What it shows
The record states the result this way:
The published record says, word for word (an excerpt)
A fast predictor may answer only when a sound bound admits it; otherwise the system refines or escalates, so there is never a silently uncertified fast answer.
In plain words: each fast answer comes with a bound on its error, computed from the solver’s own residual. If the bound is too wide, the answer is not given as fast; it is refined with the real solver or passed on.
On 40 fresh layouts the bound was never violated. That is the useful half. The honest half is that the raw fast answers were certified on 0% of layouts, so in this test every answer went through refinement, and starting from the fast answer saved only 0.5% of the solver’s iterations.
Why it matters
A design flow that uses a fast stand-in needs to know which answers to trust without re-running the slow solver every time. Zero violations on the fresh layouts tested, against our own solver’s residual, is the property that would make a stand-in safe to put in a flow. How often it lets an answer through is what decides whether it saves time, and here it does not yet.
What is ours, and what is not
Predictors that abstain, and fast solves that carry a certified error bound, are known (see the prior art below). What is ours is this bound for our coupling stand-in and the test that it holds on fresh layouts.
Who should care
- Teams putting fast stand-ins into signoff flows, where a silent wrong answer costs more than a slow one.
- Reviewers of our results. The negative half is stated, not hidden.
The limits, in the record’s words
The published record says, word for word (an excerpt)
Blunt honest negative: raw certified coverage at τ_rel = 1e-3 is 0% — the model residual of order 1–10 far exceeds the ~1.2e-2 compression ceiling — and warm-start iteration reduction is 0.5%, NEGATIVE at N=8.
The published record says, word for word (an excerpt)
Zero bound violations on 40 fresh layouts; refined C versus frozen scipy ≤1.4e-11; fault-A inflates the bound ~1.2e11×.
In plain words: zero violations of its error bound on 40 fresh layouts against our own solver’s residual; not yet fast. It is a soundness contract over our own solver’s residual, not accuracy against silicon.
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.