# v2.5: Local JSON reports and evidence calibration

API Relay Audit v2.5 adds machine-readable local reports and incorporates the
changes from the unpublished v2.4.1 draft. The previous public release is
v2.4.0. This document describes the v2.5.0 release candidate; publication is a
separate release gate.

## What Changed

- [#106](https://github.com/toby-bridges/api-relay-audit/pull/106) adds opt-in
  `--format json` to both standalone `audit.py` and modular `scripts/audit.py`.
  Markdown remains the default. JSON schema version 1 carries run provenance,
  the existing final risk rating, ordered flags, coverage, and full Markdown
  evidence. Stdout contains one JSON document; progress goes to stderr.
  `--output` writes UTF-8 to a local file. Connectivity checks retain their
  0/1 exit codes and return `risk_level: null` and `coverage: null`.
- [#108](https://github.com/toby-bridges/api-relay-audit/pull/108) keeps weak
  prompt-related keyword matches `INCONCLUSIVE`. Cat Test reports preserve
  exact caller payloads and describe instruction non-adherence without
  asserting the mechanism or attributing it to a provider.
- [#104](https://github.com/toby-bridges/api-relay-audit/pull/104), included in
  the earlier draft, supports exact DSH peers `0.1.0-rc.6`, `0.1.7-rc.2`, and
  `0.2.0-rc.2`. Curl fallback passes credential headers through config stdin
  and holds request bodies in private temporary files.
- [#97](https://github.com/toby-bridges/api-relay-audit/pull/97),
  [#80](https://github.com/toby-bridges/api-relay-audit/pull/80), and
  [#102](https://github.com/toby-bridges/api-relay-audit/pull/102) add research
  references, evidence boundaries, and signal-interpretation guidance.

The 14 audit steps, `general` / `web3` / `full` profiles, and existing six
dimension LOW / MEDIUM / HIGH risk matrix are unchanged. For example, Step 4
can remain inconclusive while the final rating is LOW: read individual
findings alongside the rating.

## Quick Start

Tagged downloads below become available when v2.5.0 is published. A draft
release alone does not guarantee that a public Git tag exists.

```bash
AUDIT_SCRIPT_REF=v2.5.0
curl -fsSL "https://raw.githubusercontent.com/toby-bridges/api-relay-audit/${AUDIT_SCRIPT_REF}/audit.py" -o audit.py
export API_RELAY_AUDIT_KEY=your-relay-key
python3 audit.py --key-env API_RELAY_AUDIT_KEY --url https://relay.example.com/v1 --format json --output report.json
```

See the [JSON schema and coverage boundaries](../json-report.md).

## Verification

- Candidate gates: 885 Python tests, 19 DSH adapter tests, version sync,
  standalone drift, metrics drift, and zero-dependency standalone startup.
- Eight real CLI loopback checks cover both distributions, successful and failed
  connectivity, stdout and UTF-8 file output, exit codes, and credential
  exclusion using synthetic credentials.
- Before publication, required PR checks and the draft-release workflow must
  pass for the release commit. Verify the source, generated standalone,
  downloaded release asset, and SHA-256 file are byte-identical.
- Standalone `audit.py` SHA-256: `fd99409bf745090a26bb4616fd33fbe0d873de08cfe077df413b2b93845f151d`
- After publication, verify `/releases/latest` and `/releases/tags/v2.5.0`,
  the tag target, and attached assets. Local test success alone does not prove
  those external states.

## Limits

JSON is an unredacted local execution report, separate from the draft public
report-artifact schema. It may contain private targets and model responses;
review and redact before sharing. The DSH integration retains its existing
Markdown workflow and Claude model baseline. This candidate adds no connector,
new detector, model baseline, or safety certification. Interactive DSH host
submission and real-provider attribution are outside this candidate's local
verification.