Parliament · The reasoning faculty

Many models.
One decision.

Multi-model deliberation for AI agents. The transcript is the artifact.

Parliament is the Reasoning faculty of the Ginnung cognitive runtime. Several models argue a question — proposing, critiquing, conceding, refining — until a synthesizer rules. The full transcript, with every objection and concession, is captured as a structured event. You don't just get an answer. You get the deliberation.

What is Parliament

Reasoning is a process,
not a token.

A single LLM call is a vote without a debate. Whatever the model happened to attend to wins. No record of the alternatives it considered. No way to know which arguments held weight.

Parliament runs the question through a configurable bench of models — GPT, Claude, Gemini, your own — each proposing, critiquing, and conceding under structured rounds. A synthesizer reads the transcript and renders judgment. When the bench can’t converge, you get a best-effort synthesis with the dissent attached, not a silent failure.

The transcript is the product. Every objection, every concession, every confidence score, sealed into a SonderEvent that the rest of your stack — memory, governance, audit — can replay.

What it does

Capabilities

Bench

Bring your own models

OpenAI, Anthropic, Google, OpenRouter, local. Configure the bench per question. No model lock-in on the reasoning layer.

Rounds

Structured deliberation

Proposal, critique, concession, synthesis. Bounded rounds prevent echo loops. Each turn is typed, addressed, and replayable.

Synthesis

Best-effort, never silent

On consensus, the synthesizer rules. On stalemate, you get the highest-confidence synthesis attempt with dissent attached — not a null.

Adversarial

Red-agent built in

An optional red agent argues the contrarian case every round. Surfaces failure modes the bench would have agreed past.

Transcript

The deliberation is the audit

Every turn lands in Engram as a SonderEvent. Replay any decision against the exact arguments the synthesizer saw.

Compose

A faculty, not a SaaS

Memory grounds the bench. Governance gates the action. Parliament is the reasoning step in a larger cognitive loop.

Why deliberation

The single-model failure mode.

One model, one answer, one confidence number. When it’s right, you ship. When it’s wrong, you have no way to tell the difference. The model that hallucinates an API and the model that quotes the docs verbatim sound exactly the same at the call site.

Deliberation makes disagreement visible. Three models that converge are evidence. Three models that diverge are a flag — the answer is load-bearing on a judgment call, and the call should reach a human or fall through to a safer default.

The cost is more tokens. The benefit is calibrated confidence and a transcript a reviewer can read. For the decisions worth getting right, that’s a trade you take every time.

Get started

Two ways in.

Self-host

github.com/heybeaux/parliament

TypeScript server, bring your own model keys. POST /deliberate with a question and a bench. MIT-licensed.

Compose

As a Ginnung faculty

Already running Engram or another Ginnung faculty? Parliament drops in on the same SonderEvent bus. No new SDK.

Built by

Parliament is the Reasoning faculty of Ginnung, the cognitive runtime built by heybeaux. Built around one bet: the most defensible thing an AI system can produce is a record of how it changed its mind.