The brain spec · v0.1 draft

A brain that outlives its harness.

B‑Spec is a versioned, machine‑readable specification of what a self‑learning agent knows — its feedback ledger, its authority graph, its calibration record. It sits between the surfaces people talk on and the executors that do the work. Owned by neither.

Versionv0.1draft · public repo coming soon
First interface provenDiscordlive — a surface we didn’t build
Owned by a harness0the brain belongs to your team, not a tool

The stack

The brain is the layer in the middle.

Because the brain owns neither end, both ends are swappable — and the learning survives the swap. That is what makes this a standard rather than a product feature.

Buses become standards.

Interfaces where humans review

Work arrives and judgments happen where your team already talks. Any surface that can carry a four‑part answer and a reason field can be an interface.

Discord · liveSlack · nextTeams · nextOur workspace · in design
feedback in  ·  answers out

The brain the spec — diagnoses, prescribes, learns

Answers in four parts. Records every judgment with who gave it, on what, and why. Weights feedback by domain authority. Calibrates its confidence against what reviewers actually decided.

feedback ledgerauthority graphcalibration recordfour‑part schema
approved paths  ·  outcomes back

Executors tools that work for the brain

Once a human confirms the path, the agent drives execution tooling to carry it out — a Hermes setup updating a website, a hosted OpenClaw instance running a campaign. The executor works for the agent; the agent answers to your team.

HermesOpenClawYour APIs

Two surfaces

What the spec actually specifies.

Surface 1 · Feedback in

How an interface reports a judgment: which part of which answer, right or wrong, the reason given, and the identity and domain of the person who gave it. Approvals count as much as corrections — a system that only learns from failure learns to avoid mistakes, not to recognise good work.

Surface 2 · Prescription out

How a confirmed path is handed to an executor: the steps, what each touches, what it would change, what’s reversible — and how the outcome reports back. Execution results are the second supervision signal: the human judged the diagnosis; reality judges the path.

Why four parts — the mechanical answer. Part four is machine‑executable. A reviewer isn’t approving prose; they’re approving a plan that will run. Parts one to three exist so a person can check the reasoning before the path is compiled into actions. The shape has older provenance — it’s a 2,500‑year‑old diagnostic — but the justification here is engineering, not philosophy.

The draft

Read v0.1 before it freezes.

The spec goes public on a versioned repo at freeze — it lands there, versioned, as it stabilises. Until then we share the working draft with researchers, integration partners, and anyone building a harness or executor who wants the brain to plug in.

  • 01Feedback event schema — part, verdict, reason, principal, domain.
  • 02Authority graph — how per‑person, per‑domain weight is learned and stored.
  • 03Calibration record — how 98 is made to mean 98, and audited.