Qe Court

QE-Court: Adversarial Review as a Verdict

Purpose

One reviewer — even a strong one — is one Einstein squinting at the chalkboard. QE-Court convenes independent adversaries from different vendors, with different roles and different probe sets, makes them attack the delivery, and then forces any SHIP verdict to survive an escalating deeper review before it stands. You stay in the loop as the final judge. Implements ADR-124; composes ADR-117..122.

The one rule that makes it a court and not a rubber stamp: a passing grade is the claim under attack, not the finish line. A shallow 91/100 SHIP that a deeper reviewer can overturn is a bug in the review, not a delivery that shipped.

When to convene the court

• Pre-merge gate on a risky PR or diff
• Release go/no-go ("deployment readiness with a jury")
• You got a PASS that felt too easy and want it stress-tested
• A test suite claims coverage you don't trust (does it kill mutants?)
• Any high-stakes "is this actually done?" call

For a quick single-lens review, use /sherlock-review, /brutal-honesty-review, or /code-review instead — the court is for when one opinion isn't enough.

The court roster (composes existing skills/agents — do NOT reimplement critics)

⚠️ Do NOT run the reduced 3-dimension QCSD workflow as the court engine. It can falsely rate SHIP by skipping the security / mutation / defect lanes. Spawn the specialized qe-* prosecutors above, or you reproduce the exact false-SHIP the court exists to catch. (ADR-124; see qcsd-development-swarm.)

Model routing (configurable — this is the point of the court)

The court's guarantees come from who reviews, not just how. Routing is user-configurable in config.json under routing; the defaults below enforce the invariants. Every model call goes through AQE's provider layer (ADR-123), so budget caps and cost receipts apply automatically.

Provider menu (mix freely in routing): claude-code (Claude subscription), cognitum-{low,mid,high} (Cognitum's own multi-model tiers — one option, it routes internally), openrouter (use when you want many distinct models across vendors for breadth), codex (GPT via ChatGPT subscription), claude/openai/gemini (metered APIs), ollama (local). Cognitum and OpenRouter are separate options: Cognitum already resolves multiple models behind its tiers; OpenRouter is the lever when you explicitly want to name several different models.

Enforced invariants (defaults; do not weaken without reason):

1. ≥2 distinct vendors across the panel — Claude / Cognitum / GPT-via-Codex — not just tiers.
2. Jury provider ∉ {writer, defense} — no model grades its own or its writer's output. Vendor is compared coarsely: cognitum-low and cognitum-high are the same vendor, so pairing them across defense and jury is a violation, not a diverse panel.
3. All calls routed through ProviderManager → ADR-123 budget cap + receipts.

Invariants 1–2 are machine-checked by validateCourtConfig() in referee.ts, which the court MUST call before seating a panel (see How to run it). The options block below binds directly to that check — minDistinctVendors and writerIsNeverJuror are read, not decorative.

The protocol

DELIVERY (diff / PR / test-suite / artifact)
   │  1. DEFENSE  — writer model states the case for shipping, from evidence only.
   │  2. PROSECUTION (parallel, blind) — N specialized reviewers, DIFFERENT vendors,
   │     each generates its OWN probe set and files CHARGES (finding + reproduction).
   │  3. KILL ROUND — blind refuters attack each charge; weak/unreproducible ones dropped.
   │  4. JURY — two-gate judge, cross-model, writer≠juror. 3-valued verdict.
   │     A numeric score is emitted ONLY if its rubric passed the ADR-122 ANOVA screen.
   │  5. OVERTURN ROUND — if verdict == SHIP, escalate ONE deeper reviewer. Loop-until-dry:
   │     SHIP only STANDS if K consecutive deeper rounds find nothing new. Surviving fatal → flip.
   │  6. SIGNED COURT RECORD — provenance-tier surviving charges (ADR-121), sign (ADR-118).
   ▼  HUMAN JUDGE (you) — rules SHIP / REMAND / BLOCK on the strongest case both ways.

Verdict states (three-valued — never a bare pass/fail)

Self-learning — the court feeds the flywheel (ADR-124 M0.B)

A verdict is not the end; it is training signal. After each court run:

1. Sign the court record → a flywheel receipt. Use the ADR-118 signer (src/learning/qe-flywheel/receipt.ts createSigner / platform-signer.ts, persisted via receipt-store.ts). The verdict + surviving charges are the body.
2. Each reproduced, surviving charge → a qe_pattern (ADR-110) at provenance tier oracle:test-exec (ADR-121 — it reproduced, so it is oracle-grade). Killed/refuted charges are NOT stored as positives (noise control).
3. An overturned SHIP is the highest-value signal there is — persist {shallow: SHIP, true: BLOCK|REMAND, charge} as a discriminator training pair for the frozen anchor (ADR-117) and two-gate judge (ADR-119). This is what makes the court harder to fool over time.
4. Retrieval-augment the next panel: seed each prosecutor's probe set with the HNSW-nearest prior charges for similar deliveries (qe_pattern_embeddings).

All writes are appends to existing stores — never destructive to memory.db.