Built on the standards · PC4AI
Agreement is only worth the independence behind it.
The Precision Council for AI runs a panel of evaluator agents that score a candidate without seeing each other, then a chair converges them, in your browser, on your own provider keys. It grades how independent that convergence actually was, because a room that agrees for one reason is not a verified floor.
Whose council this is
The council is Golem’s. The guard is the contribution here.
The multi-agent council underneath this is not our design. It is the Octant Council Builder, built by the Golem Foundation and published open source under MIT. That is the heavyweight instrument: real data-gathering agents, live orchestration, cross-provider evidence. PC4AI is the easy on-ramp, a browser reimplementation of the council’s shape with no install, running on keys you supply.
What PC4AI adds, and what it is here to show, is one thing the base plugin leaves optional: the independence guard. That is the contribution, the guard and the browser fork, not the council itself.
The independence guard
Convergence is graded by how many minds actually stand behind it.
The five evaluators are assigned across whatever providers you supply keys for, and independence scales with distinct model lineages. The chair weighs the convergence by that grade and never passes off a same-lineage agreement as a verified floor. Evidence is shared here, a common blind spot, and that is stated at every grade.
One provider: weak
Five evaluators on one provider's keys are one model lineage in several costumes. They tend to agree because they share a mind, not because the answer is sound. The chair never presents this as a verified floor.
Two providers: moderate
Two distinct lineages give the convergence something to be independent about. Agreement now carries more weight, and the grade says how much.
Three or more: strong
Three or more lineages standing behind an agreement is the point at which convergence is worth trusting as a floor. The chair weighs the result by this grade and states it on the face of the output.
Where it sits
A demonstrator under the Precision Toolkit.
PC4AI is the toolkit’s three-plate method run at scale: independent reads ground against each other, and the verified floor is only as high as the independence behind it. Keys live in the browser tab’s session storage, are sent straight to the named provider, and are never logged or stored server-side; you pay your own usage.