NeuralRing

The Quorum

Coordinated multi-model inference over your pinned-eligible pool.

neuralring/quorum convenes several pinned-eligible models in a coordinated Thinker→Worker→Verifier loop, stopping when the Verifier accepts. It buys better quality per sovereignty-constrained token — not a cheaper call. The sovereignty pin is applied ONCE before the pool forms and binds every turn; the attestation carries the full trajectory and claims the conservative minimum tier/assurance across every model used.

Effort:

Pool preview

Which eligible models the engine would pre-select per function for this pin + tier. Computed for a general-domain request; a live request re-triages by the actual prompt.

Loading…

Test chat

Side-by-side compare

One prompt, two separately billed and attested requests: The Quorum vs a single-model baseline. Both costs are shown — no hidden double-spend.

Programmatic

curl
curl https://neuralring.eu/api/v1/chat/completions \
  -H "Authorization: Bearer $NEURALRING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "neuralring/quorum:fast",
    "messages": [{"role":"user","content":"..."}],
    "quorum": {"effort":"fast"},
    "stream": true
  }'
TypeScript
import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://neuralring.eu/api/v1", apiKey: process.env.NEURALRING_API_KEY });
const res = await client.chat.completions.create({
  model: "neuralring/quorum:fast",
  messages: [{ role: "user", content: "..." }],
  // @ts-expect-error NeuralRing coordination extension
  quorum: { effort: "fast" },
  stream: true,
});
Python
from neuralring import NeuralRing
client = NeuralRing(api_key=os.environ["NEURALRING_API_KEY"])
res = client.chat.completions.create(
    model="neuralring/quorum:fast",
    messages=[{"role": "user", "content": "..."}],
    extra_body={"quorum": {"effort": "fast"}},
    stream=True,
)

How it works

  • ▸Function-matched heterogeneity (X-MAS). Each role goes to the model measured best at that function (plan / generate / verify) — not by reputation, by our own probe scores.
  • ▸Quality-gated pools (Self-MoA). Weak members are dropped before diversifying: N samples of the best model beat a mix that includes a weak one. Diversity never excuses weakness.
  • ▸Verification-weighted selection. The answer is chosen by a verifier, not a vote — execution against tests where the task is checkable, else a rubric-structured judge that never grades its own author. Order is randomised to fight position bias.
  • ▸Adaptive tree search (AB-MCTS). The deep tier runs an adaptive tree search that decides per step whether to try wider or refine deeper, and which model to use, from the rewards seen so far — spending more only where it pays off.

What we deliberately DON'T do: naive multi-model debate (it fails to beat single-model self-consistency at equal compute) and naive LLM-as-judge (position/verbosity bias). We report selected@1 — what you actually receive — never pass@N.

Proof

selected@1 accuracy per suite and tier, from stored evaluation runs — never a hardcoded number. Empty until a run is published. A tier at or below the baseline is shown as-is (amber).

Loading…