| name | reviewing-cybernetic-control-structures |
| description | Use when an existing v1 requirements/design/goal/plan chain needs compatibility review. |
Reviewing Cybernetic Control Structures
Legacy v1 compatibility only. V2 reviews the commitment before approval and the
final evidence subject before completion, without reviewing strategy artifacts.
Overview
Review whether approved work can execute.
This skill does not execute target work and does not start /goal.
Detailed review dimensions live in references/control-review-detailed-rules.md
and references/review-rubric.md.
Stage-routing gate results follow ../references/transition-gate-protocol.md.
Independent Review Rule
Do not self-review and mark Approved.
Use authorized subagents, explicit human approval, or another independent
reviewer. Otherwise mark Needs Independent Review.
No Approved state is allowed after an unreviewed substantive artifact
mutation. If approved files change after review, mark Dirty or Needs Re-review until final independent re-review passes.
Semantic Review Verdicts
Subagent semantic review verdicts must be one of:
`Approved` / `NeedsRevision` / `Blocked`
NeedsRevision is not a simple fail. It means the chain contains repairable
intent drift, obligation downgrade, or artifact misrouting.
runtime compilation is forbidden until the verdict is Approved; route to the
returned stage, repair, and review again.
Non-passing checks persist invariant, typed owner, stage, counterexample, and
artifact hashes. Route by owner; return_to_stage is compatibility metadata
and mechanism suggestions are nonbinding. Owner taxonomy lives in the detailed
rules and schema.
Required Checks
Always check:
- requirement traceability;
- Intent Preservation / Obligation Preservation Review;
- source-requirement-preservation when
source_requirements exist;
- required outcome coverage;
- Counterexample Gate;
- design match when design exists;
- goal correctness;
- plan controllability;
- evidence governance;
- runtime suitability;
- review independence;
- final observer.
Use the detailed reference for the full rubric.
Intent Preservation / Obligation Preservation Review
Reject when implementation, measurement, repair,
diagnosis, or decision work is weakened into readiness, future work, allowed
action, compatibility-only behavior, a framework, or a plan.
Regression examples that must be NeedsRevision:
- required
/api/v2 implementation must not be accepted as legacy Drogon compatibility readiness;
- source request to measure E, S, A, M, Q, K, Se, Nout, Cckpt growth curves must
not become define scan framework and dominance rules;
- implement /api/v2 download/extract/preview API family must not become
compatible with future v2 exposure.
Route these to ReturnToRequirementsAnalysis if the source requirement itself
was weakened.
Counterexample Gate
Every approved review must include counterexample-gate. It attempts to
disprove the target decomposition and runtime claim.
The gate executes the requirements-approved counterexample_gate_contract.
Review must not invent, weaken, or replace that quality contract. If the
contract is missing for a controlled run, return to requirements analysis.
It must also cover each blocking outcome's per-outcome counterexample_gate.
It must include reviewer provenance: reviewer.kind, reviewer.id, and
reviewer.evidence_ref. Self-written evidence is not enough.
Required checked_transformations include at least:
source_requirements->required_outcomes
required_outcomes->required_steps
required_steps->work_packages
required_steps->runtime_steps
pre_runtime_compile
blocked_or_goal_achieved
Plan And Runtime Review
Each blocking required step needs a credible chain:
required step -> what would make it true -> producing action -> mainline work package -> evidence after action
Deterministic Lint
If scripts are available, run:
python3 .agents/skills/reviewing-cybernetic-control-structures/scripts/control_artifact_lint.py \
--requirements [REQUIREMENTS] \
--design [DESIGN] \
--goal [GOAL] \
--plan [PLAN]
Lint is structural evidence, not meaning approval.
Output Format
This output format is response-only. Do not write $skill ... commands,
runtime /goal prompts, or conversational next-step prompts into the review
artifact.
Created or updated review:
`docs/cybernetics/runs/YYYY-MM-DD-slug/review.control.json`
Review status:
- `Approved` / `Needs Revision` / `Needs Independent Review` / `Dirty` / `Needs Re-review` / `Blocked`
Review verdict:
- `Approved` / `NeedsRevision` / `Blocked`
Key findings:
- ...
Response-only next step:
- return to `$orchestrating-cybernetic-pregoal`, or
- revise the named artifact and rerun review, or
- report the smallest unresolved decision, dependency, or unavailable fact.
Validation Checklist
