autonomous-audit

Tamper-evident audit log for AI trading agents — a SHA-256 hash chain plus a human-readable decision report. Give any agent-driven trading workflow (Alpaca or otherwise) a lightweight integrity + traceability record for its decisions — with zero dependencies.

Extracted from the autonomous_ "Senate Protocol" decision-audit log into a standalone, stdlib-only package.

Why

An AI agent that trades makes decisions you may later need to review and check were not altered. autonomous-audit appends every decision to an append-only JSONL file under a SHA-256 hash chain: each record carries the hash of the previous one, so an in-place edit, mid-chain deletion or reorder breaks the chain and is detected. (It does not by itself detect truncation of the newest records or a wholesale rewrite from genesis — anchor the chain head externally if you need that.) A one-command report turns the machine log into a readable HTML document.

• Zero third-party dependencies — pure Python stdlib (no external supply-chain surface).
• Language-agnostic pre-image — the canonical pre-image is json.dumps(entry, sort_keys=True) (Python defaults: ensure_ascii=True, ", "/": " separators); a verifier in another language that reproduces it byte-for-byte validates the same chain — a Python-emulating serializer is required (naive JSON.stringify differs). See reference.md for the exact byte contract.
• Fail-loud — refuses to write when disk is low, rather than silently dropping evidence.

Install / run

# no install — run straight from PyPI with uv
uvx autonomous-audit demo                       # write + verify + report a sample chain
uvx autonomous-audit verify path/to/audit.jsonl # exit 0 = intact; 1 = problem (see message)
uvx autonomous-audit report audit.jsonl -o report.html

# or install it
pip install autonomous-audit

Use as a library

from autonomous_audit import AuditChainWriter, verify_chain

w = AuditChainWriter("audit_log.jsonl")
w.append({"symbol": "AAPL", "signal_action": "BUY", "consensus_score": "0.71",
          "timestamp": "2026-07-02T09:30:00+00:00"})

ok, err = verify_chain("audit_log.jsonl")   # (True, None) if intact

Keep record values float-free (use strings) if the same chain will be verified in JavaScript — float formatting can differ between json.dumps and JSON.stringify.

Async & throughput. append() does one blocking, synchronous write plus a disk check per record (pure stdlib, no async). In an event loop, run it in a thread — e.g. await loop.run_in_executor(None, writer.append, record) — so a fast market's order path is never blocked by audit I/O. It targets per-decision logging, not high-frequency (hundreds/sec) streaming.

Report

report / render_html_report() produces a self-contained HTML page: an integrity banner, one readable row per decision (subject, verdict, reason, short hash) and the required disclosure. Print it to PDF from any browser — no PDF dependency required.

Scope, limits & regulatory context

This tool gives you tamper-evidence and a readable trail of what your agent decided — integrity and traceability. It does not judge whether decisions were correct, and it is not investment advice, a regulatory control, or a compliance certification.

• Tamper-evident, not tamper-proof. A local JSONL file is inherently mutable; the chain lets you detect alteration (in-place edits, reorders, mid-chain deletions — not tail-truncation or a genesis rewrite), not prevent it. It does not on its own satisfy MiFID II statutory record-keeping (Art. 16 MiFID II; Art. 72/76 Reg. (EU) 2017/565; RTS 6). For regulated retention, pair it with WORM / durable-medium storage and anchor the chain head externally.
• Not explainability. Recording what was decided is not explaining why. Model explainability/interpretability (e.g. EU AI Act Art. 13/14; MiFID II algo-governance) is a separate obligation this tool does not meet.
• AI Act orientation. The EU AI Act's event-logging concept is Art. 12 (record-keeping), which binds high-risk systems; an AI trading agent is not high-risk merely by trading (Annex III's finance items are credit scoring and life/health-insurance pricing). Assess your own use case.

Chain integrity is not authorization. Keep the raw log as the source of truth.

License

Apache-2.0.