Description

The audit skill already emits insights (durable/time-bound). Extend that schema to capture the richer analysis requested: historical/era context (the 2010-2015 moment), conference + topic context, interviewee context, “lessons for now,” followup questions, short-worthy quote timestamps, and cross-interview link anchors. Update the audit prompt (backlog/audit/outbox generation) to produce the extended fields, keeping backward compatibility with already-ingested insights.

Acceptance Criteria

• #1 The insight schema is extended with the new fields and validated
• #2 The audit prompt is updated to generate them
• #3 Existing ingested insights remain valid (backward-compatible)
• #4 Verified end-to-end on 2-3 interviews

Implementation Plan

Scope

Extend the ONE canonical audit path only. Leave bin/archive/modules/enrich.rb and bin/cerebral_enrichment.rb untouched — they are direct-LLM scripts TASK-242 is retiring; do not spread the schema onto engines about to be deleted.

Investigation findings

• Canonical audit prompt = src/generators/transcript_processor.rb SYSTEM_PROMPT (OUTPUT SCHEMA lines 33-48). bin/prepare_audit_prompt.rb extracts it into backlog/audit/outbox/<slug>.md; rake transcript:process uses it directly. Single source of truth.
• Ingest bin/ingest_audit.rb copies ONLY speaker_map/turns/insights/youtube (lines 53-56); new top-level keys are dropped unless added here.
• Current insight shape: insights: [{statement, type: durable|time-bound, confidence: high|medium}].
• TIMECODE CHECK (verified corpus-wide): 196/196 transcripts have NO content, NO timecoded turns. So timestamp has NO source until TASK-247 diarization lands (this is why 251 deps on 247). The field is RESERVED/optional, populated later from diarized turns — the LLM must NEVER invent MM:SS.

Extended schema (all additive/optional → backward compatible; add schema_version: 2, absence = v1)

• Per insights[] item, add: quote (verbatim short-worthy line, audit-selected), timestamp (MM:SS — reserved, filled from 247 diarized turns, OMIT until available), anchor (stable id e.g. insight-1, the cross-link target 254 references).
• New top-level context: block — subkeys era, conference, topic, interviewee, each {summary, sources: [], uncertain: bool}. 251 defines the container + the audit emits a transcript-grounded first pass; TASK-255 fills the citable/sourced external layer.
• New lessons_for_now: [{statement, from_anchor}] — “reads now” framing.
• New followup_questions: [string].
• New cross_links: [{anchor, related: []}] — shape reserved; TASK-254 populates related.

Files to touch

1. src/generators/transcript_processor.rb — extend SYSTEM_PROMPT OUTPUT SCHEMA + rules: quote verbatim only; DO NOT invent timestamp (leave absent unless diarized turns supplied); context summaries grounded in transcript, mark uncertain: true for anything not directly stated.
2. .gemini/skills/transcript-conversational-audit/SKILL.md — update step 2 generated-field list.
3. bin/ingest_audit.rb — persist new top-level keys (context, lessons_for_now, followup_questions, cross_links, schema_version) alongside existing four.
4. Rakefile transcript:process apply block (~L335) — same additive persistence.
5. Validator — extend rake validate:audit_transcripts (Generators::ArchiveManager::TranscriptAuditor) or add bin/validate_insight_schema.rb wired into validate:all: new fields well-formed when present; v1 records (statement/type/confidence only) still pass (AC#3 backward compat); assert no timestamp present unless a diarized turn exists.

Verify

rake audit:prepare[slug] → generate → rake audit:ingest[slug] on 2-3 FRESHLY audit-processed slugs (not the stale near-duplicate enrich.rb output). Validator green on both new + existing v1 files. rake build clean.

Definition of Done

• #1 AC criteria is completed and the change has been verified