Conciseness

The body is mostly efficient and decision-oriented, but re-explains concepts a competent model already knows: the Overview walks through Fisher's principles ("assign treatments at random so that confounders... are balanced in expectation"), the mistakes section explains basics like "Evaporation and thermal gradients make plate edges differ", and the Installation/Resources sections carry version and date detail (pyDOE3 archived May 2026, tested-version lists, Windows interpreter paths). This is anchor 3's 'includes some unnecessary explanation or could be tightened' rather than anchor 4's 'minor instances', since the conceptual re-explanations span several sections.

Actionability

Both script sections give copy-paste-ready code with real arguments (e.g., `block_randomization(n=60, arms=["treatment", "control"], seed=42)`, `central_composite({"temp_C": (20, 60), "conc_mM": (1, 10)}, center=(2, 2), face="inscribed", seed=42)`) whose imports, function names, and keyword arguments all match the bundled scripts, plus install commands and CSV export. Specific examples cover the common randomization and DOE cases; this matches anchor 5, exceeding anchor 4 because the examples are verified-executable and comprehensive.

Workflow Clarity

The 8-step Workflow is clearly sequenced (state question/unit → list nuisance factors → pick design → decide replication → generate layout → randomize and verify → document → match analysis) with explicit validation checkpoints: step 6's "verify roster IDs, completed-block balance, range limits, and rank", the `arm_balance(sched)` sanity check, and "verify every ID exactly once". This matches anchor 5's 'explicit validation steps' rather than anchor 4, since checkpoints are named rather than implied; the destructive/batch cap does not apply as verification steps are present and the skill generates rather than destroys.

Progressive Disclosure

SKILL.md is a genuine overview: a decision tree for choosing a design, working examples for the two scripts, a mistakes checklist, and clearly signaled one-level-deep pointers ("→ references/randomization_and_blocking.md", etc.) with a Resources section repeating them. All four referenced files and both scripts exist on disk and the reference files contain no nested references. Content is appropriately split with easy navigation — the anchor 5 pattern, not anchor 4's 'minor organization gaps'.

Specificity

"Designs experiments and studies BEFORE data is collected — choosing a design, randomizing, blocking, and laying out treatment combinations so results are interpretable" lists multiple specific concrete actions (design choice, randomization, blocking, treatment layout) with comprehensive coverage of the skill's capabilities. It clearly matches the anchor 5 example pattern rather than anchor 4, since no meaningful action is missing.

Completeness

It explicitly answers both questions: the 'what' in the opening sentence (concrete design-time actions) and the 'when' via "Use whenever someone is planning a study..." with concrete trigger phrases, plus boundary routing to statistical-power and statistical-analysis. This matches anchor 5 exactly; it exceeds anchor 4 because the 'when' is explicit and specific rather than merely present.

Trigger Term Quality

The description covers natural terms exhaustively with synonyms: "randomization, blocking, stratification, controls, factorial or fractional-factorial designs, design of experiments (DOE), screening many factors, response-surface optimization, crossover or repeated-measures or split-plot designs, cluster/group randomization, Latin squares, plate layouts, batch/run-order effects, replication vs. pseudoreplication", plus informal phrasings users would actually say ("how should I set up this experiment", "assign these mice to conditions"). This is the anchor 5 'comprehensive coverage including synonyms' case, not anchor 4's 'a few natural terms missing'.

Distinctiveness Conflict Risk

"BEFORE data is collected" carves out a clear niche, and the description actively disambiguates adjacent skills ("For computing the sample size or power once the design is chosen, use statistical-power; for analyzing data already collected, use statistical-analysis"), minimizing wrong-skill triggers. Distinct specialized triggers like pseudoreplication and split-plot designs place this firmly at anchor 5 rather than the minor-overlap anchor 4.