Verified Agent Skill record

Integrity Forensics

A thin launcher skill that runs the Anti-Autoresearch integrity-forensics sweep against a paper via a SHA-pinned clone, then converts the verdict into a typed policy gate (BLOCK/WARN/NO_NEW_BLOCKER) and an append-only obligations ledger.

Tasks

Run the integrity forensics sweep against a paper by cloning the Anti-Autoresearch repo at the pinned commit and following the upstream workflow end to end
Post-process the sweep verdict into a typed policy gate (BLOCK/WARN/NO_NEW_BLOCKER) and an append-only obligations ledger using forensics_gate.py

Inputs

A paper directory, PDF, or arXiv id given as the audit target ($ARGUMENTS)

Outputs

Upstream report.json and REPORT.md with a verdict in CLEAN_GIVEN_EVIDENCE / SOFT_FLAGS / HARD_FLAGS / REVIEW_UNAVAILABLE
gate.json recording the typed policy decision and paper_fingerprint
An append-only obligations ledger with typed, hashed receipts for each dispositioned obligation

Limitations and checks

The launcher exposes no reviewer model/effort parameters; the pinned upstream runs its own configuration (gpt-6-astra + xhigh) and this must not be overridden
Observability level (L0/L1/L2) is whatever upstream derives from the artifacts present; do not promise L2
There is no cryptographic binding between the report and the paper; a determined executor with shell access can hand it a stale report — the gate defends against sloppy executors and honest crashes, not the machine's owner
human: / checker: / cross-family-review: labels are accountability, not authentication
Confirm the clone is pristine after reset+clean (git status --porcelain empty) at the pinned commit before running
Confirm the eval-gate marker (${CLONE_DIR}.aris_eval_ok_${ANTI_AR_COMMIT}) exists outside the clone before accepting any verdict
Confirm forensics_gate.py resolved through the full canonical chain (.aris/tools/ → tools/ → $ARIS_REPO/tools/ → $ARIS_REPO/tools/ via ~/.aris/repo) and was not improvised
Confirm evaluate was run immediately after the sweep and before any paper file was touched (evaluate refuses a report older than any paper file)