AGENT SKILL

verify-against-rubric-fleet

Orchestrator-workers generalization of verify-against-rubric: parallel stateless sub-agents judge one artifact each against ONE locked rubric, merged into VERIFICATION_LOG.md. Use when N >= 4 artifacts share one rubric (multi-slice retrospective, batch PR review). Do not use for single artifacts (use verify-against-rubric), when rubrics differ per artifact (run separately or refactor the rubric first), or on LOW/MEDIUM complexity slices (use self-critique-and-revise).

License 许可
MIT
所属仓库
Mozurok/fhorja.dev
最近更新
2026-10-10
Stars
★ 10
Forks
0

01 前言 Front Matter

name
verify-against-rubric-fleet
metadata
{ "tools": "Read, Write, Edit, Bash, Glob, Grep, Agent", "workers": "[{\"role\":\"rubric-verifier\",\"dispatch-role\":\"judgment\",\"contract_ref\":\"commands/_shared/worker-contract.md\"}]", "category": "planning-and-validation", "max_fanout": "9", "provenance": "first-party", "convergence": "{\"pattern\":\"barrier\",\"timeout_ms\":\"600000\",\"partial_ok\":\"true\"}", "orchestrator": "true", "merge_strategy": "union", "x-wos-profiles": "full", "suggested-model": "claude-opus-5-5", "multi-repo-aware": "false", "primary-cursor-mode": "Plan", "worker_input_schema": "{\n \"type\": \"object\",\n \"required\": [\"artifact_id\", \"artifact_path\", \"rubric\", \"criteria\"],\n \"properties\": {\n \"artifact_id\": {\"type\": \"string\", \"pattern\": \"^[a-z0-9-]+$\"},\n \"artifact_path\": {\"type\": \"string\"},\n \"rubric\": {\"type\": \"string\"},\n \"rubric_source\": {\"type\": \"string\"},\n \"criteria\": {\n \"type\": \"array\",\n \"minItems\": 1,\n \"items\": {\n \"type\": \"object\",\n \"required\": [\"criterion_id\", \"criterion\"],\n \"properties\": {\n \"criterion_id\": {\"type\": \"string\"},\n \"criterion\": {\"type\": \"string\"},\n \"threshold\": {\"type\": \"string\"}\n }\n }\n }\n }\n}\n", "worker_output_schema": "{\n \"type\": \"object\",\n \"required\": [\"status\", \"artifact_id\", \"overall_verdict\", \"per_criterion\"],\n \"properties\": {\n \"status\": {\"enum\": [\"satisfied\", \"needs_revision\", \"max_iterations_reached\", \"failed\", \"interrupted\"]},\n \"artifact_id\": {\"type\": \"string\"},\n \"overall_verdict\": {\"enum\": [\"satisfied\", \"needs_revision\", \"failed\"]},\n \"per_criterion\": {\n \"type\": \"array\",\n \"minItems\": 1,\n \"items\": {\n \"type\": \"object\",\n \"required\": [\"criterion_id\", \"verdict\", \"reason\"],\n \"properties\": {\n \"criterion_id\": {\"type\": \"string\"},\n \"verdict\": {\"enum\": [\"satisfied\", \"needs_revision\", \"failed\", \"n/a\"]},\n \"reason\": {\"type\": \"string\", \"maxLength\": 240}\n }\n }\n },\n \"subagent_identifier\": {\"type\": \"string\"}\n }\n}", "context-layers-consumed": "memory", "context-layers-produced": "memory" }
description
Orchestrator-workers generalization of verify-against-rubric: parallel stateless sub-agents judge one artifact each against ONE locked rubric, merged into VERIFICATION_LOG.md. Use when N >= 4 artifacts share one rubric (multi-slice retrospective, batch PR review). Do not use for single artifacts (use verify-against-rubric), when rubrics differ per artifact (run separately or refactor the rubric first), or on LOW/MEDIUM complexity slices (use self-critique-and-revise).

03 SKILL.md 正文

Output contract, in brief. This body is over the per-skill re-injection cap, so after a compaction the sections below are truncated away while this summary survives. They remain authoritative in full; re-read this file before emitting if you need them.

  • Standard output layout (required): Produce the command output using this structure (English only):
  • Artifact changes: Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file...
  • Command transcript: Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).
  • Handoff: Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text b...
  • Definition of done (command output): Rubric is locked and discrete (>= 2 independently verifiable criteria).

Act as a senior/staff engineering orchestrator dispatching N stateless rubric-verifier sub-agents (one per artifact) and synthesizing their independent verdicts into a single VERIFICATION_LOG.md cohort entry with failure clustering.

Goal: For cohorts of N >= 4 artifacts that share ONE locked rubric (multi-slice retrospective, batch PR package review, design system spec audit), dispatch N stateless judgment workers in parallel. Each worker's evaluation context is ONLY its assigned artifact + the rubric; the artifact ID, output schema and selected-carrier envelope supply delivery metadata (no TASK_STATE.md, no DECISIONS.md, no sibling artifacts, no prior conversation). Workers return structured per-criterion verdicts independently. The orchestrator merges into one VERIFICATION_LOG.md cohort entry with: per-artifact verdicts, aggregate pass/fail/needs_revision counts, AND failure clustering -- which criteria fail across multiple artifacts (signal: likely a system-level issue not an individual artifact issue). Closes the same-context bias gap at cohort scale (Anthropic Outcomes 2026-05-06; the +10pp uplift compounds across N artifacts since each verification is independent).

Mandatory context bootstrap (before any output):

  • Read these sections in WORKFLOW_OPERATING_SYSTEM.md first:
    • ## LLM execution contract
    • ## Editor mode policy (mode definitions only; the tool mapping table is lazy-loaded in wos/editor-mode-mappings.md and needed only for non-Claude-Code tools)
    • ## Global output contract (including Adaptive handoff and Mode selection rule)
    • ## Cross-cutting workflow guardrails
  • Bootstrap tiers: the light-weight commands (branch-commit, what-next, where-we-at, slice-closure, compact-task-memory) plus the high-frequency implement-approved-slice and sync-task-state (v3 wave1 item D) read the four sections above with two subsections of ## Cross-cutting workflow guardrails skipped: ### External web access (centralized) and ### Sequencing heuristics (by phase). Everything else is read at every tier, including ### Proposal vs approved persistence and ### Substrate peer ownership (per ADR-0034), since all seven write substrate sections and reason about PROPOSED (state-reconcile stays on the full tier for cross-artifact judgment). The full tier is measured at 12109 tokens, the four always-read sections combined; the two skipped subsections are 1,034 of those (measured 2026-09-28), so the reduced tier is about 11,074. The leaf-reviewer tier (verify-against-rubric, ADR-0226) reads only ## Global output contract, measured at 5215 tokens, plus its rubric.
  • Session bootstrap reuse (skip-if-unchanged; v3 wave1 item D): WHEN this conversation already read the bootstrap sections in an earlier turn still VISIBLE in the context window AND WORKFLOW_OPERATING_SYSTEM.md has not changed since, the command MAY skip the re-read and cite the earlier one, emitting one Command transcript line: Bootstrap: reusing turn <N> read, WOS unchanged. Scoped exception to the context-budget re-fetch rule (wos/context-budget.md, "The re-fetch rule"), because these sections are one large, static, byte-identical read repeated every turn; every other tool result still re-fetches. VISIBLE means the section text itself is still present and quotable now, not merely that a record of the earlier read exists: a harness that clears a tool result while the record survives (ADR-0114) has not satisfied VISIBLE, and self-declared memory after a compaction never qualifies. A stateless-per-turn harness is excluded. The transcript line is mandatory; a silent skip is invalid output.
  • Resolving the spec and wos/<topic>.md: canonical repository FIRST. Otherwise use active-adapter docs: ~/.claude/workflow-docs (CLAUDE_WORKFLOW_DOCS_DIR), ~/.cursor/workflow-docs (WORKFLOW_DOCS_DIR; project .cursor/workflow-docs on project installs), ~/.codex/workflow-docs (CODEX_WORKFLOW_DOCS_DIR), or <kimi-home>/workflow-docs (KIMI_WORKFLOW_DOCS_DIR). Nonempty overrides win. Kimi home: KIMI_CODE_HOME, else existing ~/.kimi-code, else existing ~/.kimi, default ~/.kimi-code. Use one root for both paths; try other adapters if absent or unknown. Report resolved root, fallback to other adapters, or missing mandatory context in ### Command transcript. Installed snapshots refresh and prune retired topics only on sync.
  • Read additional sections only when relevant to this command's role.
  • Align all routing recommendations and next-command suggestions with the current command set.
  • Official next-command names only: recommendations, including Run now, use registered names from commands/<name>.md or commands/<name>/SKILL.md; installed mode uses the available Fhorja command or skill inventory. No invented aliases. Run now: none with Mode: N/A ends only if no honest next step exists, defined under ### Official command names (routing integrity) (ADR-0126). Required inputs:
  • task folder path
  • artifact manifest: array of {artifact_id, artifact_path} per artifact (caller pre-enumerates; orchestrator does NOT auto-discover artifacts)
  • rubric source: EITHER inline rubric (verbatim) OR reference to a section in IMPLEMENTATION_PLAN.md or DECISIONS.md (path + section anchor)
  • optional: explicit criteria list (overrides the criteria extracted from rubric; useful when rubric prose mixes criteria with prologue)
  • optional: explicit max_fanout override, lower only (defaults to 9, the review and verification cap in WORKFLOW_OPERATING_SYSTEM.md ## Parallel workflow, ADR-0236; never above 9 here)
  • optional: cohort_label (string used in VERIFICATION_LOG.md entry header; defaults to date + N artifacts)

Task repository files to update:

  • VERIFICATION_LOG.md (orchestrator is the SOLE writer per wos/substrate-peers.md; one cohort entry appended per fleet run)
  • TASK_STATE.md ## Last completed step per the canonical 5-section write pattern in commands/_shared/task-state-slice-closure-pattern.md
  • .wos/fleet-inbox/<run_id>/ directory (gitignored; one JSON payload per worker: native return or orchestrator copy of the runtime result)
  • .wos/VERIFICATION_LOG.jsonl (one line per merged section + one per per-worker classification event; distinct from VERIFICATION_LOG.md which holds human-readable verdicts)

Operating rules:

  • Handoff: end with the adaptive ### Handoff block per WORKFLOW_OPERATING_SYSTEM.md ## Global output contract.
  • Step 1: Lock the rubric. Either accept inline as-is OR extract referenced section. Rubric MUST decompose into discrete criteria (each verifiable independently). If criteria list not provided, parse the rubric: each bullet/numbered item -> one criterion with auto-generated criterion_id (c1, c2, ...). Refuse with NO_OP_TRACE if rubric is vague ("looks good", "feature works") -> route to resolve-contract-gaps first.
  • Step 2: Validate manifest. Confirm N >= 4 (else NO_OP_TRACE: route to verify-against-rubric per-artifact). Confirm every artifact_id unique. Confirm every artifact_path exists. WHEN N > max_fanout, split the cohort into sequential sub-batches of at most max_fanout and run them in order, stating how many sub-batches the overflow produced (ADR-0201). This is the review and verification fan-out of WORKFLOW_OPERATING_SYSTEM.md ## Parallel workflow (ADR-0236): read-only judgment workers, at most 9 at once, and this command the only writer of VERIFICATION_LOG.md and every task artifact; a worker writes only its one assigned return file.
  • Step 3: Verify prerequisites. VERIFICATION_LOG.md exists (or create with header from convention); rubric has >= 2 criteria (single-criterion rubric collapses to a trivial pass/fail; route to verify-against-rubric instead).
  • Step 4: Verify the orchestrator constraint. Confirm this command's suggested-model resolves to a model at or above the model the workers' dispatch-role resolves to in wos/model-routing.md ## Dispatch roles (wos/sub-agent-orchestration.md ## Role-aware dispatch protocol); refuse to dispatch if it does not. Workers are judgment: each returns a verdict per criterion, and a verdict is never routed down. This command's suggested-model is Opus-class so it stays at or above them; the failure clustering and aggregation it runs itself are rule-based.
  • Step 5: Dispatch workers (the load-bearing step). For each artifact, invoke a stateless sub-agent (Claude Code Agent tool with subagent_type: general-purpose). Each worker MUST receive ONLY: its artifact ID and path (read-only), the rubric verbatim, the criteria list, the output schema and selected-carrier delivery metadata. The assigned return file is its only write permission. NO TASK_STATE.md, NO DECISIONS.md, NO sibling artifacts, NO prior conversation history. Per ADR-0033 isolation rule. Per ADR-0038 Rule 1 (structured output mandatory), the worker returns a typed payload matching worker_output_schema and never free-form prose. Name the path you are on before dispatch (ADR-0158 D-1). Resolve <task_root>/.wos/fleet-inbox/<run_id>/<worker_id>.json for each worker and create the run inbox. On the dynamic-workflow path, declare agent(prompt, {schema}) with worker_output_schema; consume the runtime's typed result and persist its JSON copy at that path. Never tell the worker to call StructuredOutput. On the Agent path, pass the resolved path as fleet_inbox_artifact in the existing worker-contract envelope, outside task_input; the worker writes one schema-conforming JSON payload there and the orchestrator reads it. This return-file permission grants no other writes (ADR-0158 D-2). Supply the output schema and selected-carrier instruction outside task_input. End the dynamic worker prompt with Return one payload matching worker_output_schema and nothing else; end the native worker prompt with Write one JSON payload matching worker_output_schema to fleet_inbox_artifact and nothing else. The orchestrator validates the selected payload against worker_output_schema; the native file or runtime copy remains available for audit/replay.
  • Step 6: Each worker (instruction template). Worker reads the artifact; for each criterion, evaluates independently and assigns satisfied | needs_revision | failed | n/a with 1-2 line reason (<= 240 chars). Worker MUST NOT consult sibling artifacts, MUST NOT defer to "the cohort", MUST treat its artifact as if it were the only one. Worker assigns overall_verdict: satisfied (all criteria satisfied or n/a), needs_revision (one or more needs_revision, none failed), failed (one or more failed). Final-instruction reminder (ADR-0038 Open follow-up mitigation against schema-skip): the worker prompt MUST end with the selected-carrier reminder from Step 5 to defeat the 10-of-12 schema-skip failure mode observed on 2026-06-05. Worker returns {status, artifact_id, overall_verdict, per_criterion: [{criterion_id, verdict, reason}, ...], subagent_identifier} through the selected carrier. Free-form prose returns are CONTRACT VIOLATIONS and MUST be classified interrupted in ### Command transcript.
  • Step 7: Wait for convergence. Barrier pattern: wait for all N workers OR timeout_ms (10 min default). Read each expected worker's validated runtime result or assigned <task_root>/.wos/fleet-inbox/<run_id>/<worker_id>.json according to the selected carrier. Count that result once; unrelated files or replay copies cannot stand in for it. Classify per commands/_shared/convergence-policy.md.
  • Step 8: Aggregate verdicts. Apply union merge: collect all per-artifact verdicts. Compute: count of artifacts satisfied / needs_revision / failed; cohort_overall_verdict (satisfied iff all N satisfied; otherwise needs_revision or failed by worst-of).
  • Step 9: Failure clustering (the novel deliverable). For each criterion, count needs_revision and failed verdicts ACROSS the cohort. Sort criteria by failure-rate descending. Cluster criteria with failure-rate >= 50% as SYSTEMIC (signal: rubric criterion likely points at a system-level issue, not per-artifact). Criteria with failure-rate < 50% are LOCALIZED. Emit the clustering as a separate output section AND in the VERIFICATION_LOG.md cohort entry.
  • Step 10: Write VERIFICATION_LOG.md cohort entry. Append a single entry with: cohort_label, date, rubric source, N, per-artifact verdict table, aggregate counts, failure clustering (SYSTEMIC vs LOCALIZED), recommended remediation routing per cluster. The orchestrator is the SOLE writer of this section. Every bullet MUST be anchored to an existing parent heading or list per ADR-0038 Rule 3.
  • Step 10.5: Scan substrate orphans (ADR-0038 Rule 3 enforcement). Immediately after the merge in Step 10 and BEFORE updating TASK_STATE.md, run scripts/scan-substrate-orphans.py VERIFICATION_LOG.md, with scripts/scan-substrate-orphans.py (resolved against the WORKFLOW ROOT, ADR-0218); its exit 2 means a named file was absent, never a pass. If exit code != 0 (orphan bullets detected between sections), the apply step is NOT successful: emit NO_OP_TRACE listing the orphan offsets in ### Command transcript, refuse to update TASK_STATE.md, and route to state-reconcile to repair the structural drift. Only on exit code 0 may the orchestrator proceed to Step 11. This closes the substrate-bullet-orphan failure class (wos/bug-classes/substrate-bullet-orphan.md, detector commit 5840755).
  • Step 11: Emit VERIFICATION_LOG.jsonl. One line per per-worker classification event plus one line for the merged cohort entry (event=fleet-merge, partials=[worker_id, ...], strategy=union, orphan_scan=passed).
  • Step 12: Update TASK_STATE.md. Per the canonical 5-section write pattern. Reference the cohort_label. Gated on Step 10.5 orphan-scan exit code 0.
  • Workers MUST NOT consult each other or share context. The independence is the load-bearing property: if workers see each other's verdicts, the +10pp Anthropic Outcomes uplift collapses (groupthink).
  • Failure clustering is INFORMATIONAL, not deterministic. A SYSTEMIC label is a hypothesis: the orchestrator MUST suggest a follow-up (typically decision-interview on the rubric itself or direction-adjust on the underlying spec) but MUST NOT auto-route.
  • Mixed rubrics in one fleet run are FORBIDDEN. Re-run per rubric. (For multiple rubrics on the same artifact set, run the fleet N_rubrics times; the rubric is the load-bearing dimension, not the artifact.)

Required output:

  1. Rubric source + criteria count
  2. Artifact count + manifest summary
  3. Dispatch summary: N dispatched, M satisfied, K needs_revision, L failed, P interrupted, T timed out
  4. Per-artifact verdict table (artifact_id, overall_verdict, count_failed, count_needs_revision)
  5. Aggregate cohort verdict + counts
  6. Failure clustering: SYSTEMIC criteria (>=50% failure rate) with affected artifacts; LOCALIZED criteria
  7. Recommended remediation routing: SYSTEMIC -> decision-interview on rubric or direction-adjust on spec; LOCALIZED -> implement-slice-complement per artifact (one per artifact with non-satisfied verdict)
  8. Path to updated VERIFICATION_LOG.md cohort entry + orphan-scan exit code (MUST be 0 to proceed)
  9. Recommended next command (typically decision-interview if SYSTEMIC clusters present, else per-artifact follow-ups; state-reconcile if orphan-scan failed)

Claim grounding (active epistemic humility)

Claim grounding (active epistemic humility). This block governs what you may assert and how you record it. It is keyed to the substrate section you are writing, not to which command is running, and it is INERT on any output that writes none of the claim-bearing sections below. Full contract and rationale: wos/active-epistemic-humility.md.

  1. When this applies. This block fires ONLY while you are writing a claim-bearing substrate section: TASK_STATE.md ## Current known facts, ## Risks to watch, ## Observations, ## Active files in scope, ## Canonical decisions; DECISIONS.md ## Locked decisions; IMPLEMENTATION_PLAN.md ## Current gaps, ## Risks and mitigations; IMPACT_ANALYSIS.md; EXTERNAL_RESEARCH.md; REFERENCES.md; or any section whose content is a statement a later command or a human decision will act on. WHEN your output writes none of these, this block imposes nothing: skip it and proceed. This is the D-13 inert clause; a fully-grounded or claim-free output pays nothing.

  2. The unit is the load-bearing claim. A load-bearing claim is one a downstream command or a human decision consumes. A passing aside is not load-bearing; a statement someone will act on is. Apply the rest of this block per load-bearing claim, not per sentence.

  3. Ground it or abstain. Before you assert a load-bearing claim, trace it to the enumerable grounded set: a captured REFERENCES.md entry, a file read in this session, command output actually seen, or a passing deterministic gate. A claim supported only by model memory is OUTSIDE the grounded set, including when you are right, because that support is not observable. WHEN a load-bearing claim falls outside the set, do NOT assert it: either investigate until it is grounded, or abstain per rule 6.

  4. Status records provenance, never confidence. WHERE you attach an epistemic status to a claim, the status names WHERE THE CLAIM CAME FROM: a REFERENCES.md entry title, a file path plus line, or the gate output it came from. It SHALL NOT express a degree of certainty. Do NOT add a confidence field, a numeric threshold, or a self-assessment prompt anywhere; a self-reported confidence signal is not a usable control signal (wos/active-epistemic-humility.md Part 1.3). A status whose referent slot is empty is read as UNKNOWN, not as a weak yes.

  5. Persisted claims carry the status; chat-only claims carry it when they route. Every load-bearing claim you write into a task-memory artifact carries its provenance referent, and that referent travels with the claim so a later command reads it too; do not drop it at the write boundary. A load-bearing claim that appears only in a chat-turn output carries a status only when it crosses the grounding boundary and triggers a route (an abstention, an escalation).

  6. Abstain as a routed continuation, never a bare refusal. WHEN you abstain, name the specific investigation that would settle the question AND route to the command that runs it (capture-references, code-locate, incident-triage, or the fitting one). A withholding that stalls the work is invalid output. Abstention is distinct from NO_OP: NO_OP means there is no work to do; abstention means there is work and the grounding to do it is missing.

  7. An unfired gate is not evidence. The absence of a fired check does not mean grounding existed. Do not read silence here as a pass.

Standard output layout (required)

Produce the command output using this structure (English only):

Artifact changes

Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file carries one of those three tokens, in Lean output too; a prose verb like "written" is not a label.

Command transcript

Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).

Handoff

Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text block with all four lines, Run now:, Mode:, Work complexity: and Reason:, on a stop and on a refusal too; the terminal form is Run now: none with Mode: N/A. A new Mode:, a model or a fresh session is never a stop, an offered choice is one, and Reason: names a role, never a model.

Definition of done (command output)

  • Rubric is locked and discrete (>= 2 independently verifiable criteria).
  • Every artifact in the manifest has exactly one verdict in the VERIFICATION_LOG.md cohort entry; missing artifacts explicitly classified per failure taxonomy.
  • Per-criterion verdicts grounded in worker output; main thread did NOT re-evaluate or override worker verdicts.
  • Failure clustering computed: SYSTEMIC (>=50% failure rate) and LOCALIZED (<50%) buckets emitted.
  • Workers received only their artifact and rubric inputs plus the artifact ID, output schema and selected-carrier delivery metadata; no task memory, sibling artifacts or prior conversation leaked into worker context.
  • Workers returned typed payloads through the selected ADR-0158 carrier; schema-skip (free-form prose) classified as interrupted and listed in ### Command transcript.
  • Per-worker structured payloads persisted in .wos/fleet-inbox/<run_id>/<worker_id>.json for audit/replay.
  • VERIFICATION_LOG.md has one cohort entry appended (never partial-merged with prior entries).
  • Orphan scanner (scripts/scan-substrate-orphans.py) exit code is 0 on VERIFICATION_LOG.md after the merge in Step 10.5 (ADR-0038 Rule 3 enforcement); non-zero exit blocks TASK_STATE.md update and routes to state-reconcile.
  • VERIFICATION_LOG.jsonl has one line per classification event + one per merged section (with orphan_scan=passed).
  • Worker contract violations (mid-flight writes to VERIFICATION_LOG.md, sibling-artifact reads, schema-skip) explicitly listed in ### Command transcript.
  • Substrate peer rule respected: VERIFICATION_LOG.md is owned by this command in fleet mode (single-artifact verify-against-rubric retains co-ownership for append-only sequential entries; mixed-mode is reconciled by state-reconcile).
  • Recommended remediation routing distinguishes SYSTEMIC (rubric or spec issue) from LOCALIZED (per-artifact fix) and routes accordingly.
  • Before declaring this output done, confirm it satisfies the shared Definition of done (command outputs) and Gate conditions in WORKFLOW_OPERATING_SYSTEM.md.

Quality bar: J.10 orchestrator pilot. The independence of workers is the load-bearing property: if the orchestrator inadvertently leaks the cohort's running tally into a worker prompt (e.g., "5 of the previous artifacts failed criterion C3; check C3 carefully"), the Anthropic Outcomes +10pp uplift collapses into groupthink. Workers are stateless against each other AND against the parent. Failure clustering is the novel deliverable -- it's the difference between "you have 20 artifacts to fix" (low value) and "criterion C3 fails on 14 of 20 artifacts; the rubric or the upstream spec is the root cause" (high value). When the clustering surfaces SYSTEMIC issues, the recommended remediation routes UPSTREAM (rubric, spec, decision), not downstream (per-artifact patches that would compound). Per ADR-0038 (Workflow tool as canonical parallel-orchestration primitive): Rule 1 binds workers to typed payloads through the ADR-0158 carrier (schema-skip is the 10/12 failure mode observed on 2026-06-05), Rule 2 keeps substrate writes sequential through the orchestrator's apply step, and Rule 3 binds the apply step to scripts/scan-substrate-orphans.py to prevent the substrate-bullet-orphan failure class (wos/bug-classes/substrate-bullet-orphan.md, bug class commit 615c6bb, detector commit 5840755) where bullets appended near section boundaries land between sections with no semantic parent. Step 10.5 is the load-bearing enforcement point -- it MUST run before TASK_STATE.md is updated, and a non-zero exit code MUST block the apply, route to state-reconcile, and surface the orphan offsets in ### Command transcript.

04 同仓库更多 Skills