Preflight · design-rule verification
Catch the design error
before the fab run.
Import a design. Preflight scans it against a rule database and returns violations and near-boundary warnings in seconds — before simulation time or fab budget is committed. Every rule carries the evidence that settled it.
The problem
Anyone can write down design rules.
The hard part is knowing which ones are true.
A rule database is worth what its weakest rule is worth. Published heuristics get repeated without provenance. Thresholds drift from the papers they came from. A rule that fires on working hardware costs more credibility than it saves in fab time.
The product is not the rules. The physics is discoverable by anyone. The product is being the independent checker where every rule carries its verification provenance — and where a rule that cannot be verified is not shipped as though it were.
How a rule earns its badge
Four kinds of claim need four kinds of evidence. One blended “percent verified” number hides which rules you can act on, so each modality is settled by its own criterion and carries its own badge.
| Modality | Settled by | Settled |
|---|---|---|
| Oracle | An exact ground-truth run — master-equation solve, or FEM | 29 / 78 |
| Proof | The cited theorem read, and confirmed to state the claim | 38 / 48 |
| Convention | An endorsed position recorded, with its rationale | 24 / 32 |
| Device data | Traced to a cited measurement, with a verbatim quote | 14 / 15 |
The standing constraint
Ground truth only. A surrogate model may decide what to verify next. It may never be the verification. Putting an approximation inside the trust boundary destroys the only thing an independent checker sells.
Cannot decide
A probe that cannot reach a verdict returns INCONCLUSIVE. Never a pass.
Refutation
REFUTED is a real outcome. It is written down and kept, not discarded.
Absence
Not finding a claim in a paper records where we looked. It is never evidence the claim is new.
Comparison
Every comparison is made at matched resource, or it is not made.
Where it stands today
174
Rules tracked
105
Settled · 60%
15
Real devices in corpus
7
Gates, all green
Bosonic QEC
LiveCat, GKP, binomial and dual-rail codes
153 rules
Superconducting
LiveTransmon, dispersive readout, Purcell, electromagnetics
16 rules
Materials · surface loss
Candidate queueCPW participation and two-level-system loss
— rules
Two domains, stated plainly. Sixty percent settled means forty percent is not, and those rules carry no badge — they are reported as unsettled rather than shipped as verified. The materials rung emits candidates for measurement, not rules, because there is no exact solve for a material and a surrogate may not certify.
Run against hardware that exists
A false positive on a working device
is our defect, not yours.
The corpus holds 15 real published devices with verbatim provenance. If a rule fires a violation on hardware that demonstrably works, that is treated as a bug in the rule. Cry-wolf behaviour is the fastest way to make a checker worthless: a rule that flags every published cat qubit is not conservative, it is broken.
Gates, not intentions
Seven gates run on every change. Each exits non-zero on failure, and each carries a self-test that breaks it on purpose and confirms it goes red — because a check that cannot fail is not a check.
The gates exist because new work in this repository has a high defect rate — including work written to catch defects. They are the reason that is visible rather than shipped.
Send us a design.
We will tell you what it violates.
Rule counts and settled figures are generated from the live database. Unsettled rules are reported as unsettled. Coverage figures are not rounded up.