Control plane / recorded proof

Operational proof, at a glance.

Evidence becomes a reviewed repair while the human authority boundary stays visible and inspectable.

Agent remediation is unauditable, so humans can't safely delegate without recorded proof.

Fixed execution order

Five model-driven seats, one explicit gate

Exact model identity, role, effort, and escalation policy remain visible.

  1. Prosecutor portrait
    Prosecutorgpt-5.6-luna

    Challenges the leading incident hypothesis

    Effort low · Escalations 0/2
  2. Inspector portrait
    Inspectorgpt-5.6-terra

    Builds the evidence-backed failure mechanism

    Effort medium · Escalations 0/2
  3. Counsel portrait
    Counselgpt-5.6-terra

    Proposes the smallest testable repair

    Effort medium · Escalations 0/2
  4. Magistrate portrait
    Magistrategpt-5.6-sol

    Returns the fail-closed incident verdict

    Effort medium · Escalations 0/2
  5. Bailiff portrait
    Bailiffgpt-5.6-luna

    Presents one approved warrant for execution

    Effort none · Escalations 0/2
01 / Genuine model runs10 genuine cases
02 / Recorded decisions10 human-approved
03 / Trusted outcomes10 verified repairs
Bounded control flow

Six steps from evidence to proof

Open an incident
  1. 01Sanitize evidence

    Raw bytes stay outside model context.

  2. 02Test hypotheses

    The leading mechanism meets a real rival.

  3. 03Propose repair

    Counsel emits a bounded patch, never commands.

  4. 04Review verdict

    The decision fails closed when proof is incomplete.

  5. 05Approve warrant

    A human reviews the exact hash-bound document.

  6. 06Verify execution

    Trusted tests derive success outside candidate code.

Human gate

Magistrate CLEAR can present a warrant; only a human can approve it.

Isolated runner

Candidate code cannot see control-plane secrets or declare its own success.

Published projection

Judges browse sanitized, explicitly published case snapshots only.