FAQ

Written from objections people actually raise. Several of them are right, and those are marked.

"My model passed. Is it correct?"

No. The checks are necessary conditions: passing means nothing was caught, not that the model is right. A perfectly passive model of entirely the wrong structure passes every law here. See what this proves.

"Isn't this just checking that σ_max ≤ 1? I can write that in ten lines."

You can, and that is roughly what the passivity check is. The value is not in the ten lines — it is in the four things around them:

1. Reading the file correctly. 2-port Touchstone is column-major and 3-port and above is row-major, so the obvious parser transposes every 2-port file and turns the reciprocity check into a no-op. That bug passes every test built on a symmetric fixture.
2. The negative control. Your ten lines will report 100% compliance whether they work or not, and you will have no way to tell which.
3. The refusals. NaN in the file, fewer than three frequency points for a derivative, a sweep that runs backwards — each needs a decision, and the easy decision produces a confident wrong answer.
4. The corpus. sparam-conformance exists because "my checker works" is an assertion otherwise. It caught three wrong labels in its own ground truth during development.

If you write your own anyway, please score it against the corpus and tell us what it finds.

"The tool flagged my ferrite isolator. That's a false positive."

Half right, and the distinction matters. The check is a true positive for the law and a false alarm for the device. A ferrite isolator's medium is non-reciprocal, so S ≠ Sᵀ is it working correctly.

The tool now says exactly that rather than claiming the network is unrealizable — that was a real over-claim, found while writing the tutorial and fixed.

What it will not do is decide for you. Record that non-reciprocity is expected for that file. Do not switch the law off globally: it exists to catch transposed reshapes, and a suppressed law catches nothing.

"Why won't it just skip files it can't parse?"

Because a checker that skips what it cannot read silently approves it. Exit code 2 (could not parse) deliberately outranks exit 1 (checked and failed): "I could not check this" is a worse answer than "I checked it and it failed", so it wins.

"Why does a file with NaN get refused instead of ignored?"

A physics verdict computed on NaN is not a verdict. There is no flag to override this, on purpose.

"SARIF findings don't show up on my PR diff."

Expected. SARIF results here carry no line region, because a physics failure happens at a frequency and a Touchstone file's rows are not the unit anyone reasons about. Inventing a line would make the annotation render and would be a claim about location we cannot support. Findings appear in the Security tab and the check summary.

"Isn't the screening ceiling obvious? Of course screening reduces coupling."

Right that it is obvious; wrong that it is therefore not worth checking. The point is not that engineers do not know it — it is that fitted and truncated models violate it anyway, and no accuracy metric reports that. maxwell-lint demo shows a second-order Born correction — the standard cheap fix — predicting k > 1 on 78% of pairs. Its mean error looks fine.

"You claim a theorem. About Maxwell's equations?"

No, and this is the objection most worth pressing. It is a theorem about a monopole-closure model — a zero-parameter analytic approximation that matches a boundary-element reference to 0.081% in the exact two-cylinder limit. The closure-versus-solver gap is an additive, disclosed, unresolved term that is never folded into the bound. (That 0.081% is itself the one number here you cannot re-derive from published code — it came from a solver that is not in any release.)

The dataset's scope line travels with the data in theorem.json for exactly this reason.

"Sampling the slider demo never finds a violation. Isn't that circular?"

Sampling cannot prove the bound; it can only refute it. That is why it is offered — it is a cheap, dependency-free way for a skeptical reader to try to catch us being wrong. Failing to find a counterexample is the honest expected outcome and is not the evidence. The evidence is the branch-and-bound proof, and its soundness rests on the interval arithmetic, which ships its own self-check.

"Why is interval-core a separate package?"

Because it is general and the family-specific certifiers are not. 745 lines of screening-law machinery stayed in the private tree deliberately; mixing them in would make a general-purpose library into a confused one. The published package is the arithmetic, not the application.

"Why does abstain-bench fail a model with 100% accuracy?"

Read the table by column. overconfident has a perfect in-envelope pass rate and answers every out-of-distribution query confidently, which is the default behaviour of almost every fitted model. timid has perfect abstention recall and abstains on everything.

Neither number is a score on its own. Your accuracy is measured on the cases your model chose to answer — that is the whole point, and pooling the two metrics conceals it.

"Can I use these on measured VNA data?"

You can run them, and the laws still apply. But the corpus is synthetic closed-form networks, so passing it does not establish that a checker behaves well on measured data with noise, drift and de-embedding artefacts. That gap is real and stated in the corpus's own scope section.

"Are any of these on PyPI?"

Not yet. Every install is pip install git+https://…, which is stated wherever a bare pip install <name> appears. The release workflows use PyPI Trusted Publishing and need no stored token; what is missing is a one-time browser step by the account owner.

One name is not merely unclaimed but taken: physics-lint on PyPI is an unrelated project, a linter for trained neural PDE surrogates. pip install physics-lint will not give you the meta-package on this site, ever. The other six names were free when last checked (2026-07-28), and a different distribution name will be picked for the meta-package before anything is published.

"Has any of this touched real hardware?"

No. Every number in the portfolio is computational. Nothing here has been near a probe. Where a figure came from tooling that is not in a release, the artifact says so at the figure rather than in a footnote.

"What is the commercial angle?"

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 ChipletOS.

The split is deliberate, and it is where it is because it was measured rather than assumed: at matched training compute, spending the budget on more solver-generated data beat spending it on a larger model, and a from-scratch reimplementation of the calibration reproduced the deployed threshold exactly. The algorithms are not the moat, so they are published; the corpus, the fitted weights and the solver that generates them are not.

Those two experiments were run against the closed corpus, so their numbers are deliberately not quoted here — nothing on this site should be a figure you cannot reproduce by running the published code.

"Can I contribute?"

Yes, and the most useful thing you can send is a new pathological case for sparam-conformance — a network that breaks a checker in a way the corpus does not currently cover. Each repository's CONTRIBUTING.md states its own non-negotiable rule; they differ, and they are the reason each package is trustworthy.