Physics Lint

📖 Documentation site — the portfolio narrative, the concepts, a full walkthrough, and what all of this proves (and does not).

A physics oracle your AI agent cannot talk its way past.

An LLM has no way to distinguish a physically impossible S-parameter matrix from a plausible one — both are just numbers. Hand an agent a fabricated model and it will reason confidently about it, cite it, and build on it. Nothing in the loop objects.

These MCP tools give the agent ground truth from linear algebra instead of from its own judgement.

30-second quickstart

git clone https://github.com/nickharris808/physics-lint-mcp.git
cd physics-lint-mcp

pip install git+https://github.com/nickharris808/sparam-lint.git@main \
            git+https://github.com/nickharris808/maxwell-lint.git@main \
            .

The two checkers are separate packages and are pulled from their own repositories; the server is the . at the end. Confirm the server answers over stdio before wiring it into a client:

$ printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"smoke","version":"0"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | physics-lint-mcp | tail -1 | python3 -m json.tool | grep '"name"'
                "name": "check_touchstone",
                "name": "self_test",
                "name": "check_screening",
                "name": "pairwise_error",

Not yet on PyPI. pip install physics-lint-mcp is the intended install once published; until then the three-package command above is the one that works.

Configure

Claude Desktop / Cursor / any MCP client:

{
  "mcpServers": {
    "physics-lint": {
      "command": "physics-lint-mcp"
    }
  }
}

A ready-made mcp.json ships with the package.

Tools

Tool	Arguments	What it answers
check_touchstone	path (string, required)	Is this .sNp file a physically possible passive network?
check_screening	c_full (array, required), c_iso (array, required) — both N×N	Does this coupling matrix violate the many-body screening ceiling?
pairwise_error	screening_factor (number, required), k in (0, 1]	How wrong is a pairwise extractor at this screening factor?
self_test	none	Does the checker still discriminate?

The schemas the server advertises over tools/list are the authoritative version of that table, and a test asserts the two agree.

A worked example: an agent that cannot fool itself

The failure this prevents is specific. An agent is handed a vendor .s2p, is asked whether the link will close, and reasons fluently about numbers that describe a network which cannot exist. Nothing in the loop objects, because an LLM has no way to tell an impossible S-matrix from a plausible one.

Wire the server in, and the first thing worth doing in a session is asking the checker to prove itself:

You: Before we look at the model, run self_test.

Agent: (calls self_test) The battery discriminates — five networks built to violate one law each were all rejected.

Now the model:

You: Check vendor/lna_stage.s2p.

Agent: (calls check_touchstone with path="vendor/lna_stage.s2p") This network is not physically realizable as a passive device: it fails passivity and energy conservation. I will not reason from its values.

That last sentence is not the agent being careful — it is the interpretation field coming back in the result, written for a model to read. An agent given only {"passed": false} will often keep going.

Two things the design forbids, and both matter more than they look: the agent cannot repair the model, because every tool is read-only; and it cannot mistake a failure for a crash, because a physics failure comes back as a result rather than a transport error. The next three sections are those two properties in detail, and the third is the case that will bite you first.

Troubleshooting

physics-lint-mcp: command not found — the console script did not install. Check pip show physics-lint-mcp; if it is there, the environment's bin directory is not on PATH, which is common when a client launches the server with a different shell. Use the absolute path in mcp.json: {"command": "/full/path/to/venv/bin/physics-lint-mcp"}.

The client shows the server as failed to start — run the stdio smoke test from the quickstart by hand first. It is the same code path with none of the client's supervision, so the traceback is visible.

No module named 'sparam_lint' — the two checkers are separate packages and are not pulled in automatically. Install all three, as the quickstart shows.

A tool call returns isError: true and the agent stops — that is a physics verdict, not a fault. The result carries the failed laws and an interpretation; your agent should read them rather than treat the call as failed.

Nothing at all comes back — the server speaks JSON-RPC over stdio, one object per line. A client that batches several objects into one line, or that writes without a trailing newline, will hang. The initialize request must come first.

The tool list is shorter than the table above — you are running an older build. tools/list is generated from the same definitions the table is tested against, so they cannot disagree within one version.

What the agent sees

Every result carries an interpretation field written for a model to read, because an agent acts on prose, not on a boolean:

{
  "physically_admissible": false,
  "failed_laws": ["passivity", "energy_conservation"],
  "interpretation": "This network is not physically realizable as a passive
    device. Do not use it as a reference and do not reason from its values.
    One legitimate exception: a non-reciprocal device such as a ferrite
    isolator will correctly fail the reciprocity law by design."
}

That last sentence matters. An agent told only "reciprocity failed" will helpfully "fix" a perfectly good isolator.

And on a clean result the interpretation says what the verdict does not mean:

All five laws hold; the network is physically admissible. Note this does NOT mean it is accurate — a passive model of the wrong structure passes every law here.

Every tool is read-only

Nothing here writes, fabricates, or repairs a model. An agent that could silently patch a failing network would defeat the purpose of having an oracle, so the absence of I/O is enforced by a test.

A physics failure is a result, not an error

A file that fails the laws returns a normal result with isError: true and the full verdict, rather than a transport-level error. The agent needs to see the failure to reason about it; an opaque protocol error teaches it nothing.

Try it without an agent

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' | physics-lint-mcp

Scope, honestly

These tools verify physical admissibility, not accuracy. A passive model of entirely the wrong structure passes every check. They are a floor, not a substitute for validation against a field solve or measurement.

The rest of the toolkit

Eight artifacts that answer one question in different places: is this model physically possible? Each is a grader — it can tell you a model is wrong; none can tell you one is right.

sparam-lint	Is an S-parameter model physically possible? Five laws + a negative control.
maxwell-lint	Does a coupling extractor predict impossible physics? Screening ceiling k ≤ 1.
abstain-bench	Does a model know when to shut up? Abstention recall, never pooled with accuracy.
sparam-conformance	11 labelled networks with verified ground truth. Grades the graders.
screening-ceiling	A certified impossibility result + 27 counterexamples. Zero-dependency verifier.
physics-lint-action	The same checks, in your CI.
physics-lint-mcp ← you are here	A physics oracle your AI agent can call.
Try it in your browser	All three checks, no install, runs client-side.

These tools grade a model. Producing one that is passive by construction — so it cannot fail these laws whatever its parameters — and accurate at speed in the many-body regime, with calibrated abstention and a fail-closed signoff certificate, is the commercial core: ChipletOS.

Licence

Apache-2.0. See LICENSE; copyright in NOTICE.

The signoff certificate and passive-by-construction synthesis these tools were written alongside are the ChipletOS closed core.

Contributing

One non-negotiable rule here: every tool stays read-only — an agent that could silently patch a failing model would defeat the point of having an oracle. CONTRIBUTING.md has the detail. Each sibling repository states its own, and they differ — that is deliberate, and it is why each is trustworthy on its own terms.

Citation

CITATION.cff is machine-readable; GitHub renders a “Cite this repository” button from it.