Quality Model

rd-skills quality is the combination of control correctness, professional judgment, bounded context, observable evidence, and maintainable packaging.

Control Correctness

• Direct Tasks begin after one classification.
• Analyzed Work receives one complete source-backed initial Analysis.
• Accepted decisions remain authoritative until new evidence invalidates a protected decision; then Delta Analysis covers only its transitive impact.
• The First Executable Slice is found before unnecessary plan breadth.
• Task completion, Task switching, ordinary discovery, and an unreached Review Boundary do not repeat Analysis.
• Writes are parallel only with safe host isolation and non-overlap.

Professional Judgment

• Every Task is a complete semantic change with one primary Professional Skill; mechanically different files, functions, layers, tests, or edit steps do not create separate Tasks.
• Work with materially different Primary Professional Skills remains separate even when Tasks share a combined Review Boundary.
• Implementation Layer 3 Skills are triggered by Task risks or technologies. Review assignments independently select zero to three Layer 3 Skills from review risk; review selection is not a Task-side union.
• Combined review preserves all required Review Skills, Specialist obligations, and professional-risk dimensions.
• Owner, invariant, placement, failure behavior, and validation are explicit.
• New material risk stops or escalates instead of silently widening scope.

Evidence Quality

• Validation runs after the latest material edit.
• Fresh, scope-correct validation with a trustworthy oracle is reused unless a concrete freshness, coverage, oracle, flake, environment, reviewer-doubt, or independent-reproduction trigger requires another run.
• Implementation review examines the actual diff and every changed file; artifact review examines the bounded artifact, decision criteria, and supporting source evidence.
• Effective Level sets review depth; minimum sufficient Review/Risk Boundaries set frequency. L1-L3 related scope defaults to one combined independent final review, L4 adds triggered specialist depth rather than automatic rounds, and L5 retains its pre-implementation and final requirements.
• A current Review Boundary carries its Boundary and Review Round IDs, strategy, Effective Level, required Review Skills, Specialist obligations, Covered Task IDs, changed scope, professional-risk dimensions, current validation binding, assignment schedule, and primary-close ordering. Exactly one primary and zero or more specialist review-agent assignments share the round; each assignment has one registered Review Skill and bounded review-risk Layer 3 selection. Specialists do not close Tasks or add rounds. The primary consumes their current results and emits one artifact referenced exactly by every covered Task projection.
• An ordinary material finding does not end its fixed Review Boundary or add scope. The reviewer completes required changed scope, base dimensions, and professional-risk dimensions, then returns every evidence-backed finding in one Review Handoff. Only a fundamental architecture error, invalid public contract, major security defect, or fundamentally unmet Acceptance may stop early as blocked, with Reviewed and Unreviewed Scope; pass still requires the complete required changed-scope review.
• Material current-task findings from one Review Round and one Task ID form exactly one bounded Repair assignment with that Task ID unchanged. Each finding retains its relation, affected scope, Acceptance or risk impact, required validation, and covering re-review obligation. A scope-blocker returns through Main to Analysis, adjacent remains record-only, and Repair never batches findings across Task IDs.
• Any scoped material edit invalidates validation and review evidence for intersecting scope and transitive Task dependencies only, retaining unaffected current evidence. Repair receives fresh targeted validation and scoped independent re-review; a re-review covering the final obligation subsumes another Final Review.
• Unverified scope and residual risk are stated plainly.

Evidence Privacy

rd-skills has no required product-runtime or private telemetry channel. Ordinary agents use visible task contracts, source, diffs, commands, validation, review, repair, and handoff evidence. They do not write private event streams, internal task identities, prompt transcripts, or hidden evidence records.

Evidence artifacts exclude raw prompts, secrets, environment variables, full command logs, personal archives, and user-specific content. Local evaluation report retention follows repository and user policy. A retained report must identify its evidence type and source freshness; persistence does not make stale output current. Benchmarks owns interpretation and proof limits.

Content Quality

Professional and Layer 3 Skills are concise AI instructions, not background articles. Their decision rules, gotchas, execution checklist, stop conditions, output contract, and targeted references must change a current engineering decision.

Claim Quality

Structural validators prove structure. Deterministic and captured fixtures prove only their declared contracts. Default code-generation commands prove only case definition integrity, checked-in harness execution, and starter negative-control behavior; they do not evaluate a generated candidate. These artifacts do not prove real-host startup, wall-clock performance, provider behavior, production accuracy, or installed user experience. Generated catalogs are local discovery assets. The handwritten scorecard records release expectations, not generated status. Repository evidence does not support project rankings, popularity, external-adoption, privacy-guarantee, or competitor-superiority claims beyond the specific evidence it contains.