consensus-of-N pro placeholder · concept

Several models screen it. Disagreement is the signal.

A conceived capability, holding its place. Instead of one model screening a claim, several do — and the point isn't just a sturdier verdict when they agree. It's that when the models disagree, that disagreement is itself the tell that you're on a boundary.

◇

This is a placeholder, kept deliberately model-agnostic. The AI landscape shifts week to week, so this page shows the shape of the idea with generic model slots rather than naming models or pinning behavior that would date. It's a marker for a conceived Pro feature — not a designed screen like paper mode or the specialist audits, and not yet built.

the shape of it

Two outcomes, both useful.

N models screen the same claim independently. What you learn depends on whether they land together or apart.

when they agree

“Grounding bare feet to the earth reduces inflammation.”

model A

Proceed

model B

Proceed

model C

Proceed
↓
✓ a sturdier call

All N agree, so the verdict doesn't hang on one model's quirk. The screen is as trustworthy as a single-model screen can't be on its own.

when they split

“A 432 Hz tuning sits closer to the resonant frequency of the natural world than 440 Hz.”

model A

Proceed

model B

Failed emp.

model C

Proceed
↓
◆ a boundary, surfaced automatically ON A BOUNDARY

The models don't agree — and that's not noise to average away. It's the machine detecting the same ~7% seam that split independent human judges, without needing to be told in advance which claims are hard. Disagreement is the contested-band flag.

why it's worth a tier

A single model gives you one reading. Consensus-of-N gives you a reading and a measure of how settled that reading is — and it derives the uncertainty from the models themselves, rather than asking you to trust one. Agreement earns confidence; disagreement earns a flag.

It also does something the project already values elsewhere: it surfaces its own uncertainty instead of hiding it. The contested band stops being a number from the validation study and becomes a live property of each screen.

The reason it sits in a paid tier and not the free screener is purely mechanical: running several models at once needs real multi-model access and a backend to coordinate it — which is the same unbuilt foundation the other Pro capabilities wait on.

Placeholder for a conceived capability — not built, not yet designed in detail. Kept model-agnostic on purpose: naming specific models here would date within weeks, and which models to include is itself a decision for build time, informed by the measured reliability map. A within-single-model consistency check is possible sooner but is deliberately weaker, and is not what this describes. Like the other Pro features, the real version waits on the hosted backend and on the core screener proving useful first. See where it sits in the tiers →