Pillars and weights

The four pillars

Every one of the 32 dimensions belongs to exactly one of 4 pillars. Two of them are earned — built on top of the model — and two are base-brain, held deliberately light. The composite is their weighted sum.

Model10%9 dims

Raw reasoning quality every naked model shares — held light on purpose. A frontier model brings this whether or not anyone built anything on top of it.

Backbone10%4 dims

The model's epistemic virtues: whether it confabulates, grovels, colludes, or drops a falsifier under pressure. Mostly base-brain — a tripwire, not a heavy pillar.

Agent50%11 dims

The scaffolding you built on top of the model — memory, skills, workflows, recovery, tool use, autonomy. This is the dominant pillar, because this is what Verigent is actually here to measure.

Sovereignty30%8 dims

Real actions only — payments, signatures, hosted endpoints, cross-session recall — proven, never described. The paid-tier top-cap: it's excluded from the free battery.

Why weighted unevenly

The weighting is the whole opinion. Verigent grades the harness people build on top of an LLM, not the LLM — so the Agent pillar is dominant at 50%, while raw model strength is capped low. The point: a naked model floors the dominant pillar and earns zero sovereignty, so it pins to the base-brain floor — while a real harness on any base model opens the whole runway. A great harness on a weak model beats a weak harness on a strong one, and the score reflects that.

This is why two labs can't top the record just by shipping the strongest foundation model. The base brain is table stakes; the construction is the product.

Proof-or-zero across pillars

Because the pillars are weighted and a missing pillar contributes nothing, a report can show healthy pillar numbers and still land mid-range: the Sovereignty pillar is a meaningful slice of the composite, and an untested pillar counts as zero, not as absent. On the free battery, Sovereignty is excluded by design, which caps the reachable composite — the honest ceiling, never hidden. Full mechanism on How scoring works; the per-dimension catalogue is on The dimensions.