JUDGE PANEL

Abstract

The judge panel is the layer that replaces human-in-the-loop approval with adjudication plus audit. A decision is posed as a question with its context and routed to a panel of models, most of them running locally, subset by which lens the decision needs and how much scrutiny it warrants. The verdict is computed from their votes, not self-reported by any one of them: approval takes agreement plus a high confidence bar, and a single well-argued dissent vetoes outright. Anything the panel cannot clear falls to needs-revision rather than sliding through. Every verdict that lands is appended to a single audit log with the question, the answer, which authority decided, and the reasoning.

Provenance

Came out of the least glamorous constraint in an autonomous factory: I was approving everything, so nothing moved faster than I could read. The directive I gave myself was blunt. Use judges instead of me for decisions, and log every one for review.

Declassified

The trade is the whole design. Delegation is only safe because it is logged, so the deal is that judges decide and I audit asynchronously, with the standing right to revert anything. Routine judgment (wording, prioritization, technical approach inside the guardrails) is delegated. The things guardrails can't cover (IP exposure, rule conflicts, anything irreversible or outward-facing) are surfaced, never silently decided. In the first month, 759 decisions were logged and roughly nine in ten were decided without me in the loop. This is honestly still in progress: the escalation tier is meant to shrink as the audit record earns it, and it hasn't shrunk much yet. The panel composition, the thresholds, and the routing taxonomy stay private.

Also in FORGE

• AARIA The orchestrator. Plan first, route always.
• SF TEAM A senior delivery team, expressed as agents.
• SYNARK The flight panel. What the factory is actually doing.