pathproof-ai 0.2.0rc1

Local-first assurance and observability for AI agents

PathProof

Evidence-driven path assurance for AI agents.

Release status: 0.2.0rc1 is a public release candidate for testing. It is not yet a stable production release.

Current MVP version: 0.1.2

PathProof is a local-first Python library for monitoring, evaluating, and enforcing policies on AI-agent runs.

Quick start

pip install pathproof-ai==0.2.0rc1

from pathproof import AgentMonitor, Policy

policy = Policy(
    allowed_tools={"verify_customer", "search_orders", "create_ticket"},
    max_tool_calls=10,
)

monitor = AgentMonitor(
    agent_name="support_agent",
    version="0.1.0",
    policy=policy,
)

with monitor.run(task="Resolve customer complaint") as run:
    run.tool_call(
        name="verify_customer",
        arguments={"customer_id": "C-102"},
        function=lambda customer_id: {"verified": True},
    )
    run.record_output({"status": "resolved"})

report = run.evaluate()
print(report.to_text())

Included in v0.1

• Run and event recording
• Tool-call monitoring
• Policy checks
• Repeated-call detection
• Missing-prerequisite detection
• Human-approval gates
• Trust score with explanations
• Text and JSON reports
• In-memory, JSON, and SQLite storage
• Basic run comparison

Easier tool wrapping

with monitor.run(task="Demo") as run:
    @run.tool()
    def search(query: str):
        return {"results": [query]}

    result = search(query="agent assurance")
    run.record_output(result)

Runtime limits are enforced before an extra step or tool call in warning, approval, and blocking modes. Tool failures can be explicitly marked as handled with run.handle_tool_failure(...).

Evidence bundles (v0.1.3)

bundle = run.evidence()
assert bundle.verify()
bundle.to_json()

The bundle fingerprints the complete run and evaluation. Any later payload change causes verification to fail. This is tamper evidence, not yet a digital signature or legal certificate.

Regression cases (v0.1.3)

case = run.regression_case(name="Prevent repeated search")
case.save("regressions/repeated-search.json")

result = check_regression(case, candidate_report)
assert result.passed

Deeper comparisons

compare_runs() now detects tool and step sequence changes, new or removed tools, new failure classes, new policy violations, and trust-score changes.

Causal evidence graph

PathProof now links steps, tool requests, tool results, and outputs. It can distinguish a successful-looking output from an unsafe execution path.

Development installation

python -m pip install -e ".[dev]"
python -m pytest

Privacy

PathProof is local-first. It does not require a cloud service and does not transmit recorded data by default. Review stored inputs, outputs, and tool arguments before using it with sensitive information.

Important limitation

PathProof provides technical monitoring, evidence, and configurable policy enforcement. It does not by itself certify that an AI decision is legally, ethically, or factually correct.