Skip to main content

Validation Gate (Read-Only, 3 Checks)

Every recommendation passes all 3 checks before delivery. Any failure blocks delivery per HC#8. Fail-open is NOT applied since the skill is read-only.

V1: Classification Confidence​

Check. Every question-type classification carries a High / Medium / Low tag with the signals that support it.

Required per classification.

  • question_type from the 9-type taxonomy (one of: Temporal, Categorical, Composition, Distribution, Correlation, Ranking, Geographic, Status/KPI, Sequential) OR refused_no_match with rationale.
  • confidence (High, Medium, or Low).
  • supporting_signals list (non-empty for High and Medium; may include a single weak-signal note for Low).
  • classification_notes (required when confidence is Low; explains why the closest match is weak).

Sentinel values acceptable. null for classification_notes when confidence is High or Medium. "unscanned" for confidence when get-worksheet-xml fallback failed and mark type could not be inferred.

Fail action. Re-derive the classification from the taxonomy. If a second run still fails, refuse delivery per HC#8. Never leave the confidence blank.

V2: Anti-Pattern Gate Completeness​

Check. Every run evaluates the full anti-pattern catalog against the current chart.

Verification steps.

  1. Every anti-pattern ID in skill-expertise/anti_patterns.md (AB-1 through AB-6, W-1 through W-7, I-1 through I-2) is evaluated. Skipped IDs are a V2 fail.
  2. Auto-Block anti-patterns cannot be silently downgraded to Warning or Info. HC#11 enforcement is a V2 check.
  3. Multiple matching anti-patterns all fire; the verdict is the highest severity.

Fail action. Identify the skipped or downgraded anti-pattern. Re-run the gate. If a second run still fails, refuse delivery per HC#8.

V3: Handoff Completeness​

Check. Every field in chart_swap_recommendation/1 populated per skill-expertise/handoff_schema.md.

Verification steps.

  1. Every required field present with a non-blank value or a sentinel (null, [], "unscanned").
  2. rebuild_handoff_payload populated when verdict is Swap Recommended or Anti-Pattern Block; empty (null) when verdict is Keep.
  3. anti_patterns_matched list is empty ([]) when no anti-pattern fires, never blank.

Fail action. Identify the missing field, populate with the correct value or sentinel, and re-run V3. If a second run still fails, refuse delivery per HC#8.

Gate Execution Order​

  1. V1 (classification confidence) first. Cheapest to run; catches most failures.
  2. V2 (anti-pattern gate completeness) next. Validates the recommendation logic ran.
  3. V3 (handoff completeness) last. Depends on V1 and V2 passing to have complete data.

What V-Gate Does NOT Check​

Unlike calc-family's 8-check gate, this gate does NOT check:

  • Whether the recommended chart type will render correctly in Desktop (runtime; worksheet-builder owns this).
  • Whether the user will accept the recommendation (user-side decision).
  • Whether the current chart's underlying data is correct (route to data-quality-review).
  • Whether the recommended chart's calc dependencies exist (route to calc-orchestrator).
  • Format-string drift on the current chart (route to format-drift-scanner).

Refuse to Fabricate​

If the V-gate cannot pass after one retry, refuse delivery and diagnose. Never fake a V-gate pass. Never invent a supporting signal to reach a Medium or High confidence tag.