| name | orc-judge |
| description | Judgment gates for the ORC ultra lane. One judge skill, one agent (orc-judge-opus-5-xhigh), three dispatch contexts: gate=analysis (after the analyst), gate=plan (after the planner), gate=implementation (after verify — the anti-miss-implementation fidelity + strict-quality gate). Returns a structured verdict (APPROVE | REVISE | ESCALATE) with anchored, consequence-cited findings; REVISE loops the author with a hard 2-loop cap and a convergence rule. Dispatched only from /orc-ultra — no slash command of its own, never in orc or orc-mini. The orchestrator dispatches the judge subagent — it never judges itself.
|
ORC-JUDGE (ultra-lane judgment gates)
A model-based quality gate on top of ORC's deterministic gates. The judge
handles what determinism cannot: scope-interpretation correctness, plan
soundness, implementation fidelity, and — at the final gate — ultra-strict
code quality. Deterministic gates ALWAYS run first; the judge never re-checks
what a string-compare catches.
Ultra lane ONLY. No /orc-judge command; the orchestrator dispatches
orc-judge-opus-5-xhigh at the three ultra gates.
Hard rules
- The orchestrator never judges — it spawns. One agent, three contexts
(
gate=analysis|plan|implementation).
- Deterministic first. A judge dispatch is only legal AFTER the gate's
deterministic checks passed (evidence spot-check + derivation lint /
coverage + graph checks / verify + traceability matrix + static analysis).
- The judge is read-only and blind to the author's internal reasoning or
self-assessment (anti-anchoring): it sees the artifact, its evidence, the
advisor brief + rubric, and the original request.
- Judge gates ADD to user sign-offs, never replace them. Gate approvals
never skip the analyst challenge round or the plan sign-off.
- Gate scope is one-directional. gate=plan takes the approved spec as
fixed ground truth; gate=implementation takes the approved plan as fixed —
no gate re-litigates an earlier approved gate.
- The judge never fixes. Authors fix: analyst (gate=analysis), planner
(gate=plan), a scored executor fix wave (gate=implementation).
Verdict handling (orchestrator side)
Validate the return against the agent's verdict contract, then:
- Downgrade enforcement: any blocking finding missing its verbatim
anchor or its class-appropriate justification (failure_consequence for
correctness/security; named category + concrete alternative for
smell/simplification/placement) → downgrade to advisory and tell the user.
EXCEPTION: a security finding with a concrete consequence is always
blocking — never downgraded.
- APPROVE → record the verdict artifact, continue the pipeline.
- REVISE → re-dispatch the author with the original slice + the blocking
findings verbatim. The author's return must echo
finding_id → resolution
per finding; a missing echo is a malformed return (requeue). Then RE-JUDGE
under the convergence rule: block only on unresolved prior findings by id
or on lines the revision changed; new findings on untouched material are
advisory-only. Hard cap: 2 revision loops per gate, counted in the
checkpoint.
- ESCALATE (or a third failure) → present the ESCALATE menu:
accept-and-proceed (unresolved findings recorded in the Phase 7 summary) /
user fixes manually then re-judge (does not consume a loop) / grant one
extra loop / stop-and-checkpoint.
- Advisory findings NEVER loop — they ride forward attached to the next
phase's slices as notes.
- Persist every verdict as
run/{run-slug}/ultra/verdict-<gate>-<round>.md
(mined by /orc-retro; survives resume). When logging, emit
JUDGE <gate> <verdict> and the judgment GATE line (pass|bounce|escalate).
The three gates
| Gate | After | Judges | REVISE target |
|---|
| analysis | analyst-return deterministic gates | scope interpretation, requirement coherence, rubric misses, stale-doc risk | analyst |
| plan | Phase 1 exit gate + blast-radius map | decomposition, declared-file plausibility, deps, right-sizing, blast-radius coverage | planner |
| implementation | Phase 6/6.5 + traceability matrix + static analysis | fidelity to the user's ask + ultra-strict quality (security, smells, simplification, placement, pattern invariants) | scored executor fix wave → re-verify → re-judge |
Return fields include actual_model + actual_effort (standard
claimed-vs-actual check) and rubric_items_checked[] +
unconfirmed_assumptions_touched[] — validate them like any worker return.
