What it shows
The record states the result this way:
The published record says, word for word (an excerpt)
Three demonstrations turn the physics into signoff consequences: a searched worst case at 116.07×, a measured pass/fail flip rate up to 19.4%
In plain words: the figure above is the worst case a search of 1400 layouts found, where the shortcut’s error was that many times our operator’s, against our own full-array solver. Across 72 designs, the shortcut’s pass or fail verdict disagreed with full-array truth for 19.4% of them at the isolation spec where it did worst. The record’s third demonstration is a large synthetic field of connections solved with no wrong verdicts against our own solver, by declining to answer when unsure, and that costs escalations.
Why it matters
An extraction error matters most where it changes a verdict: a design that passes when it should fail, or one redesigned for nothing. Showing the flips, not just the error, is what makes the case.
What is ours, and what is not
Calibrated uncertainty and declining to answer are known methods (see the prior art below). What is ours is the operator that replaces the shortcut, the searched worst case, and the measured flip rates.
Who should care
- Signoff teams who rely on pairwise coupling estimates for interconnect-dense packages.
- Reviewers. The record’s partial coverage of its own reproduction is stated below.
The limits, in the record’s words
The published record says, word for word (an excerpt)
Signoff consequences against full-array BEM truth, not against silicon. Zero wrong signoffs is achieved BY ABSTENTION — 25–65% of cases escalate, which is the disclosed cost.
The published record says, word for word (an excerpt)
The 116.07× hero is at the CORRECTED provenance: the prior '110.7× CMA-ES hero' was a random_reference draw MISLABELLED as searched
The published record says, word for word (an excerpt)
The robust calibrated family figure is 37.4× [31.6, 45.5] over 42 geometries at conformal coverage 0.901. WallFinder's honest negative is surfaced: the BINDING worst-aggressor isolation error is sub-dB to a few dB, so the value is removing the over-design tax, not a dramatic headline. HBM4 shipped with its own honest negative — the random-layout conformal sidecar under-covers on a regular grid — which entry 13 (`conformal-risk-stack`)'s Gate 123 then fixed. PARTIAL COVERAGE (found 2026-08-27): this entry's reproduce command verifies the HBM4 demonstration ONLY - one of the three named in the claim. The 116.07x searched hero lives in benchmarks/wsx/manybody_failure_atlas_2026_06_27.json and the 19.4% signoff-flip rate in benchmarks/wsx/package_wallfinder_2026_06_27.json; neither is re-derived by the command above.
In plain words: these are simulations against our own full-array solver, not silicon. The worst case is a single searched layout, and the steadier family figure is lower. The command that reproduces this record re-checks only the third demonstration; the worst case and the flip rate rest on their own files, served here as extracts.
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.