PHM Physics Check

phm-lite-check

Deterministic physics plausibility checking for LLM outputs and robotics / embodied-AI datasets. Zero dependencies, pure standard library, Python >= 3.8.

This package is an independent, community-provided library: it is the offline subset of the PHM physical-judgment engine (proprietary core methods are intentionally excluded). A companion plugin (phm-physics-check) is listed on the Dify Marketplace.

Not affiliated with, endorsed by, or sponsored by LangChain, AutoGen,

or Dify. Integration adapters are community-provided.

MCP server

This package doubles as a local MCP server (stdio transport). After installing, point your MCP client at:

python3 -m phm_lite_check.mcp_server

Exposing three tools: physics_text_check, physics_quantity_check, robotics_data_check.

What it checks

Text / quantities (check_text, full_check)

• Dimensional / unit consistency (SI + common units)
• Universal speed limit (c)
• Absolute temperature floor (0 K)
• Non-negativity of mass / length / energy
• Probability range [0, 1]
• Kinematic relations (free-fall family, tolerance-checked)
• Conservation checks (energy / momentum, tolerance-checked)

Robotics records (check_robotics_record)

• Quaternion normalization (|q| within tolerance of 1.0)
• Timestamp sanity: monotonicity, duplicates, NaN/Inf, outlier gaps (possible frame loss)
• NaN / Inf scan over numeric arrays
• Unit annotation presence for common motion fields

Design guarantees

• Deterministic: same input -> same output, always reproducible.
• Zero dependencies: pure standard library; framework adapters import their host lazily.
• Verdicts: PASS / WARN / FAIL with structured violation records.
• No network calls: everything runs offline in-process.

NOTE: this package contains only the deterministic offline subset of the

PHM physical-judgment engine. Proprietary core methods are intentionally

excluded. (Deployment-specific deep checks are available separately via a

configurable remote engine endpoint, not included in this package.)

Install

Option A — from the GitHub repo (no package index needed):

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

Option B — from a Release wheel (faster, no git clone):

pip install https://github.com/PHM-Audit/phm-lite-check/releases/download/v1.0.0/phm_lite_check-1.0.0-py3-none-any.whl

Optional extras (once installed):

pip install "phm-lite-check[langchain]"   # + langchain-core (adapter tests)

pip install "phm-lite-check[autogen]" # + autogen-agentchat

PyPI listing is planned; both options above work today.

Quickstart — library

from phm_lite_check import check_text, check_robotics_record

check_text("The drone descends at 450 km/h and confidence is 1.8.")

{"verdict": "FAIL", "violations": [...], ...}

check_robotics_record({ "quaternion": [1.395, 0.0, 0.0, 0.0], # |q| >> 1 -> FAIL "timestamps": [0.0, 0.1, 0.05], # not monotonic -> FAIL "position": [1.0, 2.0], # no unit annotated -> WARN })

Quickstart — LangChain

from phm_lite_check.integrations import langchain_tool

tools = langchain_tool.get_langchain_tools() # 3 @tool callables agent = create_react_agent(llm, tools) # your usual wiring

Quickstart — AutoGen (v0.4+)

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(), )

Tools exposed by both adapters

Tests

pip install -e ".[dev]"

pytest tests/ -q

License

MIT

Benchmark

The judgment rules shipped here are exercised in public by [DPC-Bench v0.2](https://github.com/PHM-Audit/dpc-bench) — a paired clean/corrupted physical-consistency detection benchmark (7 task families incl. a real NGSIM I-80 family, three-caliber L1/L2/L3 labels).