Council ✦

Evidence-first multi-agent decision making.

Council is an AI decision system that checks reality before reasoning, forces agents to disagree before converging, and makes important recommendations falsifiable rather than merely confident.

The decision loop

tools
  ↓
evidence
  ↓
multiple specialist views
  ↓
forced counter-case
  ↓
resolution
  ↓
confidence + falsification test
  ↓
human-approved action

What it proves

• tool use is required when a decision depends on external reality
• evidence is graded instead of silently treated as equally reliable
• specialist agents contribute different decision lenses
• a critic is required to attack the emerging consensus
• a chief agent resolves disagreement rather than adding another vote
• final recommendations include confidence and what evidence would reverse them
• consequential external actions stay human-approved

Evidence grades

Grade	Meaning
A	tool-verified state
B	repository- or source-asserted claim
C	independently validated external proof
D	Council inference or hypothesis

That distinction matters: successfully reading a README should never turn a claim inside it into independent proof.

Example: portfolio decision

A request such as:

Check my projects, cluster related work, challenge the ranking and pick one priority.

runs through:

GitHub scan
   ↓
venture clustering
   ↓
Builder
   ↓
Capital
   ↓
Humanist
   ↓
Critic
   ↓
Chief of Staff
   ↓
falsifiable decision

Reliability behaviour

Council fails visibly rather than inventing missing reasoning. The runtime includes recovery for empty model outputs, bounded portfolio context and explicit tool/evidence gates.

API surface

• GET /api/status — runtime and capability state
• GET /api/portfolio?username=... — GitHub portfolio evidence + clustering
• POST /api/council — evidence-gated multi-agent execution

Tests

npm run check
npm run test:decision
node tests/github-smoke.mjs

The deterministic tests verify routing order, evidence requirements, Critic-before-Chief behaviour, recovery paths and explicit separation between observed evidence and inference.

Run locally

npm install
npm start

Requires Node.js 20+ and an OPENAI_API_KEY for live model execution.

Stack

JavaScript · Node.js · OpenAI API · multi-agent orchestration · tool calling · evidence gates · deterministic tests · GitHub integration

---

Built by Michael Ninh in Berlin.