Review Design

Review Design Skill

Validate the quality and feasibility of an experiment plan before compute is spent. Runs a triage-first, fail-fast multi-level analysis hierarchy with parallel subagents and an adversarial red-team, then synthesizes a GO/REVISE/STOP verdict.

When to Use

Use when the research recipe's review_design ingredient is true (the default). The recipe calls this skill after plan_experiment to gate execution on a quality check. This skill is bounded by retries: 2 — on exhaustion the recipe proceeds with the best available plan.

Arguments

/autoskillit:review-design {experiment_plan_path}

• experiment_plan_path — Absolute path to the experiment plan file produced by /autoskillit:plan-experiment. Scan tokens after the skill name for the first path-like token (starts with /, ./, or .autoskillit/).

Critical Constraints

NEVER:

• Modify the plan file, any source code, or any file outside {{AUTOSKILLIT_TEMP}}/review-design/
• Halt the pipeline for a REVISE verdict — emit the verdict and let the recipe route
• Proceed to Level 2, 3, or 4 analysis when any Level 1 finding is classified as STRUCTURAL (halt at fail-fast gate). ADDRESSABLE L1 criticals continue L2-L4.
• Spawn SILENT (S) dimension subagents — they are not run and not mentioned in output
• Exit non-zero — GO, REVISE, and STOP are all normal outcomes (exit 0 in all cases)
• Include code snippets, shell commands, or specific tool invocations in findings or revision guidance — findings describe gaps and risks, not implementation instructions
• Prescribe HOW to fix an issue — findings must describe WHAT is lacking or at risk; the fix is the plan author's responsibility
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use model: "sonnet" when spawning all subagents via the Task tool
• Write output to {{AUTOSKILLIT_TEMP}}/review-design/ (relative to the current working directory)
• After writing output files, emit the absolute paths as structured output tokens as your final output. Resolve relative save paths to absolute by prepending the full CWD: verdict = GO|REVISE|STOP experiment_type = {type} evaluation_dashboard = /absolute/cwd/{{AUTOSKILLIT_TEMP}}/review-design/{filename}.md revision_guidance = /absolute/cwd/{{AUTOSKILLIT_TEMP}}/review-design/{filename}.md (REVISE only)
• revision_guidance is written and emitted ONLY when verdict = REVISE
• evaluation_dashboard is ALWAYS written and emitted
• Red-team agent always sets requires_decision: true on all its findings
• Halt at Level 1 fail-fast gate: if any Level 1 finding is classified as STRUCTURAL, emit STOP immediately — do NOT proceed to Level 2, 3, or 4. If all L1 criticals are ADDRESSABLE, tag them as REQUIRED priority fixes and continue L2-L4 analysis.

Context Limit Behavior

When context is exhausted mid-execution, output files may be partially written or absent. The recipe routes to on_context_limit, abandoning the partial review.

Before emitting structured output tokens:

1. If evaluation_dashboard was not fully written, emit verdict = STOP as a safe fallback
2. Omit revision_guidance if not written; the orchestrator handles the context-limit route

Workflow

Step 0: Read Plan & Setup

1. Create {{AUTOSKILLIT_TEMP}}/review-design/ if absent.

2. Extract experiment_plan_path from arguments (first path-like token starting with /, ./, or .autoskillit/). Error handling: When no path-like token is present in the arguments, emit verdict = STOP with message "No experiment_plan_path provided" and return (per the NEVER exit-non-zero constraint).

3. Read the plan file. Error handling: If the file does not exist or is unreadable at the resolved path, emit verdict = STOP with message "Plan file not found: {path}" and return.

4. Load the experiment type registry: a. Locate bundled types dir: run python -c "from autoskillit.core import pkg_root; print(pkg_root() / 'recipes' / 'experiment-types')" to get the absolute bundled directory path. b. Use Glob *.yaml in that directory, then Read each file. Parse YAML frontmatter to extract name, classification_triggers, dimension_weights, applicable_lenses, red_team_focus, and l1_severity fields from each. c. Check .autoskillit/experiment-types/ in the current working directory. If it exists, read all *.yaml files there. A user-defined type with the same name as a bundled type replaces the bundled entry entirely — do not merge fields. d. The resulting registry is a mapping of type name → spec. The set of valid experiment_type values for this run is the set of keys in the registry.

5. Parse YAML frontmatter using the backward-compatible two-level fallback:

   • Level 1 (frontmatter): Read YAML frontmatter between --- delimiters directly (zero LLM tokens). Return present fields and note which are missing. Record source: frontmatter for each extracted field. Error handling: If the YAML between --- delimiters is malformed, treat all fields as missing and fall through to Level 2 for all fields (fallback handling).

   • Level 2 (LLM extraction): For each missing field, launch a targeted LLM extraction subagent against the corresponding prose section. All extractions are independent and run in parallel. Record source: extracted for each field from this path.

   • Fields: experiment_type, hypothesis_h0/h1, estimand, metrics, baselines, statistical_plan, success_criteria

   • Missing-field to prose-target mapping:

     Missing Field	Prose Target	Extraction Prompt
     experiment_type	Full plan	"Classify using the loaded registry types: {', '.join(registry.keys())}"
     hypothesis_h0/h1	## Hypothesis	"Extract the null/alternative hypothesis"
     estimand	## Hypothesis + ## Independent Variables	"Extract: treatment, outcome, population, contrast"
     metrics	## Dependent Variables table	"Extract each row as structured object"
     baselines	## Independent Variables	"Extract comparators: name, version, tuning"
     statistical_plan	## Analysis Plan	"Extract: test, alpha, power, correction, sample size"
     success_criteria	## Success Criteria	"Extract three criteria"

   This two-level approach ensures backward compatibility with plans that lack frontmatter: the provenance (source: frontmatter or source: extracted) is tracked for each field and included in the evaluation dashboard.

Step 1: Triage Dispatcher

Launch one subagent. Receives full plan text plus parsed fields. Returns:

• experiment_type: one of the type names in the loaded registry (from Step 0)
• dimension_weights: the complete weight matrix for this plan (H/M/L/S per dimension)
• secondary_modifiers: list of active modifiers with their effects on weights

Schema validation: After the subagent returns, verify that experiment_type is a key in the loaded registry (from Step 0). If the returned value is not in the registry, default to exploratory and log a warning — do not silently pass an invalid type into the weight matrix lookup, as this would corrupt all subsequent spawning decisions.

Triage classification rules (first-match):

Use the classification_triggers list from each type in the loaded registry to classify the experiment. Apply first-match: iterate types in registry insertion order (bundled types sorted alphabetically, then user-defined types sorted alphabetically). The first type whose trigger description matches the plan is selected. If no trigger matches, default to exploratory.

Secondary modifiers (additive, increase dimension weights):

• +causal: mechanism claim in non-causal type → causal_structure weight +1 tier
• +high_cost: resources > 4 GPU-hours → resource_proportionality L→M
• +deployment: motivation references production/users → ecological_validity floor = M
• +multi_metric: ≥3 DVs → statistical_corrections weight +1 tier

Dimension weights:

Use the dimension_weights dict from the matched type's registry entry (loaded in Step 0). Each key is a dimension name; each value is one of weight=H (High), weight=M (Medium), weight=L (Low), or weight=S (SILENT — dimension not spawned, not mentioned in output). Pass the full dimension_weights dict to the triage subagent so it can return the complete weight matrix for this plan.

Silent Type Handling (after Step 1)

After Step 1 classification, check whether the experiment type is "silent" — an experiment type where quantitative-audit dimension scoring does not apply.

Detection rule: An experiment type is silent when >=6 of 8 dimension_weights are equal to S. Use is_silent_type(spec) from experiment_type_registry (or count the S values directly from the loaded registry entry).

Reference: docs/research/silent-type-convention.md defines the shared convention consumed by both review-design (this skill) and vis-lens-methodology-norms (#846).

When is_silent_type returns True:

1. Skip Steps 2-7 entirely — do not launch L1, L2, L3, L4, or red-team subagents.

2. Emit verdict = GO with requires_decision: false.

3. Write evaluation_dashboard with a "Scope Advisory" section containing:

   verdict: GO
   advisory_context:
     subject_kind: experiment_type
     subject_name: {experiment_type}
     reasoning: "Quantitative audit framework does not apply to {experiment_type}. Design rigor is assessed via domain-appropriate criteria (e.g., trustworthiness, transferability, dependability, confirmability for qualitative research)."
     reference_framework: "SRQR / COREQ"
   requires_decision: false
   

4. Write machine-readable YAML summary in the dashboard:

   # --- review-design machine summary ---
   verdict: GO
   experiment_type: {type}
   critical_count: 0
   warning_count: 0
   blocking_count: 0
   required_count: 0
   advisory_count: 1
   red_team_count: 0
   active_dimensions: 0
   warning_threshold: 0
   

5. Proceed to Step 8 (Emit Output Tokens) — the recipe routes to plan_visualization.

The advisory is appended to the evaluation dashboard file (which is later copied to research/{slug}/audit/design-review-dashboard.md by create_worktree.sh).

When is_silent_type returns False: Continue with the standard path (Steps 2-7).

Subagent Evaluation Scope (applies to ALL dimension subagents)

Include this instruction block in every dimension subagent prompt.

Every finding must describe WHAT is lacking or at risk in the experimental design. Never prescribe HOW to fix it — the fix is the plan author's responsibility.

• GOOD: "The plan does not address how implementation correctness will be verified before measurement"

• BAD: "The plan must contain apply_phase1_changes.sh with inline Python that greps for function ordering"

• GOOD: "The step-timing instrumentation could introduce inter-iteration contamination if reset ordering is incorrect"

• BAD: "step_timing::reset() uses Ordering::Relaxed — change to Ordering::Release"

Findings must never include:

• Code snippets or shell commands as required plan content
• Specific tool invocations, package versions, or file paths as mandated fixes
• Implementation-level instructions (these belong in implementation phases, not design review)

Design scope boundary:

Evaluate the experimental DESIGN: hypotheses, metrics, statistical methodology, controlled variables, threats to validity, data acquisition strategy, reproducibility specification.

Do NOT evaluate:

• Correctness of implementation code snippets included as illustrative examples in the plan
• Shell script syntax, path accuracy, or command correctness in plan examples
• Package version pinning, compile-time constants, or infrastructure code correctness

If a code snippet in the plan reveals a design-level concern (e.g., the metric definition contradicts the hypothesis), flag the design concern, not the code bug.

Step 2: Level 1 Analysis — Fail-Fast (parallel)

Two subagents run in parallel. Both are always H-weight; severity thresholds are calibrated per experiment_type via the rubric below.

Each L1 subagent receives as explicit inputs:

• Full plan text
• experiment_type (from Step 1 triage output)
• Severity calibration rubric (below)

Severity calibration rubric for L1 dimensions:

Use the l1_severity dict from the matched experiment type's registry entry (loaded in Step 0). Keys are estimand_clarity and hypothesis_falsifiability; values are severity levels (critical, warning, info). Calibration anchors: causal_inference → critical; benchmark, configuration_study, robustness_audit → warning; exploratory → info.

• estimand_clarity agent: "Can the claim be written as a formal contrast (A vs B on Y in Z)?" Reference the exp-lens-estimand-clarity philosophical mode as guidance (do NOT invoke the skill — reference its lens question only in the subagent prompt). Use the l1_severity.estimand_clarity value from the registry to assign severity.
• hypothesis_falsifiability agent: "What result would cause the author to conclude H0?" Use the l1_severity.hypothesis_falsifiability value from the registry to assign severity.

Each subagent returns findings in the standard JSON structure (see Finding Format below).

FAIL-FAST GATE: After both Level 1 subagents complete, check for critical findings. If ANY Level 1 subagent returns a finding with "severity": "critical":

• Collect these as stop_triggers
• Do not proceed to Level 2, 3, or 4 analysis
• Do not start the red-team agent
• Proceed directly to Step 7 (Synthesis) with only L1 findings
• The verdict logic will produce STOP; halt and emit tokens

Subagent parse failure: If a Level 1 subagent returns unparseable output (malformed JSON, empty response, token-limit truncation), treat it as if it returned one critical finding with message: "L1 subagent did not return parseable findings". This ensures parse failures trigger the fail-fast gate rather than silently passing it.

ADDRESSABLE vs STRUCTURAL classification: After collecting critical L1 findings, classify each one before applying the gate:

• ADDRESSABLE: Concrete methodological flaw with a mechanical fix — the research question remains answerable after revision (e.g., "decompose composite hypothesis into independent pairs", "add explicit estimand contrast statement")
• STRUCTURAL: The research question is not answerable with this experimental design regardless of revision (e.g., "no observable outcome exists for this hypothesis", "estimand requires counterfactual data that cannot be collected")

Classification scope limitation: Initially, only hypothesis_falsifiability findings are eligible for ADDRESSABLE classification — hypothesis restructuring is the dimension most likely to produce mechanically fixable defects. estimand_clarity findings default to STRUCTURAL (absent estimands typically indicate deeper design flaws).

Gate behavior after classification:

• If ANY critical finding is STRUCTURAL → halt L2-L4 analysis (emit STOP)
• If ALL critical findings are ADDRESSABLE → tag each as "priority": "REQUIRED", continue L2-L4 analysis. The verdict becomes REVISE (not STOP) with the ADDRESSABLE findings at the top of the evaluation dashboard.
• If mixed (some ADDRESSABLE, some STRUCTURAL) → halt (STRUCTURAL takes precedence)

Step 3: Level 2 + Red-Team (concurrent)

When the L1 gate passes (no STRUCTURAL critical L1 findings — gate also passes when all L1 criticals are ADDRESSABLE), launch 2–3 Level 2 subagents AND the red-team agent concurrently — all at the same time without waiting for each other.

Level 2 subagents (parallel, weights from the matrix):

• baseline_fairness: "Are all compared systems given symmetric resources and tuning effort?"
• causal_structure: weight from matrix (S for benchmark/config_study, H for causal_inference). Only spawn when weight ≥ L.
• unit_interference: "Can treatments spill over between experimental units?"

Red-team agent (concurrent with L2 and L4 — does NOT block L3):

Receives: full plan text and experiment_type (from Step 1 triage output)

• Five universal challenges (challenge every plan regardless of type):
  1. Goodhart exploitation — cheapest way to score well without solving the research question
  2. Data leakage — test-set info contaminating training/hyperparameter selection
  3. Asymmetric tuning — proposed method tuned against eval while baselines use defaults
  4. Survivorship bias — cherry-picking best run from multiple seeds
  5. Evaluation collision — same infrastructure in both treatment and measurement
• Type-specific focus: use red_team_focus.specific from the matched type's registry entry (loaded in Step 0).
• ALL red-team findings must set "requires_decision": true and "dimension": "red_team"

Red-team severity calibration rubric:

The red-team agent assigns severity based on the intrinsic quality of each finding. After the red-team agent returns, cap each finding's severity to the maximum allowed by the experiment type using this rubric — identical to how L1 severity calibration works. For causal_inference: critical red-team findings remain critical (STOP-eligible). For benchmark/configuration_study/robustness_audit: critical findings are downgraded to warning (REVISE-eligible but never STOP). For exploratory: all red-team findings are capped at info (informational only).

This cap is applied in Step 7 before the verdict logic evaluates stop_triggers.

Step 4: Level 3 (parallel)

Run after Level 2 completes. Do not wait for the red-team agent before starting Level 3.

Each L3 subagent receives:

• Full plan text
• experiment_type (from Step 1 triage output) — calibrates expected statistical rigor: exploratory plans do not require pre-registered correction procedures; causal_inference plans demand formal power analysis and correction pre-specification.
• L1 and L2 findings summary as context (findings may inform statistical planning relevance)

Three subagents run in parallel:

• error_budget: "Is power analysis present? Are error rates (Type I / Type II) acknowledged?"
• statistical_corrections: "Are multiple comparisons corrections pre-specified for all DVs?"
• variance_protocol: "Are seeds fixed? Is run-to-run variance addressed?" NOTE: absent seeds IS a valid finding for this dimension at H-weight — do not suppress via foothold validation.

Step 5: Level 4 (parallel, gated by triage)

2–4 subagents. Only spawn subagents for dimensions with weight ≥ L in the matrix. SILENT (S) dimensions are NOT spawned and NOT mentioned in output.

Each L4 subagent receives:

• Full plan text
• experiment_type (from Step 1 triage output) — calibrates rigor expectations per dimension: benchmark plans have lower ecological validity expectations than causal_inference plans by design; reproducibility_spec rigor scales with causal_inference > benchmark > exploratory.
• dimension_weights (from Step 1) — provides context on why this dimension was spawned (e.g., H-weight dimensions warrant stricter thresholds than L-weight dimensions)

Level 4 dimensions (spawn when not SILENT):

• benchmark_representativeness: "Does this generalize beyond the specific test bed?"
• ecological_validity: "Do test conditions match the intended deployment context?"
• measurement_alignment: "Do the metrics actually measure what the research question claims?"
• reproducibility_spec: "Could an independent party reproduce this experiment?"
• data_acquisition: "Does the plan include a complete data acquisition strategy?"
• agent_implementability: "Is this plan executable by a code-generating agent without human intervention?"

data_acquisition — Data Acquisition Completeness

Validates that the experiment plan includes a complete data acquisition strategy:

1. Hypothesis coverage: Every hypothesis in success_criteria has at least one data_manifest entry specifying its data source.
2. External source readiness: Every entry with source_type: external has an explicit acquisition command and a verification criterion.
3. Gitignored path handling: Every entry with source_type: gitignored has an acquisition/generation step — gitignored paths are empty in fresh worktrees.
4. Dependency ordering: If entry A's depends_on references entry B's acquisition, B must be listed before A (or the dependency chain must be acyclic).
5. Directive compliance: If the research task directive specifies particular data, the data_manifest must include acquisition steps for that data.

Findings format:

• STOP if: a hypothesis has no data source at all, or directive-specified data has no acquisition step
• REVISE if: an external source lacks verification criteria, or gitignored path handling is unclear

agent_implementability — Agent Execution Feasibility

Validates that the experiment plan can be implemented by a code-generating agent without human intervention:

1. Step atomicity: Each implementation step has a single unambiguous action. Multiple interleaved actions in a single step create ambiguity about ordering and completion criteria.
2. File path resolvability: All referenced files and modules are named with locatable paths. Vague references like "the utils module" or "the config file" are insufficient — the agent needs exact paths or unambiguous naming conventions.
3. Performance feasibility: Pseudocode and algorithms are viable at the specified data scale. A nested loop that is O(n^2) at n=50K is a design-level concern, not an implementation detail.
4. Verification criteria completeness: Each implementation phase has a concrete, runnable acceptance test or verification criterion. Phases without verification criteria cannot be confirmed complete by an agent.
5. Dependency ordering: No step requires an artifact (file, dataset, model checkpoint) that has not yet been produced by a prior step. Circular or forward dependencies make deterministic execution impossible.
6. Absence of human-only actions: No steps require subjective judgment, visual inspection, or domain expertise that a code-generating agent cannot execute deterministically. Examples: "review the output and decide if it looks reasonable", "manually inspect the plot for anomalies".
7. Artifact continuity: No planned artifact from a prior plan version was silently removed without replacement or explicit justification. Removals without explanation are data provenance regressions.

Findings format:

• REVISE if: any of checks 1–7 identifies a gap in the plan
• This dimension does not produce STOP-eligible findings (L4 contract)

Level 3 and Level 4 may run concurrently with the red-team agent (do not block on red-team).

Three-layer silencing (prevents orphan warnings):

1. Static SILENT from matrix — dimensions not spawned, not mentioned in output.
2. Foothold validation — before spawning M/L dimensions, check plan text has relevant content. If absent: M→L, L→S. Exception: absent seeds IS a finding for variance_protocol at H-weight.
3. Finding-count suppression — L-weight dimensions with zero findings: omit from output entirely. H/M dimensions with zero findings: emit "No issues identified" (deliberate clean bill of health).

Step 6: Wait for Red-Team

After Levels 3 and 4 complete, wait for the red-team agent if still running. All red-team findings are merged into the finding pool with their "requires_decision": true flag preserved.

Step 7: Synthesis

One synthesis pass (no subagent — orchestrator synthesizes directly):

1. Merge all findings from L1, L2, L3, L4, and red-team into a single list.
2. Deduplicate by (dimension, section, message) — identical findings from parallel agents are collapsed into one entry.
3. Consolidate related findings: If multiple findings address the same underlying methodological concern (e.g., four separate findings about multiple comparisons correction from different dimensions), group them as sub-findings under a single parent finding with the highest severity and priority of the group. The parent finding's message should describe the shared concern; sub-findings retain their dimension-specific detail as bullet points beneath the parent. This prevents the same issue from inflating finding counts and obscuring distinct problems.
4. Apply red-team severity cap, then verdict logic:

   # RT_MAX_SEVERITY is built from the registry loaded in Step 0 (dict-of-dicts from YAML parsing):
   RT_MAX_SEVERITY = {name: spec["red_team_focus"]["severity_cap"] for name, spec in registry.items()}
   SEVERITY_RANK = {"info": 0, "warning": 1, "critical": 2}
   rt_cap = RT_MAX_SEVERITY[experiment_type]
   
   for f in findings:
       if f.dimension == "red_team" and SEVERITY_RANK[f.severity] > SEVERITY_RANK[rt_cap]:
           f.severity = rt_cap  # downgrade before verdict
   
   # Reclassify after cap
   critical_findings = [f for f in findings if f.severity == "critical"]
   warning_findings = [f for f in findings if f.severity == "warning"]
   
   active_dimensions = count_of_spawned_dimensions  # tracked from Steps 2-6
   
   # Proportional warning threshold: each active dimension gets a budget of 5
   # warnings before the plan is flagged for revision.
   WARNING_BUDGET_PER_DIM = 5
   warning_threshold = active_dimensions * WARNING_BUDGET_PER_DIM
   
   # L1 fail-fast path: only STRUCTURAL defects trigger STOP
   l1_criticals = [f for f in critical_findings if f.dimension in {"estimand_clarity", "hypothesis_falsifiability"}]
   # Tag ADDRESSABLE L1 criticals as REQUIRED (scope: hypothesis_falsifiability only)
   for f in l1_criticals:
       if f.fixability == "ADDRESSABLE":
           f.priority = "REQUIRED"
   # Scope guard: estimand_clarity always STRUCTURAL; None fixability defaults to STRUCTURAL
   structural_stop_triggers = [
       f for f in l1_criticals
       if f.fixability == "STRUCTURAL" or f.fixability is None or f.dimension == "estimand_clarity"
   ]
   
   # Red-team STOP path: adversarial critical findings after full analysis (L2-L4)
   # These fire only when the L1 gate passed AND the severity cap still allows critical.
   stop_triggers = structural_stop_triggers + [f for f in critical_findings if f.dimension == "red_team"]
   
   if stop_triggers:
       verdict = "STOP"
   elif critical_findings or len(warning_findings) >= warning_threshold:
       verdict = "REVISE"
   else:
       verdict = "GO"
   

5. Write evaluation_dashboard_{slug}_{YYYY-MM-DD_HHMMSS}.md — always written. Must include:
   • Verdict banner and classification summary
   • Dimension Rationale subsection — renders a per-dimension rationale table sourced from spec.dimension_weight_rationale (loaded in Step 0 registry). Table format:

     ## Dimension Rationale (why these weights apply to {experiment_type})
     
     | Dimension | Weight | Rationale |
     |---|---|---|
     | {dim} | {weight} | {rationale from spec.dimension_weight_rationale[dim]} |
     

     Include ALL non-SILENT dimensions (H, M, L). For each dimension:
     • If dimension_weight_rationale[dim] exists and is non-empty: render the rationale string
     • If no rationale entry exists (legacy types with empty dimension_weight_rationale): render the Weight column only, leave Rationale blank Omit SILENT (S) dimensions from the table (consistent with finding-count suppression). If dimension_weight_rationale is entirely empty (no entries for any dimension), omit the Dimension Rationale subsection entirely — do not render an empty table.
   • Dimension scorecard table (dimension → weight → findings count → severity summary)
   • Adversarial findings section (red-team findings, each marked requires_decision: true)
   • Cannot Assess section with at least 2 items (dimensions where evaluation was impossible due to absent plan content; minimum 2 entries, e.g., "Randomization mechanism not described — cannot assess unit interference risk", "No resource budgets stated — cannot assess resource_proportionality")
   • Mechanizable check log — fixed checklist items (always evaluated) plus ad-hoc additions from subagents:
     • Fixed: "All implementation phases have runnable verification criteria"
     • Fixed: "All file paths in the implementation plan resolve to valid locations"
     • Ad-hoc entries contributed by dimension subagents during this review cycle
   • Machine-readable YAML summary block at end:

     # --- review-design machine summary ---
     verdict: GO|REVISE|STOP
     experiment_type: {type}
     critical_count: {n}
     warning_count: {n}
     blocking_count: {n}
     required_count: {n}
     advisory_count: {n}
     red_team_count: {n}
     active_dimensions: {n}
     warning_threshold: {n}
     

   The YAML summary block enables downstream recipe steps to parse verdict and counts without re-reading prose.

6. Write revision_guidance_{slug}_{YYYY-MM-DD_HHMMSS}.md — written ONLY when verdict = REVISE. Must include:
   • Required revisions: critical findings with gap and risk descriptions (not implementation instructions)
   • Recommended revisions: warning findings with gap and risk descriptions
   • Red-team findings with risk descriptions and decision points Revision guidance describes WHAT needs to change in the experimental design, never HOW to implement the change. The revision_guidance path is passed back to plan-experiment in the recipe loop.

Step 8: Emit Output Tokens

Emit these lines as your final output:

verdict = GO|REVISE|STOP
experiment_type = {experiment_type}
classification_timestamp = {ISO 8601 UTC timestamp}
evaluation_dashboard = /absolute/path/{{AUTOSKILLIT_TEMP}}/review-design/evaluation_dashboard_{slug}_{YYYY-MM-DD_HHMMSS}.md
revision_guidance = /absolute/path/{{AUTOSKILLIT_TEMP}}/review-design/revision_guidance_{slug}_{YYYY-MM-DD_HHMMSS}.md

classification_timestamp is the UTC timestamp (e.g., 2026-04-13T15:32:00Z) at the moment the experiment-type classification is finalized, before writing the evaluation dashboard.

revision_guidance line is emitted ONLY when verdict = REVISE. When verdict is GO or STOP, omit the revision_guidance line entirely.

Finding Format

All subagents must return findings in this JSON structure:

{
  "section": "## Hypothesis",
  "dimension": "estimand_clarity",
  "level": 1,
  "severity": "critical | warning | info",
  "priority": "BLOCKING | REQUIRED | ADVISORY",
  "fixability": "ADDRESSABLE | STRUCTURAL | null",
  "message": "{describes what is lacking or at risk — never prescribes how to fix}",
  "requires_decision": false
}

Fixability classification (L1 findings only — all other levels use JSON null):

See ADDRESSABLE vs STRUCTURAL classification in Step 2 for authoritative definitions and scope limitations. null (JSON null literal, not the string "null") is used for all non-L1 findings where fixability is not applicable.

Priority tiers (supplementing, not replacing, severity):

• BLOCKING: Must be addressed before implementation proceeds. Structural invalidity that would make results uninterpretable (e.g., "metric |Delta T_B| is mathematically invalid for Approach B — entire experimental arm produces meaningless data")
• REQUIRED: Must be addressed in revision. Methodology gaps that weaken but do not invalidate results (e.g., "composite hypothesis is unfalsifiable — decompose into independent pairs")
• ADVISORY: Should be addressed but omission will not invalidate results (e.g., "gitignore coverage for generated artifacts not specified")

Priority assignment rules:

• STRUCTURAL L1 criticals → BLOCKING
• ADDRESSABLE L1 criticals → REQUIRED
• Non-L1 criticals → BLOCKING (default) or REQUIRED (if finding describes a gap rather than an invalidity)
• Warnings → REQUIRED (default) or ADVISORY (if finding is informational in nature)
• Info → ADVISORY

Red-team findings: always "requires_decision": true, "dimension": "red_team".

Output

{{AUTOSKILLIT_TEMP}}/review-design/
├── evaluation_dashboard_{slug}_{YYYY-MM-DD_HHMMSS}.md   (always written)
└── revision_guidance_{slug}_{YYYY-MM-DD_HHMMSS}.md      (REVISE only)

Emit structured output tokens (absolute paths) as your final output.

Related Skills

• /autoskillit:plan-experiment — produces the plan this skill validates
• /autoskillit:scope — first step in the research recipe chain
• /autoskillit:implement-experiment — consumes this skill's GO output (via recipe routing)
• exp-lens skills — philosophical modes referenced in dimension prompts (not invoked directly)