Quorum‑AI
“What happens when four models disagree about your question?”
A costly run cannot start on a nod: the approval is a single-use, expiring token cryptographically tied — an HMAC, a keyed hash — to the account and the exact estimated cost, so a stale or borrowed approval is rejected with the estimate reattached. And agreement is never allowed to flatter: anything Quorum wrote itself — a simulated answer, a templated critique — is excluded from the evidence before consensus is scored, a rule added after a run once showed “4 of 4 aligned” while asking nobody.
GateCost approved before anything runs; fallbacks disclosed
SAMPLE — an end-to-end acceptance fixture rendering canned data (e2e/fixtures/golden-run.ts), not a measured run. Live execution is off by default. The revision count is inferred from position movement, and the interface says so.
- Not claimed
- Latency · accuracy · adoption
What it does, and what it refuses
Works now
- Four models answer in parallel; a separate moderator pass critiques them over two rounds
- The four never read each other — there is no code path on which they could
- Cost is approved before anything runs
- Any fallback or simulation is disclosed on screen, never hidden
Known gaps
- Round 2 is skipped when the per-run debate budget runs out, and recorded as skipped
- Live execution is off by default — the interface says the output is local simulation
- Consensus is lexical: 4-gram overlap and negation heuristics, captioned “inferred, not a tallied vote”
- The moderator’s default model id is the same as answer slot 2’s — a separate call in a separate role, using the same model by default
- The mutation gate is advisory in CI. Promotion to blocking was built, measured, and reversed
Not claimed
- Latency
- accuracy
- adoption
The eval judge is a paid call several threads may need at once, so the first thread makes it and the rest wait on that one future. A reader that times out marks itself, so a verdict-less result never gets cached over a run the judge did verify. Its production readiness review records a Go dated 21 June and the No-Go it superseded five days earlier. Both are kept in the file.
Discussion
Comments need JavaScript and a GitHub account — the same account system the site's own code already lives in. Without either, read or start the thread directly on GitHub.