Design AI Benchmarking

Design-AI-Benchmarking Skill

Purpose

This skill pressure-tests an AI-vs-human-expert benchmark before any ratings are collected, so that the comparison is fair, the rubric measures distinct constructs, the scale is calibrated, and the reported reliability is interpretable. It is the AI-evaluation specialization of /design-study: where /design-study reviews a study in general, this skill owns the specific machinery of comparing AI system(s) to a panel of human experts (or to each other) on rated outputs.

Use it when:

• one or more AI systems will be scored against a human-expert reference (reader study, annotation panel, AI-output evaluation, model-vs-model bench)
• a rubric and rating protocol must be locked before reviewers begin
• a benchmark feels vulnerable to "the highest score is just the most tautological item" or "low agreement, but we cannot tell why" criticism
• a reviewer or editor asks how the evaluation controlled for rater drift, leakage, or judge bias

Do not use it for: general study/validity review (use /design-study); statistical execution such as ICC or DeLong (use /analyze-stats); reporting-guideline item audits (use /check-reporting); or reviewing an already-written manuscript (use /peer-review or /self-review).

Communication Rules

• Communicate with the user in their preferred language.
• Use English for statistical, machine-learning, and reporting-guideline terminology.
• Be direct about evaluation-validity risks, but always propose the smallest feasible fix first.
• Never invent reviewer ratings, reference labels, or agreement statistics; those come from collected data only.

Standard Output

## AI-Benchmark Design Review
Evaluation question: ...
Arms / systems compared: ...
Reference (human-expert panel): ...
Unit of rating: (item / case / output)

### Rubric (decoupled dimensions)
- dimension -> construct -> anchors (1..k)

### Calibration probes (blinded, randomized)
- positive-control / known-bad / instability / mechanism-contradiction

### Reviewer panel
- n reviewers, metadata captured, per-reviewer randomized order

### Reliability plan
- overall IRR target + control-item IRR (reported separately)

### Judge strategy
- human-as-judge / LLM-as-judge / both + adjudication rule

### Validity risks
1. ...

### Minimal fixes
- ...

### Decision
- Ready to collect / Needs rubric revision / Needs arm or judge redesign

Workflow

Phase 1: Define the evaluation question and arms

Pin down, in writing:

• the exact claim the benchmark must support (e.g., "system A's outputs are perceptually indistinguishable from expert outputs", not "system A is deployment-ready")
• every arm/system being compared, and what each arm receives as input (same items, same information access, same output format) so no arm has a hidden advantage
• the human-expert reference: who they are, and whether they set ground truth, provide a comparison arm, or both
• the unit of rating (item, case, output) and how many units each reviewer sees

Gate: Present the reconstructed evaluation question, arms, and reference to the user and confirm before designing the rubric. A wrong reconstruction misdirects the entire benchmark.

Phase 2: Design a decoupled multi-dimensional rubric

• Decouple the axes. Each rated dimension measures one construct. Keep "is the output valid/correct" separate from "is it novel", "is it feasible/measurable", "does it add value over current tools", and "would it change action". A candidate can be high-validity yet low-added-value ("real but redundant"); a single blended score hides this divergence.
• Anchor every scale point with a short verbal descriptor; pilot the anchors with at least one reviewer before locking.
• Pre-specify discriminant validity: hypothesize which dimensions should correlate vs be orthogonal, then report the full inter-dimension correlation matrix to confirm the rubric measures distinct constructs.
• A worked rubric template lives in ${CLAUDE_SKILL_DIR}/references/elicitation_rubric_template.md.

Phase 3: Insert and randomize calibration probes

Plant a small number of deliberate control items, blinded and randomized across raters (record who received which via a probe_arm flag), to (i) anchor the scale, (ii) measure rater drift/fatigue, and (iii) audit the rubric and pipeline itself. Four useful flavors:

• Positive control / "too-good" item — a known-strong or near-tautological item; tests whether raters equate "largest effect" with "best", and whether the construct-independence gate (Phase 7) works.
• Known-bad negative control — an engineered defect (fabricated reference, missing key statistic); expected to score low.
• Instability item — an estimate that reverses or fails to replicate on a holdout; tests caveat-handling.
• Mechanism-contradiction item — an empirical direction that opposes the proposed mechanism.

Probes are planted or adjudicated, never fabricated to fit a hypothesis.

Phase 4: Construct the reviewer panel

• Recruit reviewers spanning the intended expertise gradient; pre-specify any expertise stratification.
• Capture reviewer metadata (years of experience, prior AI-evaluation experience, subspecialty) for descriptive reporting and stratified analysis.
• Randomize item order per reviewer (not one global seed) and record the order; plan to analyze order and fatigue effects.
• Require each item to be judged standalone; discourage cross-item references in free-text, which signal non-independent rating.

Gate: Present the panel composition, stratification, and randomization plan for user review before recruitment is finalized.

Phase 5: Set inter-rater reliability targets

• Pre-specify the agreement statistic (e.g., ICC for continuous ratings, weighted kappa for ordinal) and a target with justification.
• Report reliability on the planted control items separately as primary evidence of rubric and scale validity. A low overall ICC is interpretable only if raters at least converge on the controls; surfacing both numbers prevents "low agreement => bad rubric" or "bad raters" misreads.
• Plan the minimum ratings-per-item needed for a stable agreement estimate (delegate the math to /analyze-stats).

Phase 5b: Reader allocation under burden constraints (anchor-and-rotate)

When the item pool is larger than one reader can rate in a session, do not force every reader to rate every item (that caps the total pool at the per-reader limit and discards coverage). Use an anchor-and-rotate (balanced-incomplete-block) layout: all readers rate a shared anchor set (which carries the inter-rater ICC/kappa, alongside the planted controls), and each reader additionally rates a rotating unique block, so total coverage grows independently of the per-reader cap. The usually-binding constraint is the number of available expert readers, not the item count — solve the reverse problem (max_pool = anchor + R*(cap-anchor)//m) to size the must-rate set to a realistic panel. Pre-specify anchor membership, raters-per-item, and the rotation seed before rating. Formulas, trade-offs, and a stdlib reference implementation are in ${CLAUDE_SKILL_DIR}/references/anchor_rotate_reader_allocation.md.

Phase 6: Choose the judge strategy and adjudication

• Decide human-as-judge, LLM-as-judge, or both. If an LLM is used as a judge, treat it as one more arm whose ratings must themselves be validated against the human panel on the control items.
• Pre-specify the adjudication rule for disagreement (e.g., majority, a third senior reviewer, consensus discussion) and who adjudicates.
• Blind judges to arm identity wherever feasible; record any unavoidable unblinding.

Phase 7: Construct-independence and leakage guards

• Exclude any predictor or input that is a definitional component of the outcome (mathematical definition), and flag near-tautological composites built from the outcome's defining components — they produce an inflated, near-circular result and belong as labeled probes, not discoveries.
• Verify no arm sees post-decision or outcome-derived information the others do not.
• Confirm the reference labels were not derived from the same model output being evaluated.

Phase 8: Lock a structured export schema

Define the machine-readable rating record up front: per-item ratings across every rubric dimension, free-text justifications, follow-up flags, the probe_arm flag, reviewer id and metadata, item order, and timing. A synthetic schema lives in ${CLAUDE_SKILL_DIR}/references/benchmark_export_schema.json.

Gate: Present the final rubric, probe set, panel plan, judge strategy, and export schema together; collect explicit user approval before any rating begins. Locking these before data collection is the whole point — changes afterward compromise the comparison.

Handoff Rules

• route to /analyze-stats for ICC / weighted kappa / DeLong, agreement sample size, and effect-size real-world translation of the benchmark results
• route to /check-reporting for STARD-AI, CLAIM, or TRIPOD+AI item-level reporting once the design is locked
• route to /design-study when the broader study around the benchmark (cohort logic, analysis unit, comparator) also needs review
• route to /peer-review or /self-review only after ratings exist and a manuscript is being assessed

What This Skill Does NOT Do

• It does not compute agreement statistics or run analyses directly (that is /analyze-stats).
• It does not collect or fabricate ratings, reference labels, or probe outcomes.
• It does not draft manuscript prose or run a reporting-guideline audit.
• It does not replace a full peer review of a finished manuscript.

Anti-Hallucination

• Never fabricate references. All citations must be verified via /search-lit with a confirmed DOI or PMID. Mark unverified references as [UNVERIFIED - NEEDS MANUAL CHECK].
• Never invent reviewer ratings, agreement statistics, reference labels, or probe outcomes — these come from collected data only. A reported ICC, kappa, or score with no underlying rating record is the failure mode this skill exists to prevent.
• Never invent clinical definitions, diagnostic criteria, or guideline recommendations. If uncertain, flag with [VERIFY] and ask the user.
• If a reporting-guideline item, journal policy, or evaluation standard is uncertain, state the uncertainty rather than guessing.

Reference Files

• ${CLAUDE_SKILL_DIR}/references/elicitation_rubric_template.md -- a synthetic, decoupled multi-dimension rating rubric with anchors and a planted-probe column.
• ${CLAUDE_SKILL_DIR}/references/benchmark_export_schema.json -- a synthetic JSON schema for the per-item rating export (ratings, justifications, probe_arm, reviewer metadata, order, timing).
• ${CLAUDE_SKILL_DIR}/references/anchor_rotate_reader_allocation.md -- anchor-and-rotate (balanced-incomplete-block) reader allocation: formulas, the reverse "max pool for R readers" ceiling, trade-offs, and a stdlib reference implementation (Phase 5b).