phm-lite-check — deterministic physics plausibility tools for AutoGen agents

[PHM (PHM-Audit)]

phm-lite-check — deterministic physics plausibility tools for AutoGen agents

Sharing a small open-source package I built: phm-lite-check — three deterministic, offline checkers wrapped as AutoGen v0.4 tools. Zero runtime dependencies (pure standard library), no network calls, no model calls.

The problem it addresses

Agents routinely produce numeric claims ("the drone descends at 450 km/h, confidence 1.8") and consume robotics/embodied-AI datasets. Some failures are purely mechanical — a probability of 1.8, a quaternion with norm 1.395, a negative mass, a timestamp sequence that goes backwards. These don't need an LLM to catch; a deterministic rule set catches them reliably and reproducibly.

Tools exposed (plain Python functions, usable with AssistantAgent(tools=[...]))

• physics_text_check(text) — scans text for physically implausible quantities: unit consistency, universal speed limit, absolute zero floor, non-negativity of mass/energy, probability range [0,1], kinematic relations, conservation checks
• physics_quantity_check(value, unit) — one quantity against physical law
• robotics_data_check(record_json) — robotics dataset records: quaternion normalization, timestamp monotonicity/duplicates/outlier gaps (possible frame loss), NaN/Inf scan, unit annotation presence

Verdicts are structured: PASS / WARN / FAIL with per-violation records (rule, severity, detail, location).

Quickstart

pip install git+https://github.com/PHM-Audit/phm-lite-check

from autogen_agentchat.agents import AssistantAgent
from phm_lite_check.integrations import autogen_tool

assistant = AssistantAgent(
    name="checked_assistant",
    model_client=your_model_client,
    tools=autogen_tool.get_autogen_tools(),
)

Actual output of physics_text_check("The drone descends at 450 km/h, confidence 1.8."):

{
 "verdict": "FAIL",
 "n_quantities_checked": 2,
 "violations": [
  {"rule": "probability_out_of_range", "severity": "FAIL",
   "detail": "Probability 1.8 outside [0, 1]", "location": "confidence 1.8"}
 ],
 "engine": "PHM-Lite", "version": "1.0.0", "deterministic": true
}

Scope and limits

• This is the deterministic offline subset of a larger physical-judgment engine (the PHM engine); proprietary deep-check methods are intentionally excluded. The offline subset covers mechanical/deterministic checks only — it does not perform statistical or learned validation, and it does not reason about context-dependent physical plausibility.
• Determinism: same input yields the same output under the same package version.
• 36 tests pass; the LangChain adapter tests additionally run against langchain-core 1.7.1 (an optional dependency, not required at runtime). Reproduce: pip install -e ".[dev]" && pytest tests/ -q

Not affiliated with, endorsed by, or sponsored by Microsoft or the AutoGen team — just a community tool that plugs into the standard tools=[...] interface.

Repo: https://github.com/PHM-Audit/phm-lite-check