tribunal

Tribunal

The orchestrator builds nothing and scores nothing itself — it slices the work, dispatches a separate doer, convenes separate verifiers, and adjudicates on evidence. Three behaviors carry the value: an adversarial second look that catches what a single pass ships silently, as named failure scenarios; calibrated verdicts — locally fixable defects get ITERATE, not "rewrite everything", and passing math is overridden only by verified evidence; evidence before claims — nothing is "done" without fresh reproduction; refuted claims are excluded, never averaged in. Announce at start: "Running this through the tribunal pattern: doer -> verifier panel -> consensus."

When to use

Multi-part or high-stakes deliverables where a shipped defect costs more than a panel. Not for trivial edits, anything one verification command proves, or ordinary code review (dialogue improving a change vs independent measurements of a frozen artifact against pre-declared criteria, adjudicated to a ship decision). Never nest tribunals — no role runs the protocol on its own output.

The loop

1. Freeze acceptance criteria and verification commands BEFORE implementation; derive scoring dimensions, weights, and a pass target from them, recorded so re-panels reuse the rubric. Identify the operative skills in play — those bearing on how the artifact is built or judged (e.g. production-grade); never the tribunal skill itself nor pure orchestration skills — and fold their standards into the criteria. Verify-only entry (artifact already exists): write criteria from the original request — never reverse-engineered from it — freeze, start at step 3.
2. Spawn a SEPARATE doer agent (never the orchestrator itself) with the full slice spec pasted in (never "read the plan file") and the operative skills named with an instruction to load them (load production-grade, etc.) — if it cannot load a named skill it says so rather than proceeding. If the orchestrator writes or edits the deliverable, there is no independent artifact to verify and the run collapses to one context. The doer implements, runs the verification commands, materializes the artifact durably — committed, pushed, or published so it outlives the doer's session — and reports a diff summary, verbatim output, the artifact's fetchable address, and exactly one status (table below). DONE without an address is not DONE.
3. Check the report against the actual diff yourself; dispatch the panel in parallel, context-walled.
4. Adjudicate per consensus-mechanics.md; record verdict, dissents, caveats, and round count in the ledger.
5. SHIP -> next slice. ITERATE -> findings become a fix list for a fresh doer; re-panel with fresh verifiers re-scoring every dimension, prior findings as risks.

Invariants (low freedom — these are the skill)

1. Context wall. Four roles, four separate sessions: the orchestrator (slices, dispatches, adjudicates, owns the ledger — produces NO deliverable content and assigns NO scores itself), the doer, and each verifier. One agent never plays two roles (shared context = shared blind spots = no triangulation, the entire point). The doer finishes first; then the verifiers run in parallel, each RECEIVING exactly: frozen criteria; the artifact by fetchable address — a commit SHA, branch, or URI a verifier in a different sandbox can retrieve on its own, plus the diff (or the document and its predecessor) inline; bare working-tree paths are insufficient, since they resolve only inside the doer's session and die with it; reference materials; permission to run the verification commands; known risks; the operative skills (named, to load). Verifiers NEVER receive: the doer's reasoning or self-assessment, design rationale, each other's first-round views or identities, expected or prior scores. Naming shared standards is not a wall breach — the wall withholds the doer's reasoning and scores, not the bar everyone is measured against.
2. Evidence anchoring. High scores require a verbatim quote from the artifact; failure claims require file:line or command output; evidence-free findings are discarded. Before any consensus math the orchestrator greps every verdict-driving citation against the artifact and spec: a quote that cannot be located verbatim discards its finding; factually refuted scores are EXCLUDED and logged as refuted dissent. Adjudicate on evidence — never average disagreement away.
3. Explicit opposition. At least one panel member must oppose: build the strongest case against shipping as NAMED failure scenarios (which input, what wrong behavior). Attack vectors: hidden assumptions; over-engineering hiding bugs; missing constraints; drift risk; spec gaps; composition fragility. Generic skepticism is worthless — discard it. Confirm every accusation against the spec text first; never flag behavior the spec mandates. An unrebutted correctness or safety scenario from any member is never silently overruled: surface it (ESCALATE).
4. Criteria precede the artifact — frozen before implementation, or from the original request in verify-only mode.
5. No completion claims without fresh verification output. "Should pass" is not evidence; agent-reported success is checked against the actual artifact and diff.
6. Bounded iteration. Max 3 panel rounds per slice (fresh verifiers each round) plus a separate doer-dispatch budget (default 5, complexity-weighted); whichever exhausts first -> ESCALATE to the human with the full evidence package. Both counters travel in the handoff or a durable ledger, never in orchestrator context alone — every dispatch states the round index and the remaining budget. A cap an orchestrator restart or context reset forgets is not a cap: the run re-panels indefinitely.

Calibration: BLOCK is structural. Many severe defects that are each a localized fix in a sound architecture make ITERATE, not BLOCK — severity sets fix priority, never the verdict class. A blocked dimension (majority at 3 or below) never lets a passing overall slip through as SHIP: with localized causes it is ITERATE plus a mandatory fix list, even when overall ≥ target; only structural causes reach BLOCK. Consistency check before issuing BLOCK: if the report itself describes every defect as locally fixable, BLOCK contradicts the evidence — issue ITERATE. One member's low score is a logged dissent, not a block.

Panel composition (derive per artifact)

Independent vantage points triangulate ground truth; the rest is sizing. Default three lenses: quality (spec compliance first — missing / extra / misunderstood — then correctness), fitness-for-purpose (real consumer, edge cases, integration), and the adversary (must oppose). Scale to stakes — two reviewers for a small low-risk slice, five lenses for a high-stakes irreversible one — picking lenses from the artifact's actual risk surface. Example: a data-migration script risks data loss and partial failure -> transform-correctness, a rollback/idempotency adversary, operator experience; an API doc -> consistency, consumer ergonomics, a breaking-change adversary. Never give the adversary a weaker model than the rest of the panel; when available, a different model family on the panel buys uncorrelated blind spots.

Verifier prompt skeleton (write each prompt from this)

Role: <lens> verifier. You have NOT seen the builder's reasoning or any other
reviewer's views. The report you judge may be wrong — verify everything against
the artifact itself, as-is.
[Adversary only: You MUST oppose — build the strongest case against shipping;
name the exact failure scenario per concern. Write "ESCALATE: <reason>" for a
correctness/safety concern you believe cannot be rebutted.]

Records

The deliverable is the ONLY durable artifact a tribunal run produces. Report the verdict, dissents (incl. refuted, verbatim), caveats (fixed | deferred + reason), and panel-round counts in your closing summary — and fold them into the PR description or commit message when one exists. Delivering that report to the requester, over the channel the request arrived on, is a step of the loop, not a side effect of finishing — a verdict nobody received is not a verdict. Never leave loose files in the tree. If a run genuinely needs working files (a running ledger across many slices, temp fixtures), put them ALL inside a single gitignored .tribunal/ directory and never commit it; nothing else is written. Never silently drop a caveat.

Plan vetting (same machinery, earlier)

Before executing a non-trivial plan, run it past three independent lenses: will it work (are the checks specific?), can the doer follow it, is it the cheapest sufficient path. 2-of-3 approval proceeds; otherwise merge the critiques into exactly ONE refinement pass, re-vet once, then escalate. Vetting freezes the criteria the panel later audits.

Platforms

With parallel subagents, dispatch the panel as parallel calls in one message — separate agents, real independence. Without parallel dispatch, run each role as its own sequential fresh-context session (still distinct agents), enforcing the wall by what each session receives. One agent playing every role forfeits the triangulation benefit — measured equal to no panel at all (a weak model self-simulating a panel scores at its solo floor); use only when no separate-session option exists, and label the verdict "single-context (no independence)".

Detached orchestration (the run outlives, or is suspended by, any single turn): where the runtime's wait primitive ends the orchestrator's turn, whatever is passed to that primitive is bookkeeping, not communication — nobody has been told anything yet. Carry the artifact address, round index and remaining budget in the handoff, since the resumed orchestrator may not be the one that wrote them. On resume, re-establish what the requester has actually received before deciding the next move: a run ending is not delivery.

End-to-end example: worked-example.md. Failure catalogue: anti-patterns.md.](streamdown:incomplete-link)