HOW THE COUNCIL WORKS

Two chambers.
One governed record.

Models may work concurrently inside their own chamber. Between chambers, execution is strictly sequential: no critic reads a solution until every solution is committed. The council produces adversarially reviewed argument, not machine-checked proof.

Solvers commit→The record is sealed→Adversaries begin

SUBMISSION

One canonical mathematical problem

Text, equations, definitions, assumptions, and up to six screenshots or images.

FIRST CHAMBER

Solver Chamber

Selectable privacy routing

Every Council model uses standard provider routing by default. Before convening, Zero Data Retention can be required independently for any model; models without a compatible ZDR route may become unavailable for that run.

01

Image Reading and Solution

sequential

When images are attached, the visual reader reconstructs the complete problem, records ambiguity, and produces the first visible solution.

Gemini 3.1 Pro

Primary image reading and complete solution

Optional ZDR
02

Visual Record Handoff

The reader’s complete visible output is supplied verbatim to the remaining solver seats.

  • No separate OCR or visual-intake model
  • Image ambiguity remains explicit in the reader’s record
  • Text-only submissions begin with parallel solving
03

Independent Commitment

parallel

Every solver commits a sealed solution with assumptions, numbered claims, dependencies, confidence, and weak points.

Gemini 3.1 Pro

Visual and geometric solver

Optional ZDR

GPT-5.6 Sol

Independent research solver

Optional ZDR

Claude Fable 5

Independent proof solver

Optional ZDR
  • On text submissions no solver sees any other seat’s work
  • On image submissions the reader’s published reading is the one exception, so the other seats can challenge the transcription
  • No solver, in either case, sees another seat’s criticism of its own work
04

Sealed Record

Candidates are recorded as committed. Nothing is revised under criticism, because no criticism has been written yet.

  • Commitment precedes critique
  • Dissent between solvers is preserved, not averaged
  • An unavailable seat is recorded, not silently dropped
05

The Seal Closes

barrier

A hard execution barrier closes the chamber. No adversary provider request has been dispatched before this point.

The SealSOLVER ───────────► ADVERSARY

Sealed record transmitted one way

SECOND CHAMBER

Adversary Chamber

Sealed record received

Seats are roles, not exclusive model assignments. A model may hold a seat in both chambers, and two do; each invocation is independent and stateless. What the seal guarantees is narrower than total blindness: no adversary reads the record until every solver has committed, and no committed solution is ever revised under criticism.

06

Record Received

sequential

The complete sealed record crosses in one direction. Adversary findings and the final verdict are never returned to the solver chamber.

07

Adversary Derivation

parallel

Each adversary seat derives the result itself and compares that derivation against the sealed record, identifying conflicts and unresolved gaps. The sealed record is in context while it derives, so this is a second opinion, not a blind reproduction — only the solver chamber works unsighted.

Grok 4.5

Unanchored adversary — committed no solution of its own to defend

Optional ZDR

GPT-5.6 Sol

Cross-examiner and final synthesis

Optional ZDR

Claude Fable 5

Proof and formalization critic

Optional ZDR
08

Adversarial Review

The chamber searches for the first invalid inference, missing hypotheses, counterexamples, specification errors, and unsupported conclusions.

  • Demonstrated errors are separated from gaps
  • No-error-found is a valid outcome
  • The process preserves disagreement
09

Evidence-Graded Verdict

barrier

A non-voting synthesis reports what survived and what remains unresolved. Truth is not decided by model majority.

  • Up to ten model calls per run: three solve, three derive, up to three review, one synthesis
  • Only adversaries that delivered a derivation are seated for review, so a failed seat lowers the count
  • The grade is argued from the record, not computed by an instrument

SCIENTIFIC COUNCIL RECORD

Evidence, not a vote.

AFormally certifiedBMachine-verifiedCAdversarially reviewedDUnresolvedERefuted

A and B are not currently issuable. They require a kernel-checked formalization and an exact-computation instrument, and neither is connected to this council yet. Every run today lands at C, D, or E, and the synthesizer is instructed never to claim otherwise. Adversarial review is not machine verification.

Delivered as readable Markdown and typeset mathematics with original source preserved for Markdown and JSON download.