Is this model physically possible?¶

Eleven artifacts. One question.

An S-parameter file, a coupling extractor, a machine-learned surrogate — each arrives as numbers, and numbers do not announce that they describe something which cannot exist. A model that produces power from nothing, responds before it is excited, or claims that adding a grounded conductor between two others increases their coupling looks exactly like a model that does not. Simulators believe them. Accuracy metrics do not object.

This is a set of tools that object.

The one thing to understand before using any of it¶

Every tool here is a grader. It can tell you a model is wrong. None of them can tell you a model is right.

That is not modesty, it is the actual logical shape. These checks test necessary conditions — properties any physically realizable network must have. Passing them means nothing was caught. A perfectly passive model of entirely the wrong structure passes every law in the portfolio.

If you take one thing from these docs, take that. Everything else follows from it, including several design decisions that look inconvenient until you see why.

Start where your problem is¶

You have	Start with
A folder of `.s2p` files from a vendor	`sparam-lint`
A parasitic extractor you are writing	`maxwell-lint`
An ML surrogate about to ship	`abstain-bench`
Network files to read, write or convert	`touchstone-tools`
A CI pipeline	`physics-lint-action`
An AI agent reasoning about hardware	`physics-lint-mcp`
A checker of your own, and no way to grade it	`sparam-conformance`
A proof that needs sound enclosures	`interval-core`
Thirty seconds and curiosity	the browser demo

Or install the front door and let it route:

pip install git+https://github.com/nickharris808/physics-lint.git
physics-lint doctor
physics-lint check .

How the pieces relate¶

                    sparam-conformance          screening-ceiling
                  (labelled ground truth)     (a certified theorem)
                            │ grades                    │ produced by
                            ▼                           ▼
  .sNp files ─► touchstone-tools ─► sparam-lint ──┐   interval-core
                (read/write/convert)              │
                                                  ├─► physics-lint ─┬─► the Action (CI)
  an extractor ──────────────► maxwell-lint ──────┤   (one CLI,     └─► the MCP server (agents)
                                                  │    SARIF)
  any model ─────────────────► abstain-bench ─────┘

Two of these are not tools at all. sparam-conformance is a corpus of labelled networks that grades the graders — without it, "our checker works" is an assertion. screening-ceiling is a machine-certified impossibility result, and interval-core is the arithmetic that produced it.

The four documents worth reading¶

What this proves — and what it does not The necessary-versus-sufficient distinction, in full. Read this before you quote a green result to anyone.
The concepts Passivity, reciprocity, the screening ceiling, certified bounds — explained for someone who did not do the course.
An end-to-end walkthrough A folder of unknown vendor models to a green CI badge, passing through most of the portfolio.
FAQ Written from the objections people actually raise, including the ones that are right.

Every checker ships proof that it still discriminates.

A physics checker reporting 100% compliance is indistinguishable from one that has quietly stopped working — a tolerance widened until nothing fails, a transposed reshape, an aliased phase unwrap. So:

sparam-lint --self-test builds a network violating each law and requires each check to reject it;
maxwell-lint ships a reference model that must fail;
abstain-bench ships a model that games the metric and must still be refused;
interval-core's self-check is tested against a deliberately broken rounding;
sparam-conformance re-derives every ground-truth label by independent linear algebra — which caught three wrong labels in its own corpus during development.

Licence and provenance¶

Code is Apache-2.0; the two datasets are CC-BY-4.0. Everything is synthetic or computational — no proprietary or measured data anywhere in the portfolio.

These tools grade a model, or move one between formats. 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.