Guardrail Rules

Guardrail Check Script

Run python3 skills/presentation-creator/scripts/guardrail-check.py <outline.yaml> <speaker-profile.json> to compute guardrail checks. The script handles the three-outcome PASS/WARN/FAIL logic:

• [FAIL] — value exceeds the limit
• [WARN] — value is under the limit but within 5 percentage points of it
• [PASS] — value is under the limit by more than 5 percentage points

The script checks pattern-history authorization, slide budget, Act 1 ratio, closing completeness, cut lines, data attribution, profanity, and branding. Its recurring_antipatterns array is the complete catalog recurrence output. Render those records without re-filtering them or adding movement data. Add the remaining checks (time-sensitive, current-outline contextual antipatterns, illustration coverage, pattern strategy, AI writing patterns) to the report manually.

The AI-writing-pattern scan is delegated to blog-writer, never reimplemented here, and it reports findings without applying them. Its execution contract — invocation, voice suppression, classification, and the absent-scanner path — is in skills/presentation-creator/references/phase4-guardrails.md.

A disabled pattern-history result affects only catalog-derived speaker history. Keep independent profile configuration and guardrails enabled. When history is partially available, require the exact domain for each derived field:

• Use mastery_and_novelty for tiers, strengths, and novelty.
• Use underuse for underuse.
• Use signature_combinations for combinations.
• Use the antipattern_recurrence domain only through the script's recurring_antipatterns output.
• Use trends for movements and score/breadth trend.
• Use modes for by-mode history.
• Always retain the current-taxonomy scan of the new outline.

Schema-v4/v5 top-level guardrail_sources.recurring_issues[] and badges[] are usable only when each consumed entry explicitly declares source_lane: "non_pattern". Suppress legacy or ambiguous entries. Catalog-derived warnings and reinforcement come only from authorized pattern_profile history.

Antipattern Tags

Every antipattern flag MUST be tagged as one of:

• [RECURRING]. Render one guardrail-check.py recurring_antipatterns record.
  • Preserve its recurrence_classification, evidence, and optional trend fields.
  • Do not reclassify the record.
• [CONTEXTUAL] — detected in the current outline but NOT in the speaker's authorized historical profile. It is a current-outline finding, not necessarily a first-time issue when history is unavailable.

Never use generic unlabeled antipattern warnings.

4-Tier Pattern Strategy

When pattern_history_status.py reports the mastery_and_novelty domain, organize recommendations into exactly four tiers:

1. Signature. Use current-cohort mastery_levels.signature patterns.
2. Contextual history. Use current-cohort regular/occasional patterns worth considering here.
3. New to You. Use exactly mastery_levels.never_tried / never_used_patterns entries that fit this talk.
   • Do not treat a first detection in the newest talk or not_yet_observed as New-to-You.
4. Shake It Up. Use exactly 1-2 wild card patterns for experimentation.
   • Do not use zero or three or more wild card patterns.

When the mastery domain is unavailable, do not manufacture those tiers from legacy profiles or unprovenanced prose. Present a flat relevant-pattern list from the current taxonomy without usage or novelty claims. A catalog fingerprint or scoring-schema change is a generation reset, never evidence of improvement or regression.

history_enabled: true means at least one policy-bound domain is available, not that all classification fields are authorized. opportunity_rows_available: true is an audit signal, not history authorization. Raw occurrence rows do not establish novelty, mastery, recurrence, or trend. Profile schema v4 and Section 15 v2 are occurrence-only; profile schema v5 and Section 15 v3 bind classifications to the versioned policy.

A catalog/scoring identity change is a generation reset. Within one generation, a changed policy_semantic_sha256 is a classification-comparison reset. Neither reset is ever evidence of speaker improvement or regression.