Analysis quality audit judgements (0.9.7 draft)

Status: provisional agent draft; Hub-card residual dropped 2026-08-07 — not a blocking owner Hub pass Scaffold (machine rows): analysis_quality_audit_scaffold.md
Column authority: analysis_quality_audit.md
Severity rules: release_severity_triage_1_0.md

Judgements live here so make docs-gen can regenerate the scaffold without wiping human columns.

Mandatory scrutiny

Module id	Recommendation	Severity	Notes
highlights	retain	known limitation	Deterministic quote-forward; useful on English meetings. Relabel in UI as deterministic when LLM summary is also present. Multilingual quality varies.
summary	improve	must-fix	Deterministic executive brief derived from highlights. Must not read as AI-authored. Overview now badges it Deterministic vs Local AI. Keep below Local AI when both exist (existing precedence).
llm_summary	retain	must-fix (labelling)	Local AI abstractive summary. Local AI badge required on Insights/Overview (shipped 0.9.7).
narrative_summary	retain	must-fix (labelling)	LLM narrative grounded on deterministic summary. Same Local AI labelling.
llm_action_items	retain	must-fix (labelling)	Keep HUMAN_REVIEW_BANNER + Local AI badge. Do not present as ground truth.
insights	retain	post-1.0	Content-first layer; prominence/tuning can continue post-1.0.

Full registry (provisional)

Module id	Recommendation	Severity	Notes
acts	retain	known limitation	Heuristic dialogue acts; English-biased.
conversation_loops	retain	post-1.0	Light T0; niche.
contagion	retain	known limitation	Depends on emotion; experimental-ish interpretation.
emotion	retain	known limitation	NRC lexicon association — not neural affect.
contextual_emotion	experimental / known limitation	known limitation	Experimental classifier; keep off Balanced defaults.
fine_grained_emotion	experimental / known limitation	known limitation	Experimental; opt in via Thorough / Custom.
entity_sentiment	retain	known limitation	Needs NER+sentiment; sparse transcripts weak.
affect_tension	retain	post-1.0	Derived indices.
interactions	retain	—	Core interaction analysis.
ner	retain	known limitation	spaCy model / language limits.
semantic_similarity	retain	—	Epoch-1 module id (ex-v2).
sentiment	retain	known limitation	VADER default; transformers optional.
epistemic_markers	retain	post-1.0	Light T0.
keyphrases	retain	—	Shipped 0.8.x; YAKE/KeyBERT optional backends.
stats	retain	—	Baseline stats.
topic_modeling	retain	known limitation	Heavy; min-data sensitive.
bertopic	retain	known limitation	Stack is optional ([bertopic]/[full]/Docker) after leaving base so core/clean-env is not blocked by llvmlite; module still registered — see bertopic_optional_module.md.
transcript_output	retain	—	Human-readable export helper.
simplified_transcript	retain	post-1.0	Convenience.
understandability	retain	known limitation	Heuristic readability.
lexical_diversity	retain	—	TTR/MTLD family.
wordclouds	retain	post-1.0	Visual; not analytical truth.
tics	retain	—	Feeds insight eligibility.
transcript_quality	retain	known limitation	Needs ASR confidence when present.
insight_eligibility	retain	—	Shared pipeline; not user-primary.
temporal_dynamics	retain	post-1.0
qa_analysis	retain	known limitation	Depends on acts quality.
pauses	retain	—	Timing/silence.
echoes	retain	known limitation	Embedding-sensitive.
politeness	retain	known limitation	Marker heuristics; cultural limits.
momentum	retain	—	Stall/flow.
topic_shift	retain	—	Chapter segmentation.
moments	retain	post-1.0	Ranked revisit moments.
llm_speaker_summary	retain	must-fix (labelling)	Local AI badge (0.9.7).
llm_custom_qa	retain	must-fix (labelling)	Optional Ollama; label AI.
chart_descriptions	retain	must-fix (labelling)	Finalize-phase LLM; label AI.
voice_features	retain	—	Prerequisite for voice stack; privacy-gated.
voice_mismatch	retain	known limitation	Needs voice + text.
voice_tension	retain	known limitation
voice_fingerprint	retain	known limitation	Speaker-identity sensitive — privacy notice v2.
prosody_dashboard	retain	post-1.0
voice_charts_core	retain	post-1.0
voice_contours	retain	known limitation	Slow; audio decode.

Hardening backlog (severity-tagged)

Finding	Severity	Action in 0.9.7
LLM surfaces without clear AI label	must-fix	Local AI badges on Insights/Overview LLM blocks
Deterministic summary looking like AI	must-fix	Overview badge Deterministic
Voice privacy copy missing	must-fix	VOICE_PRIVACY_USER_NOTICE + notice v2
Experimental emotion modules in Guided defaults	known limitation	Applied 0.9.8: removed from Balanced heavy allowlist (fine_grained_emotion / contextual_emotion off Guided defaults); Thorough/Full still may include. No module delete.
Perf envelopes incomplete on large library	closed	Large-library UI soak pass 2026-08-07 (200+ transcripts); Medium Balanced batch pass 2026-08-07 (~9.3 min / 6 transcripts)
Chart LLM narratives unlabelled	must-fix	Local AI badge on Charts gallery LLM text (0.9.7)
Custom Q&A unlabelled	must-fix	Local AI badge on Insights custom QA block (0.9.7)
RTD hostname not live	known limitation	Flip checklist ready; owner slug gated

Owner sign-off

• Judgements reviewed

• Any remove / deprecate decisions confirmed (none proposed as hard delete in this draft)

• Severity backlog accepted for RC triage