The settle is the consensus.
ljos vote ID --for OPTION records a ballot. One ballot per identity; --as NAME casts as a persona. A recast replaces. A missing identity is no.
ljos consensus ID is the seat verb: the model crate, then the tracker, both under the trust rows the pack holds that apply to the issue's topic and every persona's anchor. The model is DeGroot / Friedkin–Johnsen, or bounded confidence when the issue is tagged broad, or the Seldon ODE when --seldon and seldon are on PATH. Every outcome says how the group split, in polarization and disagreement; a count says only who won.
On a hard question add a forecast beside the ballot, ljos predict ID --expect OPTION; with two or more forecasts the settle also names the surprisingly popular answer, the option whose actual share most exceeds its forecast, and shows each voter's standing from the rows.
The rows are memory. ljos learn ID --outcome OPTION weighs each voter by its record's accuracy, shrunk toward the panel's pooled accuracy, keeps the outcome, and loosens the anchors of refuted personas; ljos calibrate -p PROJECT writes rows from the project's history when nobody names an outcome; ljos trust FROM TO W --about DOMAIN sets one by hand, scoped. ljos consensus discounts the voters who err together once five issues have an outcome. Under equal rows a settle is a count.
A panel is one subagent per persona, each started from ljos brief NAME ID, each casting one ballot as itself, then the settle. ljos persona NAME --anchor A --view TEXT --about DOMAIN writes the persona once; it stays in the pack.
Where the machine turned Jev on, ljos panel ID --jev asks Jev for each persona's ballot first. Jev returns what the ballot always carried and a subagent had to guess: the chosen option's probability is the confidence, and its forecast of the panel is the prediction, so the settle and the surprisingly popular reading need nothing new. Personas answered by one model are one correlated voter, so the panel casts only when every seat is sure (confidence 0.8 or above) and all agree. Otherwise it casts nothing and writes the briefs, and the subagents take the contested question. learn scores a Jev ballot like any other.
Four green ticks are a host rule for landing work. They are not the model. Do not report a count as if it were the crate.
The explanation is where a settle's claim comes from: the closed form the iteration solves for, the round count guaranteed for an anchored settle, the SymPy check and the interval enclosure under Derivations, proofs and certificates, and when to use each model.
consensus · ljos · seldon-code