Profile Construction Rules

Read this reference before aggregating or constructing speaker-profile.json, and again before diffing an existing profile. The profile schema remains authoritative for the complete JSON shape; this file defines source ownership, merge behavior, and fail-closed edge cases used by the eight-step workflow in ../SKILL.md.

Contents

• Loader and cohort semantics
• Non-catalog aggregation
• Template-layout merge
• Profile field ownership
• Pattern-profile construction
• Empty-cohort behavior
• Pacing and Section 15
• Existing-profile diff
• Achievement badges

Loader and Cohort Semantics

baseline_talks is the exact active Presentation Pattern scoring generation whose persisted source-located artifacts still match their recorded identities. The loader passes the vault root and database-configured source roots to the shared read-only freshness assessor. The loader has already excluded scoring-v5 records whose persisted citation artifacts are missing, replaced, or inconsistent with their recorded root and digest; never restore them from another cohort or a prior profile.

pattern_baseline.eligible_talk_count is the occurrence-cohort size. Its raw-score count and average are available only when every eligible talk shares one opportunity_coverage_identity containing at least one evaluable opportunity. Otherwise retain the explicit mixed-coverage or no-evaluable-opportunities unavailable state. pattern_opportunities contains the canonical exhaustive positive and negative rows. pattern_classification contains the canonical policy stamp, per-domain availability, exhaustive classifications, trend audit, and derived projections. Copy both payloads rather than recalculating them. pattern_scoring_exclusions explains every otherwise eligible talk omitted from the cohort, including artifact drift.

Copy the loader's complete classification_policy stamp unchanged. Do not construct or repair any field in that stamp during profile assembly.

current_instrumentation_talks and stale_instrumentation_talks form a separate extractor cohort for pacing-sensitive analysis. Instrumentation membership never confers pattern-baseline eligibility.

Legacy database schema 0 and current schema 1 are read-only inputs. An unsupported future database or record schema has no usable prior state. Compatibility stays internal, adds no payload fields, and never migrates the database.

A missing rhetoric-style-summary.md blocks profile generation and requires vault-ingress. A missing slide-design-spec.md yields design_spec: ""; leave the design-spec section empty and continue.

Non-Catalog Aggregation

Keep these cohorts distinct:

• processed_talks supplies talks_analyzed and non-catalog narrative/profile data.
• baseline_talks is available for source review only; deterministic pattern_opportunities and pattern_classification supply catalog rows.
• current_instrumentation_talks supplies pacing data only.

For non-catalog dimensions, skip processed talks with empty structured_data and use matching summary prose when needed. If every processed talk lacks structured data, warn the speaker and use prose only for non-catalog fields. Never use prose, processed_talks, excluded_pattern_scoring_talks, either instrumentation cohort, or a prior profile to fill a missing pattern aggregate.

Template-Layout Merge

Accept extraction schema v1, v2, v3, or v4 for template_layouts only. Missing or legacy v0 output and unknown future versions are unusable; obtain current extraction output. A v1 record has no timing contract and must not be interpreted as zero timing.

Merge fresh layouts into infrastructure.template_layouts. Structural fields (index, master_index, name, and placeholders) come from extraction; speaker-curated use_for does not. Match prior use_for values by (master_index, name), because layout names are not unique across masters. Drop prior layouts absent from fresh extraction. When config.template_pptx_path is unset, emit an empty layout list.

Profile Field Ownership

Use every top-level key required by speaker-profile-schema.md. Set schema_version to 5, generated_date to today's YYYY-MM-DD date, and talks_analyzed to the count of all processed_talks.

The required top-level keys are schema_version, generated_date, talks_analyzed, speaker, infrastructure, presentation_modes, instrument_catalog, rhetoric_defaults, confirmed_intents, guardrail_sources, pacing, pattern_profile, visual_style_history, publishing_process, design_rules, and badges.

Keep pattern_profile as the only catalog-history storage lane. Never copy mastery, signatures, scores, usage, or another pattern aggregate into rhetoric_defaults. Never materialize catalog-derived guardrail_sources.recurring_issues or badges; consumers derive those directly from validated pattern_profile. Every top-level recurring issue and badge is non-pattern data with source_lane: "non_pattern".

Pattern-Profile Construction

Copy pattern_baseline unchanged to pattern_profile.pattern_baseline. Set baseline_talk_filenames to the sorted unique exact filenames from baseline_talks, copy pattern_baseline.eligible_talk_count to eligible_talk_count, and take talks_scored plus average_pattern_score from the baseline's raw-score-comparison fields. A non-empty eligible cohort can validly have zero scored talks and a null average when opportunity identities are mixed or no outcome is evaluable. Never turn missing opportunity coverage into a zero average.

Copy pattern_opportunities.pattern_usage and pattern_opportunities.antipattern_frequency unchanged. Their denominators are pattern-specific evaluable counts, not the global eligible count. Preserve exhaustive rows even for an empty cohort, including their canonical null rate and coverage sentinels.

Copy every field of the loader's deterministic pattern_classification bundle into pattern_profile unchanged: classification_schema_version, the self-contained classification_policy stamp, classification_availability, exhaustive pattern_classifications and antipattern_classifications, trend_analysis, and all derived projections. Those projections are score_trend, pattern_breadth, underused_patterns, score_drivers, by_mode, strengths, strengths_note, never_used_patterns, signature_combinations, and mastery_levels.

skills/vault-profile/scripts/classify-pattern-profile.py is the sole authority for domain availability, thresholds, and classification arithmetic. Copy its classification_availability output unchanged. Never reconstruct availability from raw rows or restate the classifier's predicates in profile-construction guidance. Consumers must honor each domain's own {status, reason_codes} and each row's evidence and reason codes. In particular, not_yet_observed is not never_tried, and an unclassified antipattern is not recurring or confirmed absent.

Do not relabel rows, derive tiers from raw rates, truncate exhaustive classification arrays, or infer an absence where absence_conclusion_capable is false. Profile regeneration re-analyzes already persisted current-generation outcomes. It does not reparse talks or modify the raw pattern_usage, antipattern_frequency, or tracking-database rows. See the generated profile's embedded semantic policy when the applied values must be inspected.

Empty-Cohort Behavior

When baseline_talks is empty, retain the zero-count baseline, empty filenames, and exhaustive zero-opportunity rows. Set score and breadth averages to null and pattern trend/direction fields to unavailable; classification rows remain exhaustive with their classifier-owned unavailable/unclassified states, while derived lists are empty. Never borrow pattern values from a prior profile, summary prose, an excluded scoring generation, or an instrumentation cohort.

Pacing and Section 15

Pacing is quantitative but approximate. Treat marginal slide-budget overages softly; the result complements, rather than replaces, Dimension 14's transcript-evident "rushing" assessment.

Use Section 15 of rhetoric-style-summary.md only for narrative consistency review. It cannot override or fill the current pattern cohort. The Section 15 reader accepts the occurrence-only v2 current block for compatibility, but the writer always replaces it with a policy-bound v3 block copied from the validated schema-v5 pattern_profile. Regenerating that block needs no talk reparse and never changes persisted raw rows.

Existing-Profile Diff

If no prior profile exists, skip the diff.

Otherwise compare pattern_catalog_fingerprint and pattern_scoring_schema_version first. A change in either is a pattern-generation reset: do not describe catalog-derived score, trend, usage, mastery, or strength differences across it as speaker improvement or regression.

Within one exact generation, do not report raw-score movement across a changed or unavailable opportunity_coverage_identity. Compare the classifier-emitted semantic_sha256 next. A changed digest is a derived-classification reset. Raw occurrence rows remain comparable, but do not attribute tier, severity, combination, underuse, or trend changes to the speaker. Within one unchanged policy identity, report a derived change only when its own availability domain is available in both profiles.

Report these same-generation non-pattern changes:

• New instrument_catalog instruments.
• Revised guardrail_sources thresholds.
• New non-pattern recurring_issues guardrails.
• Worsening pacing.adherence.trend or rising over_budget_rate.
• New presentation modes, prominently; this is the highest-signal creator behavior change.

Achievement Badges

Generate fun, self-deprecating badges from pacing, visual, publishing, talk-count, or confirmed-intent data. Use the speaker's voice rather than corporate gamification. Never derive a badge from catalog pattern usage, mastery, strength, antipattern, score, or absence. Set source_lane: "non_pattern" on every badge.