/SOFT flags; it can NEVER return
> `CLEAN_GIVEN_EVIDENCE`. Translating upstream's reviewer calls into
> `spawn_agent` would REWRITE an upstream contract — forbidden.

## Constants

- **ANTI_AR_REPO = `https://github.com/wanshuiyin/Anti-Autoresearch.git`**
- **ANTI_AR_COMMIT = `b47af6f983b38347b6d2110379e266400597cf66`** — never
  tracks HEAD; bumping is a reviewed change (mainline Pin-bump checklist).
- **CLONE_DIR = `~/.ei/anti-autoresearch`** — host-neutral. An older clone at
  `~/.claude/anti-autoresearch` is unused; move it and its `.ei_eval_ok_*`
  receipt to keep an offline deterministic-only run working — this pack's own
  mode — otherwise delete it whenever convenient.
- **NO REVIEWER KNOBS** — and no `— effort:` mapping onto upstream settings.

## Step 0 — Bootstrap the pin (identical to mainline)

```bash
CLONE_DIR="$HOME/.ei/anti-autoresearch"
ANTI_AR_COMMIT="b47af6f983b38347b6d2110379e266400597cf66"
mkdir -p "$HOME/.ei"
if [ ! -d "$CLONE_DIR/.git" ]; then
    git clone --no-checkout https://github.com/wanshuiyin/Anti-Autoresearch.git "$CLONE_DIR"
fi
git -C "$CLONE_DIR" cat-file -e "$ANTI_AR_COMMIT^{commit}" 2>/dev/null \
    || git -C "$CLONE_DIR" fetch -q origin
git -C "$CLONE_DIR" checkout -qf "$ANTI_AR_COMMIT" || { echo "FATAL: cannot checkout pin"; exit 1; }
# pristine tree at the pin — local tampering (incl. nested-repo injections;
# hence double -f) must not run under the pin's name; verify, don't assume
git -C "$CLONE_DIR" reset --hard -q "$ANTI_AR_COMMIT" || { echo "FATAL: reset failed"; exit 1; }
git -C "$CLONE_DIR" clean -ffdxq || { echo "FATAL: clean failed"; exit 1; }
[ -z "$(git -C "$CLONE_DIR" status --porcelain)" ] || { echo "FATAL: tree not pristine"; exit 1; }
# marker OUTSIDE the clone (a marker inside a tamperable tree proves nothing)
MARKER="${CLONE_DIR}.ei_eval_ok_${ANTI_AR_COMMIT}"
if [ ! -f "$MARKER" ]; then
    ( cd "$CLONE_DIR" && python3 eval/run_eval.py ) || {
        echo "FATAL: upstream eval gate FAILED at pin — refusing an unvalidated pin"; exit 1; }
    touch "$MARKER"
fi
```

## Step 1 — Delegate: upstream deterministic-only mode, unchanged

Open **`$CLONE_DIR/workflows/anti-autoresearch/SKILL.md`** and follow its
**deterministic-only** path (its own documented degraded mode): Step 0 ingest →
Step 1 evidence ledger → deterministic auditors → adjudication with the
generated all-`review_unavailable` coverage map. Wrapper rules: run every
upstream bash block with `cd "$CLONE_DIR"` (upstream self-locates via
`git rev-parse --show-toplevel`); refer to the paper by ABSOLUTE path; never
rewrite upstream outputs.

Expected outcome: `report.json` whose verdict is `HARD_FLAGS` / `SOFT_FLAGS` /
`REVIEW_UNAVAILABLE` — by construction never `CLEAN_GIVEN_EVIDENCE`.

## Step 2 — Typed gate + obligations

Resolve `forensics_gate.py` via the canonical helper chain
(`shared-references/integration-contract.md` §2, Policy A), then:

```bash
python3 "$GATE_HELPER" evaluate --report "$PAPER_DIR/report.json" --paper-dir "$PAPER_DIR" \
    --anti-ar-commit "$ANTI_AR_COMMIT" --executor-model "codex-gpt-6-astra"
```

Policy: `HARD_FLAGS` → **BLOCK** · `REVIEW_UNAVAILABLE` → **BLOCK** (which a
deterministic-only run reports whenever it found no flags — the semantic
dimensions never ran, so nothing may wave the paper through) · `SOFT_FLAGS` →
**WARN**. The gate records `same-family` proposal provenance for a Codex
executor — informational: this gate only raises flags, it grants nothing.
The downstream preflight is ONE command:
`python3 "$GATE_HELPER" fresh --paper-dir "$PAPER_DIR" --anti-ar-commit "$ANTI_AR_COMMIT"`
— exit 0 ⟺ produced at the current pin ∧ gate
exists ∧ paper unchanged since ∧ gate matches the current ledger ∧ decision
pass-capable (WARN/NO_NEW_BLOCKER), where the decision is RE-computed from
the sha-verified archived report + live ledger (the stored token is display,
not authority). Any ledger mutation deletes the standing gate.json, and
`evaluate` refuses a report older than any paper file — neither a stale pass
nor a stale report can be replayed.

## Step 3 — Fix what it found

Identical obligations discipline to the mainline skill: append-only ledger,
`UNRESOLVED_DISAPPEARANCE` on vanished-but-unresolved findings, typed + hashed
`resolve` receipts (`corrected-from-results | claim-narrowed | claim-withdrawn |
citation-replaced`; `--verified-by` must be typed provenance —
`human:<name>` / `checker:<tool>` / `cross-family-review:<thread-id>` — and
the evidence file is RE-hashed on every later gate), human-only `waive`
(never a resolution), and **The One Forbidden Loop**: never "edit → re-sweep
→ repeat until it stops flagging".
Numeric obligations route to the result files; the rest to the matching audit
skill or the human.