CCS Receipt Batch Audit — bulk verification & tamper stats (local)

When you hold an entire batch of CCS AI-agent audit receipts — pipeline outputs, log archives, evidence sets handed across systems — and need to know in one shot which are authentic and which were tampered with or carry invalid signatures, use this skill. Verification runs locally by default (vendored open-source CCS core, no network, receipts stay on the machine). Up to 200 receipts per batch, verified against one issuer public key.

中文：本地批量审计最多 200 张 CCS 审计收据——逐张 Ed25519 验签 + JCS/SHA-256 完整性校验，输出有效/无效统计、通过率、篡改索引和逐条失败原因，可选哈希链 链接检查；默认零网络、数据不出本机。

What it does

Every receipt in the batch runs the same full pipeline as single-receipt verification:

1. RFC 8785 JCS canonicalization + SHA-256 content-hash recompute and comparison;
2. Ed25519 verification (one issuer public key for the whole batch; PEM/DER/raw 32 bytes accepted);
3. 22-field schema check;
4. Aggregate output: total / valid / invalid / pass_rate, failure-reason distribution, the tampered-index list (tampered_indexes), and per-receipt results (index, decision_id, verdict, error summary, tamper flag, duration).
5. Optional hash-chain check (--check-chain): for ordered batches, verify that previous_receipt_hash[i] equals the recomputed content hash of the prior receipt; reports the first break point (index + reason).

Stateless and idempotent: re-auditing the same batch gives identical statistics.

What it does NOT do

• It does not issue receipts and does not judge whether any individual business decision "should have allowed or blocked".
• It never receives a private key; the public-key trust anchor is managed by the user.
• Batch cap is 200 receipts (the service returns 400 beyond that); split larger sets.
• Local mode logs nothing remotely.

Usage

# Batch audit (public key embedded in the file, or supplied via -k)
python3 scripts/batch_audit.py -f examples/batch_mixed.json

# Separate public key + hash-chain check
python3 scripts/batch_audit.py -f receipts.json -k pub.pem --check-chain

# Machine-readable JSON (full per-receipt detail)
python3 scripts/batch_audit.py -f receipts.json -k pub.pem --json

# Read from stdin ({"receipts":[...]} or a bare receipt array)
cat receipts.json | python3 scripts/batch_audit.py -k pub.pem

Exit codes: 0 = all valid; 1 = at least one invalid/tampered receipt (report still printed); 2 = input error.

Arguments:

Arg	Description
-f/--file	Batch JSON file ({"receipts":[...]} or a bare array); omit or - for stdin
-k/--public-key	Issuer public key PEM file (required unless the file embeds public_key_pem)
--check-chain	Also verify hash-chain linkage between receipts
--json	Emit the full JSON report (per-item detail)
--online	Explicitly use Correctover's hosted audit (default is local, offline)
--endpoint	Online mode: override the service endpoint (self-hosted)

Privacy boundary: default batch verification is entirely local — no network request, receipts and public keys stay on the machine. The core is the open-source CCS implementation (same code as the official hosted service). Only with an explicit --online flag are receipts and the issuer public key sent over HTTPS; private keys are never needed, read, or transmitted.

Report fields

{
  "total": 5, "valid": 4, "invalid": 1, "pass_rate": 0.8,
  "fail_reason_distribution": {"Ed25519 signature verification failed": 1},
  "tampered_indexes": [4],
  "chain": null,
  "results": [
    {"index": 0, "valid": true, "decision_id": "dec-batch-000",
     "verdict": "allow", "errors": [], "tamper_detected": false,
     "duration_ms": 1.8}
  ]
}

Example

examples/batch_mixed.json holds 5 receipts (4 genuinely signed; the 5th tampered — verdict flipped to block, checks.security.status flipped to fail, signature unchanged):

  total   : 5
  valid   : 4
  invalid : 1
  pass    : 0.8
  tampered: [4]
  Per-receipt (invalid only; use --json for full detail):
    [#4] dec-batch-004 verdict=block tamper=True
        content_hash does not match recomputed SHA-256(JCS payload)
        Ed25519 signature verification failed (signature vs canonical payload mismatch)

A 5-receipt batch takes about 4–5 ms locally (measured; varies by machine).

Requirements

Python 3.7+, standard library only (urllib/json/hashlib). No third-party packages required for local mode.

About

By Correctover (Guigui Wang) — Correctover = AI Reliability. CCS (Correctover Conformance Shape) is an open, vendor-neutral runtime verification standard for AI agents: 7 dimensions with Ed25519-signed receipts that make agent tool calls auditable and tamper-evident — signed evidence you can hand across teams, pipelines, and systems. Homepage: https://correctover.com/?utm_source=clawhub&utm_medium=skill&utm_campaign=ccs_batch

Disclaimer

Results are auto-generated and apply only to the submitted data. Automated checks are an aid only and constitute no security guarantee, compliance conclusion, audit opinion, or legal opinion. No endorsement by OWASP, IETF, or any standards body is claimed. The business impact of tampered receipts is outside the scope of automated checks and requires manual review; any security or compliance decision must ultimately rely on qualified human review.

License

MIT-0 for this ClawHub distribution. Anyone may use, modify, and redistribute it, including commercially, with no attribution required.