What it shows
A design search that scores candidates with a fast model tends to find the model’s mistakes: layouts that score well but would fail a real check. The defence is to admit a candidate only after a separate, slower calculation re-computes the quantity that matters.
The record states the result this way:
The published record says, word for word (an excerpt)
Candidates generated by gradient or search must clear a BEM admission
The published record says, word for word (an excerpt)
reaching a discrete topology 10.842 dB beyond the continuous-only incumbent.
In plain words: the search was allowed to add extra grounded connections around the two signal connections, and it found a layout whose two signal paths interfere 10.842 dB less than the best layout kept to the original two connections. Every candidate it kept passed an admission check that re-computes isolation with our own solver. That guards against the search exploiting the fast model that scores it; it does not guard against an error the fast model and our solver share.
Why it matters
Automated layout search is only useful if what it outputs is real. A gate that the search cannot satisfy by exploiting its own scoring model is the part that makes the output worth checking further.
What is ours, and what is not
Searching with a fast model and checking against a slower one is a known approach, and so is the failure it guards against (see the prior art below). What is ours is this gate on glass-carrier via layouts and the record of the two mistakes caught before the number was published.
Who should care
- Chip-packaging design-software vendors building automated layout search.
- Glass-carrier makers designing layouts of vertical connections.
The limits, in the record’s words
The published record says, word for word (an excerpt)
The winning layout passed the admission check, but its own certificate records it as not admitted to signoff. The comparison is not like for like: the search may add five extra grounded vias that the two-via comparison layout may not, and the isolation target (−9.754 dB) was set past the best the two-via layout reaches (−6.754 dB), so the two-via layout fails by construction. The search covered 180 candidates, and admission is re-computed by our own two-dimensional cross-section solver, not an outside one. Credibility comes from two adversarial catches before the number was published: an unfair bounds handicap, then a confound where the small-d initial point rather than the gradient was passing consensus; both were corrected first. The synthesis tool's 'exact adjoint' is exact for 4 of 9 levers: only 5 of 10 margins take the capacitance matrix. Its best output is an impossibility: a 6 pF/m differential-pair imbalance budget SMALLER than 3σ at σ ≈ 5.2 pF/m, which is a specification result, not an optimizer failure. A separate loop without the admission check (GLM-5.2) found a better layout at −19.11 dB, but it is nondeterministic, so the claim stays with the deterministic, checked run.
In plain words: the margin compares a search that may add five grounded connections with a layout that may not, against a target of −9.754 dB that the two-connection layout cannot reach (its best is −6.754 dB), so the comparison layout fails by design. The search was small, 180 candidates, every check uses our own solver, and the winning layout’s own certificate does not admit it to signoff.
Open source for this step
Tools and datasets we publish for the package step of building a multi-chip package. They are the checkers around this work, not a copy of the result itself.
- physics-lint: One command that checks a folder of physics models against a fixed set of named physical rules, with findings straight into CI.
- maxwell-lint: Flags a coupling extractor whose answers no passive set of conductors could produce.
- sparam-lint: Is your signal-response model physically possible? Five physical laws checked from the command line.
- interval-core: The interval arithmetic core behind our proofs over whole families of layouts.
- touchstone-tools: Read, write and convert Touchstone files, the standard text files that record how signals pass through a package's connections, and refuse to write one that cannot be read back.
- physics-lint-mcp: The physics checks, callable by an AI agent.
- physics-lint-action: A GitHub Action that fails the build when a model breaks one of a fixed set of named physical rules.
- Signal-response validity corpus: A labelled corpus of physically invalid signal-response networks, and a scorer that grades any checker against it.
- screening-ceiling: The screening-ceiling family as an open dataset.