validate

name	validate
description	Validate provenance and schema for SWARM claims. Verifies run_id links, effect sizes, correction methods, confidence criteria, and _schema compliance. Triggers on "/validate", "/validate [claim]", "validate provenance", "check this claim".
user-invocable	true
allowed-tools	Read, Write, Edit, Grep, Glob, Bash
context	fork

Runtime Configuration (Step 0)

Read vault/ops/derivation-manifest.md and vault/ops/config.yaml.

---

EXECUTE NOW

Target: $ARGUMENTS

• If target is a claim path: validate that claim fully
• If target is empty: validate all claims lacking full provenance
• If target is "all": batch validate entire vault/claims/ directory

Execute these steps:

1. Read the target claim
2. Run all six validation gates
3. Report PASS/FAIL per gate with specific failures
4. For failures: provide exact remediation (what run_id to add, what effect size to compute)

---

Validate Provenance

Combined quality verification gate for SWARM claims.

Six Validation Gates

Gate 1: Run Provenance

Every evidence.supporting entry must have:

• run: pointing to a folder in runs/ that actually exists
• metric: a named metric
• detail: effect size, p-value, N — not just a description

# Check run exists
ls runs/[run_id] 2>/dev/null && echo "EXISTS" || echo "MISSING"

Gate 2: Statistical Rigor

For high or medium confidence claims:

• Effect size (Cohen's d) must be reported
• Correction method must be one of: Bonferroni, Holm-Bonferroni, Benjamini-Hochberg
• Raw p-values alone are INSUFFICIENT — must pair with correction method
• N (number of seeds/runs) must be stated

Gate 3: Confidence Criteria Match

Confidence level must match evidence strength:

• high: Bonferroni-significant AND replicated across ≥2 independent runs/seeds
• medium: Nominally significant OR single-sweep support with BH correction
• low: Suggestive but underpowered (<20 seeds) or unreplicated
• contested: Conflicting evidence from different runs

Gate 4: Schema Compliance

Check against vault/templates/claim-card.md _schema block:

• Required fields present: description, type, status, confidence, domain, evidence, created
• Enums valid: status ∈ [active, weakened, superseded, retracted], confidence ∈ [high, medium, low, contested]
• Description ≤200 chars, no trailing period, adds info beyond title

Gate 5: Boundary Conditions

Claims without boundary_conditions are incomplete:

• What topology was tested? (small-world, random, scale-free, ring, complete)
• What agent count? (8, 14, custom)
• What adversarial fraction? (0%, 10%, 25%, 30%)
• What time horizon? (epochs, steps)

Gate 6: Prose-as-Title Test

Run the claim test: "this claim argues that [title]"

• Must make grammatical sense
• Must be something you could agree or disagree with
• Must be lowercase with spaces, no filesystem-breaking punctuation

Output Format

Validation: vault/claims/[claim-name].md

Gate 1 — Run Provenance: PASS / FAIL
  [if FAIL: which run_id is missing, which entry is incomplete]

Gate 2 — Statistical Rigor: PASS / FAIL
  [if FAIL: which effect size is missing, which correction method is absent]

Gate 3 — Confidence Criteria: PASS / FAIL
  [if FAIL: what evidence level was found vs what confidence claims]

Gate 4 — Schema Compliance: PASS / FAIL
  [if FAIL: which required fields are absent, which enums are invalid]

Gate 5 — Boundary Conditions: PASS / FAIL
  [if FAIL: which dimensions are undocumented]

Gate 6 — Prose-as-Title: PASS / FAIL
  [if FAIL: why the title fails the claim test]

OVERALL: PASS (6/6) / NEEDS WORK ([N] gates failed)

If all gates pass, task is complete. If any fail, remediation must be completed before the claim is marked active.