DCP Audit

dcp-audit is the small offline verifier for the Discovery Certification Protocol. It checks a content-addressed Audit Bundle and deterministically replays the recorded DCP decision. It does not call a model, access the Web, or rerun the original Auto-Research system.

Install

Use an isolated environment because another distribution already uses the top-level dcp package name.

python -m venv .venv
. .venv/bin/activate
python -m pip install dcp-audit

The numerical dependencies are pinned in the first release because a DCP certificate is compared with its fresh replay exactly.

Verify a Bundle

dcp verify path/to/audit-bundle

The default output is a short, human-readable summary.

Replay: VERIFIED
Bundle: sha256:...
Profile: full_audit_v1
Core: certified
Evidence: certified
Provenance: local_content_addressed_only
Formal issuance: no

Use JSON for automation.

dcp verify path/to/audit-bundle --json

The installed dcp command is the public CLI. The older python -m dcp verify ... entry point remains available with its original JSON-only output for compatibility with frozen audit workflows.

Exit status 0 means that the manifest, payload hashes, and kernel replay agree. It does not mean that the scientific verdict must be certified. Refuted and inconclusive decisions can also be validly replayed. Exit status 1 means verification failed, while 2 means the command was used incorrectly.

Python API

from dcp.api import verify_bundle

report = verify_bundle("path/to/audit-bundle")
if not report["ok"]:
    raise RuntimeError(report["errors"])

verdict = report["replayed_verdict"]
print(verdict["core"], verdict["evidence"])

Scope

This release supports the Core evidence bundle and full_audit_v1 profiles. It verifies local content integrity and deterministic kernel replay. It does not authenticate who ran the experiment, validate external registry signatures, or issue a formal certificate. Those boundaries are stated in every verification report.

The GitHub reference repository contains complete offline examples under examples/audits/ and a separate generalized evidence-producing harness. Each published Bundle retains its own producer and provenance record.