The Quorum — coordinated multi-model inference

neuralring/quorum convenes several pinned-eligible models in a structured Thinker → Worker → Verifier loop and stops as soon as a Verifier accepts the answer. The idea comes from TRINITY (Sakana AI, ICLR 2026), whose result maps directly onto the sovereignty constraint: a jurisdiction pin often leaves you a pool of good-but-not-frontier EU models — and most of Trinity's gain came from combining complementary models, not from picking one better model. NeuralRing v1 ships a deterministic, heuristic coordinator (explainable, zero extra dependencies) — not the paper's trained coordinator, so we do not claim its benchmark numbers.

The honest cost framing

A Quorum call is NOT cheaper than one single-model call — it dispatches multiple turns and bills each at its model's own rate, plus a named coordination fee.
The claim it can stand behind: early termination on Verifier-ACCEPT makes it meaningfully cheaper than naive ensembling (always-run-N, majority-vote, 5× self-reflection) at comparable or better quality. Trinity measured 2–10× fewer output tokens than MoA/Smoothie while outscoring them — Trinity's numbers, on Trinity's pool, until we measure our own (Phase 14.1).
Reach for it on complex reasoning / coding / math where one pinned-eligible model isn't enough. For everything else, a single model, neuralring/auto, or :nitro/:floor is the right tool.

Sovereignty + honesty guarantees

The pin is applied once, before pool formation; every turn — Thinker, Worker, every Verifier retry, every failover — selects only from the pinned-eligible set.
The attestation carries the full trajectory (every turn's endpoint, role, tokens) and its claimed tier/assurance is the conservative minimum across every endpoint used — never the best one.
If the turn budget runs out without an ACCEPT you still get the last Worker answer, but finish_reason, headers, and the attestation all say budget_exhausted — never dressed up as an accepted answer.
Billing reserves the worst case up front and refunds down to actual usage on termination (reserve → release → usage: three transparent ledger lines). Fewer than 2 eligible models degrades to neuralring/auto behavior (pool_too_small), never an error.

Usage

const res = await fetch("https://neuralring.eu/api/v1/chat/completions", {
  method: "POST",
  headers: { authorization: `Bearer ${KEY}`, "content-type": "application/json" },
  body: JSON.stringify({
    model: "neuralring/quorum",
    messages: [{ role: "user", content: "Prove that √2 is irrational, step by step." }],
    sovereignty: { jurisdiction: ["DE"] },
    quorum: { max_turns: 5 },   // optional; server ceiling 6
  }),
});
// x-neuralring-quorum-turns: 3
// x-neuralring-quorum-terminated-by: verifier_accept
// body.neuralring.quorum → the full explainable trajectory

Non-streaming only in this phase. Not yet available on /api/v1/messages (clear rejection). The trained coordinator (Trinity's actual method) is tracked as Phase 14.1 and will only ship with quality numbers measured on NeuralRing's own catalog.