Skill eval designer

Skill Eval Designer

Produce a Skill Evaluation Program that can disprove whether an exact skill bundle improves realistic work while preserving routing boundaries, safety, cost, and evidence integrity.

Atomic boundary

Own the eval contract: candidate digests, routing/behavior tasks, controls, rubrics/oracles, artifact checks, model/provider matrix, hidden-data handling, metrics/uncertainty, provenance, replay, attestation, expiry, and regression triage. Do not own the skill procedure, repository portfolio decision, runtime capability, or adoption/outcome telemetry.

Workflow

1. Freeze the question: exact skill name/description/body/references/scripts, intended jobs/artifacts, nearest neighbours, risks, output budget, supported runtimes/tools, and claim that the eval may falsify.
2. Read references/skill-eval-systems.md. Compute an injection-contract digest over name+description and behavior digest over the complete ordered bundle. Bind candidate commit, catalog digest, task/rubric/runner/policy/model-registry digests, parameters, seed, tool availability, retries, and expiry. For repeated or continuously monitored evaluations, also apply references/holdout-sequential-evaluation.md.
3. Build full-catalog runtime auto-injection cases: at least five realistic positives and five near-neighbour negatives per candidate, abstentions, multilingual/ambiguous/correction/misleading-keyword forms, and compound tasks with an exact expected injected set, expected artifacts, one semantic owner per artifact, supporting constraints, and forbidden contradictions. Exercise realistic catalog size and the runtime's documented description shortening/omission state. Test the native mechanism; do not build a meta-router or evaluate descriptions in isolation.
4. Build unseen behavior tasks from real job shapes and adversarial failure modes. Expected behavior is observable assertions and deterministic artifact checks, not a leaked golden answer. Include unsafe shortcuts, prompt injection, missing authority, uncertainty, and recovery.
5. Compare under one declared budget: base model, length-matched generic expert instruction, exact skill, and strongest reasonable public comparator where licensing permits. Use at least two answer-model and two independent judge families for promotable evidence.
6. Separate deterministic oracles from blind judgment. Recompute metrics from complete raw artifacts in a protected eval store with declared audience, access, retention, and deletion. Retain per-case outputs, critical failures, costs, latency, variance, model/provider identities, and hashed response receipts there; a public/shareable report uses redacted aggregates, safe excerpts, and opaque locators rather than copying protected prompts or data.
7. Keep promotion holdouts hidden at inference through a one-use pre-run commitment. Disclose their content after the attested run only when the dataset was explicitly classified and authorized for that audience; confidential tasks remain protected while commitments, receipts, aggregate results, and safe excerpts support independent verification. Reject duplicate, leaked, authored-to-fit, or post-result-edited tasks.
8. Report per-skill F1, artifact-owner accuracy, abstention precision/recall, exact-set and compound exact-set accuracy, win rates versus every control, paired/task-clustered confidence, critical failures, cost/latency, and family-specific effects. Aggregates may not hide one broken route.
9. Save failures as minimal regression cases, edit the skill—not the threshold to chase green—and rerun from a newly frozen candidate. Any relevant byte change invalidates the matching proof.

When not to use

• Use source-to-skill-distiller to author or repair the procedure from source material before evaluation.
• Use skill-repository-curator to audit portfolio value, content quality, collisions, consolidation, and retirement across a repository.
• Use deterministic unit/contract/property/model tests directly when no model behavior or routing judgment is involved.
• Do not use authored evals, self-installs, catalog presence, or historical scores as adoption or current behavior proof.

Guardrails

• Do not leak the intended answer, suspected bug, fix, or expected verdict into tasks, subagents, judges, or holdouts.
• An LLM judge is additional evidence, never the sole critical-safety oracle. Calibrate it on prompt injection, sycophancy, verbosity, and control order.
• Never let the author agent self-attest provider/model identity, omit failed samples, rewrite raw outputs, relabel families, or average away critical regressions.
• Do not publish raw customer context, private prompts, hidden holdouts, credentials, internal topology/process state, provider secrets, or reconstructable protected outputs merely to make an evaluation auditable.
• A green score is invalid without exact candidate binding, complete artifacts, independent recomputation, provenance, and replay.
• A model saying that it loaded or followed a Skill is not injection evidence. Prefer runtime-native selection/load traces; where the runtime exposes none, label the result a fresh-context behavior test and do not upgrade it to an exact injection claim.
• Separate behavior improvement from demand, business outcome, current authority, and production proof.

Output contract

Return one Skill Evaluation Program containing:

1. falsifiable claim, exact candidate/catalog identities and all digests;
2. runtime auto-injection dataset design with positive, neighbour, abstention, compound, multilingual, and injection cases;
3. behavior tasks, assertions, deterministic artifacts, critical failures, and output budget;
4. base/expert/skill/competitor and answer/judge/provider matrix;
5. protected holdout/raw-artifact store, commitment and authorized disclosure rule, receipts, provenance/attestation, replay, retention, and expiry;
6. thresholds plus per-skill/per-case/per-family metrics, uncertainty, cost, and latency;
7. failure triage, regression fixtures, proof invalidation, and next run.

Complete only when an independent runner can reproduce the candidate, tasks, outputs, metrics, and verdict without trusting the author narrative.