Cad sim reviewer lenses

CAD Sim Reviewer Lenses

Own one integrated judgment for the current material candidate. Try to falsify readiness. Do not repair the design, fill expected answers, or confirm a claimed fix. If invoked directly on an active packet, route through $cad-sim-review-revise-loop unless the caller is explicitly the main reviewer.

Role And Evidence Boundary

• Treat summaries, claimed fixes, prior scores, and old routes as untrusted history. Inspect only the current canonical packet.
• Start from the current instruction and intake/project_brief.yaml.review_contract. A narrow complaint is focus, not a scope limit.
• Grade only artifacts named current by handoff.json and, when present, reviews/current_evidence_workspace.yaml.
• In source-blind review, the allowed packet is exhaustive. Source blindness covers tool use as well as files: do not inspect collaboration, thread, agent-list, task-status, parent-status, transcript, sibling, or prior-round state. The manager supplies procedural reviewer provenance.
• A fresh material candidate, complaint, round, or route judgment requires a fresh reviewer. Schema-only repair resumes the same reviewer thread, round, and file and cannot change engineering judgment without re-audit.
• Reviewer prompts may name keys, inputs, and validator errors. They must not prescribe semantic values, empty lists, conclusions, or a preferred route.

Before the deep audit, write the procedural review_status: in_progress checkpoint required by the launch prompt. Include no route, grade, pass claim, empty findings, or engineering conclusion. Publish the completed hidden pending sibling only through reviewer_lifecycle.py publish.

Audit Sequence

1. Confirm run, candidate, material-state hashes, source boundary, target maturity, approved stage, round, and reviewer UUID.
2. Read the review contract and only the current rubric, Quality Harness, evidence workspace, artifact registry, and comparison manifest that exist.
3. Open the registered primary user-facing artifact first, then every required_secondary; inspect the full surface and required runtime states.
4. Falsify current engineering and domain truth using source/model/result evidence before accepting prose, screenshots, scores, or page health.
5. Audit selected options against the strongest relevant challengers, governing scenarios, sensitivities, tradeoffs, and claim boundaries.
6. Simulate the builder/operator/analyst path, including likely misuse, failure, setup, service, recovery, and rejection states.
7. Separate route usefulness, target outcome, and grade. Identify remaining self-contained work, honest external boundaries, and the concrete next move.
8. Validate the same YAML with route_preflight.py --review-file <path> --review-contract-only before publication.

Fatal Engineering Gates

Apply every relevant gate; mark an irrelevant gate with a reason.

Architecture and configured behavior

For every master's-level review, include architecture_and_behavior_integrity. Check independent requirements and environment/reference geometry; continuous load, material, energy, signal, data, control, setup, service, and recovery paths; and a configured-state matrix of required contacts, clearances, forbidden contact, and interfaces. Check loads, self-mass, joints, contacts, materials, boundaries, scenarios, uncertainty, and materially different alternatives when exploration is in scope.

Apply the same gate with domain semantics for physical products, robotics, electromechanical, simulation, optimization, and software-connected hardware. An impossible assembly, forbidden interference, disconnected required path, self-referential proxy, or configured-interface contradiction is a fatal digital error. It cannot be demoted to a physical or external boundary.

For robotics_electromechanical, recompute the current hash-bound quality/system_interface_manifest.yaml. Local subsystem results and every declared interface combine with AND semantics; any schema, identity, evidence, or interface mismatch is fatal.

Current and release representations

Audit every release-relevant representation used by the next action. Native source, nominal JSON, matching hashes, or a successful export do not prove actual STEP, STL/mesh, slicer, manufacturing, controller, or runtime behavior. Use a verifier capable of falsifying the governing claim. Thresholds require a justified positive margin after tessellation, manufacturing tolerance, numerical uncertainty, and other applicable errors. A value exactly on the limit is not release-representation acceptance when a cheap margin or export-aware check is available. Distinguish nominal calculations from release-representation acceptance.

Quality, comparison, and evidence

When quality/quality_harness.yaml exists or is required, audit the V2 acceptance_profile, every mandatory row, verifier kind/target/expected, generated quality_harness_results.json, hashes, and primary-surface evidence promotion. Failed or missing mandatory rows block ready routing.

Emit quality_harness_used, quality_harness_audit, failed_or_missing_harness_rows, harness_evidence_promoted, and quality_reset_recommended.

For selection studies, compare the selected candidate, baseline, top two nonselected candidates on the primary objective, and every distinct scenario winner on a current reproducible common basis. Missing strongest challengers or stale/hard-coded comparison evidence blocks ready routing.

Primary surface and real action

Require exactly one primary user-facing surface and full-surface review of all primary and required_secondary artifacts. The website/report/manual must contain novice-critical procurement, tooling, build/use, evidence, rejection, and next-action information at the point of use; Markdown or audit sidecars cannot carry required user-facing proof. An interactive surface needs semantic runtime/state inspection or an adequate static fallback.

Emit explanatory_coverage_status with unexplained_user_salient_elements, overexplained_low_value_elements, and missing_story_or_decision_flow. Emit primary_surface_review_workspace_status with user_visible_next_move_use_artifacts, embedded_or_summarized_at_point_of_use, audit_only_artifacts, real_world_action_artifacts, and missing_point_of_use_evidence.

For prototype/build/use work, check sourcing, tools, fabrication, assembly, installation, calibration, inspection, first test, stop/reject gates, service, and physical/material/certification boundaries against current artifacts.

Design consequence and calibration

Require enough common-basis evidence to justify important architecture, mechanism, tool, model, and option decisions. Parameter or styling variants do not establish system design-space maturity. If a compact autonomous comparison, verifier, simulation, or evidence promotion could materially change trust, continue rather than stop.

For A+ or master's-level work, apply masters_engineering_integrated_v1, every mandatory floor, automatic cap, and submitted-now evidence maturity. Derive compact governing_analysis_rows when absent; do not edit or migrate the brief. Always emit route_usefulness, target_outcome_status, and governing_analysis_rows separately. A useful bounded packet may remain not_met_boundary.

Emit exemplar_calibrated_handoff_status, reusable_reviewer_tests, and integrated_exemplar_judgment when A+ handoff, design comparison, fuzzy dissatisfaction, or near-ready primary-surface quality is active.

Route Rules

• continue_autonomous: useful approved-stage self-contained engineering, evidence, comparison, verifier, or primary-surface work remains.
• revise: a bounded correction is required before fresh review.
• promote_maturity: a pre-approved next stage should begin.
• needs_user: a targeted choice, approval, measurement, credential, purchase, physical result, external execution, or broader stage controls progress.
• ready or bounded-ready: every applicable contract passes, no high-value self-contained work remains, and the next real action is useful and honest.
• blocked: a precise repeated workflow/tool failure prevents progress.
• max_revisions_reached: the revision budget ended with unresolved findings.

Never hide a current digital contradiction behind an external boundary. Never route ready with failed mandatory floors, stale evidence, unresolved lens gaps, unreviewed required artifacts, missing strongest challengers, or open_self_contained/open_tooling governing rows. A ready or needs_user route requires a concrete user_visible_next_move, not an instruction merely to open or read side files.

Conditional References

Do not preload all references. Inspect the domain, maturity, active complaint, artifact registry, and current evidence, then open only direct triggered files:

• output schema and canonical YAML: references/review_output_contract.md;
• active rubric, evidence promotion, profile-critical falsification: references/active_review_rubric.md;
• primary surface, runtime, workspace, and consolidation: references/primary_outcome_surface.md;
• build/use and real-world failure: references/assembly_and_prototype_handoff.md;
• decision and alternative closure: references/design_decision_evidence_closure.md;
• exemplar/adversarial reader tests: references/exemplar_calibrated_handoff.md;
• next-action worthiness and comparative fidelity: references/next_action_worthiness.md;
• spatial/state verifier evidence: references/verifier_backed_spatial_evidence.md;
• selected domain profile: references/rubric_profiles.md;
• master's grading: references/master_engineering_handoff_rubric.md and references/submission_grade_calibration.md.

The optional launch-side search index is cad-sim-review-revise-loop/references/reviewer_prompt_lens_catalog.md; it is only a router to these authoritative references.

Completion

Write one evidence-backed reviews/review_round_NNN.yaml for the current material candidate. Use the smallest applicable output from references/review_output_contract.md, integrate or reject any narrow sub-review findings, validate contract-only, and publish atomically. Stop after the route-bearing judgment; do not author a revision or create another round.