VeriTrace is a deterministic conformance and replay harness for AI-agent governance. It imports recorded agent traces, evaluates them against versioned policies, generates mutation cases, and exports evidence bundles that can be verified offline.
Record once, replay anywhere, prove which policies were satisfied.
python -m pip install -e '.[dev]'
veritrace validate fixtures/demo_trace.json fixtures/demo_policy.json
veritrace replay fixtures/demo_trace.json fixtures/demo_policy.json \
--output demo-result.json --evidence demo-evidence.json
veritrace verify demo-evidence.json fixtures/demo_trace.json fixtures/demo_policy.json
veritrace report demo-result.json --output report.htmlThe project is intentionally offline-first. It does not execute tools, call an LLM provider, or claim regulatory certification. An unmatched event is unknown, not an unsafe allow.
Guardrail proxies can block actions and observability systems can store traces, but teams still need reproducible policy regression tests and portable evidence. VeriTrace separates trace ingestion, deterministic evaluation, mutation testing, and offline verification so it can complement existing runtimes.
The core package contains strict Pydantic models, a policy evaluator with explicit verdict semantics, a one-pass replay engine, a hash-chained evidence format, and a Typer CLI. The schema is designed to align with OpenTelemetry GenAI concepts while remaining small and stable for local fixtures.
Read the architecture decision for the product rationale, threat model, requirements, and roadmap.
pip install -e '.[dev]'
pytest
ruff check .
mypy srcApache-2.0. See LICENSE.