Plumb · the trust layer
Your models should know
when they’re guessing.
Plumb flags when a Cogitan model is operating outside what it was trained on — in a single forward pass — so you never act on a confident-but-blind prediction. Part of Cogitan’s structured-design suites, and licensable as a standalone trust layer — enabled in-app and validated on your designs.
0.98
off-distribution detection (AUROC) · vs 0.48 for MC-dropout
1 pass
not 30 — no resampling, no retraining
~9×
cheaper than sampling-based uncertainty
70%
of violations caught at a 5% review budget
What Plumb is
Knows when it doesn’t know.
Every Cogitan model is a fast stand-in for something slow and physical — a SPICE run, a measurement. The dangerous failure isn’t being a little off. It’s being confidently wrong about something it has never seen.
Plumb measures how far an input sits from the model’s training distribution and flags the unfamiliar ones for fallback. One pass on the embedding the model already computes. The conventional method — MC-dropout, run 30 times — is slower and, on this exact job, no better than chance.
The operating curve
Catch more by reviewing less.
Set the review budget — the share of clean cells you’re willing to send to full simulation — and Plumb returns the share of rule-violating cells it catches. A 5% budget catches 70% of violations. A 10% budget catches 87%. One forward pass per cell.
The same approach holds beyond chip design: on language-model representations, Plumb separates familiar from unfamiliar inputs at 0.95+ AUROC, where confidence-based methods collapse to near-chance. A general capability, proven on our own silicon first.
DRV cells caught vs. review budget — RSFQ-JEPA, 4,390 cells
39%
1% budget
70%
5% budget
87%
10% budget
Head to head
The standard method can’t see novelty.
At flagging off-distribution cells, MC-dropout scores 0.48 — statistically indistinguishable from a coin flip — while spending 30× the compute. Plumb scores 0.98, in one pass. Higher is better; 0.5 is random, 1.0 is perfect.
Dashed line = 0.50, random guessing. AUROC, off-distribution (DRV) detection.
Select suites · licensable
In the suites that need it.
Licensable for the rest.
Plumb earns its place where a design’s inputs are structured enough that an unfamiliar one can’t be caught by a simple bounds check — and where a confident-but-wrong answer is expensive. That’s our structured-design suites, starting with Fluxus. We turn it on within a license and validate it on your designs first — a higher license tier, not a per-transaction charge.
Building your own surrogates in a domain with that same structural risk? Plumb is licensable as a standalone trust layer — the validated implementation, and the map of exactly where single-pass geometric detection beats your cheaper checks.
Included
your design suite
- 200–500× faster than SPICE
- 1σ confidence bound on every output
- Batch screening & SPICE-fallback shortlist
Premium tier
licensed, not per-cell
- Everything in the suite license
- Off-distribution flagging at 0.98 AUROC, 1 pass
- Enabled in-app & validated on your designs
- Catches the confidently-wrong cases simpler checks miss
Plumb is a licensing tier, not a per-transaction charge — we enable and validate it on your designs before you rely on it.
What Plumb is for
Built to catch the unfamiliar.
Plumb answers one question precisely: is this design outside what the model was trained on? That’s the failure mode where a fast surrogate is most dangerous, and where confidence-based methods are blindest.
It complements, rather than replaces, your in-distribution error checks. We build calibrated models and we’re precise about what each layer does — Plumb is the novelty detector, not a correctness oracle. That distinction is the point.
Put a trust layer
on your models.
Plumb runs in Cogitan’s structured-design suites, starting with Fluxus RSFQ verification — and we license it to teams building their own surrogates where being confidently wrong is costly. If your models stand in for something expensive, they should know their limits.