
Council
Council is my experiment in turning disputed questions into public, sourced, challengeable records. A claim, forecast, ranking, link, post, or argument goes in; Council normalizes the ask, routes it through a visible protocol, and publishes a record with the answer, evidence, model votes, dissent, cruxes, limitations, challenges, and version history attached.
The idea is not "chat with several models." It is closer to a public docket for AI-mediated judgment. The first read should be decisive, but the record underneath should stay inspectable and movable when better evidence arrives.

What It Does
Council takes broad input and tries to turn it into one of a few public record shapes: a claim verdict, a forecast, a comparison, a ranking, or a clarification state when the ask is too vague. Completed runs become shareable records instead of disposable answers.
Each record carries the answer first, then the accountability layer beneath it: confidence, split, source status, strongest support, strongest objection, main crux, what would change the verdict, challenge activity, and the protocol details behind the run.

Why It Exists
I am interested in AI interfaces that produce artifacts people can argue with. A raw model answer disappears into the chat scroll. A Council Record is meant to sit in public: sourced, versioned, challengeable, and clear about where the judgment could move.
The product tension is the fun part. It needs to feel calm and authoritative without pretending to be an oracle. It should help disagreement become more inspectable: not by forcing consensus, but by preserving the evidence, dissent, and update path in the same artifact.

The Protocol Surface
Council Protocol v1.5 is a working local vertical slice. The app has the public filing flow, confirmation step, live/replay Floor view, settled Record view, archive, challenges, reruns, version history, evidence panels, seeded records, SQLite persistence, Vitest coverage, Playwright coverage, and no-key fallbacks for local development.
The Floor view is where the record forms: model roles, transcript events, observer notes, source checks, provisional record fields, and final synthesis. The goal is for the process to be visible enough that the answer does not feel like a black box, while still ending in something someone can actually read.
