Analysis quality audit

Status: freeze sheet for the 25 core module ids
Pins: analysis_port_pins.md
Severity: release_severity_triage.md
Porting map: analysis_module_porting.md

Living sheet — one row per registered MODULE_ID. 0.9.x / 1.0 freeze: no new modules unless an audit proves a release-critical repair. Do not rewrite these to TranscriptX-depth stacks (sklearn LDA, spaCy NER, speaker momentum).

semantic_class is copied from the pin registry. A unit test fails if a registered MODULE_ID disappears from that list.

Honesty on thinner algorithms

These notebook adaptations are weaker than the TranscriptX originals and must stay labelled as such:

Module	Algorithm in Transcribe	Must not silently become
topic_modeling	Deterministic seed-bucket topics (topic_seed_buckets_v1); no sklearn LDA	A different topic engine under the same id
keyphrases	Notebook TF-IDF unigram/bigram ranking	spaCy / TX keyphrase stack
wordclouds	Baseline token frequencies (enrichment_mode: baseline)	Keyphrase-enriched clouds without a version bump
bertopic	Optional extra; missing extra → unavailable_extra	Seed-bucket / LDA substitute under this id
moments	Notebook salience fork (no TX momentum)	Speaker-momentum scoring

Optional extras and LLM modules that cannot run must abstain honestly (unavailable_extra / unavailable_model), never invent a substitute payload.

Rows

module_id	semantic_class	Intended question	Algorithm honesty	Recommendation	Severity
stats	adaptation	How long / varied is this notebook?	Page/unit distributions; no speaker rollups	retain	—
lexical_diversity	adaptation	How diverse is the vocabulary?	Document + per-unit metrics	retain	—
understandability	adaptation	How readable is the text?	Pure-Python readability (no nltk/textstat)	retain	—
wordclouds	adaptation	What words dominate?	Baseline frequency path	retain; keep baseline labelled	known limitation vs TX viz
ner	adaptation	Which names and places appear?	Notebook NER adaptation	retain	—
sentiment	adaptation	What is the tone by page?	Lexicon / unit sentiment; no speakers	retain	—
epistemic_markers	adaptation	How certain is the language?	Lexicon markers	retain	—
keyphrases	adaptation	What phrases stand out?	Notebook TF-IDF	retain; document thinner stack	known limitation vs TX
entity_sentiment	adaptation	How does tone attach to entities?	Join NER × sentiment	retain	—
topic_modeling	adaptation	What is this notebook about?	Seed-bucket topics; no sklearn	retain; document thinner stack	known limitation vs TX
semantic_similarity	adaptation	Which pages are alike?	BoW cosine; multi-speaker gate dropped	retain	—
topic_shift	adaptation	Where does the subject change?	Order-based cosine-drop	retain	—
bertopic	adaptation	Clustering when the extra is installed	Optional extra; no silent substitute	retain as optional	known limitation
emotion	adaptation	What emotions show in the text?	Lexicon emotion	retain	—
contextual_emotion	adaptation	How does emotion move across pages?	Order-neighbor window	retain	—
fine_grained_emotion	adaptation	Finer emotion labels	Optional extra → unavailable_extra	retain as optional	known limitation
affect_tension	adaptation	Where does affect tighten?	Join adaptation	retain	—
moments	fork	Which pages are salient?	Notebook salience; no momentum	retain as fork	known limitation vs TX
highlights	adaptation	Which quotes matter?	Notebook salience quotes	retain	—
summary	adaptation	What is the brief?	From-highlights executive brief	retain	—
insights	adaptation	What themes compose?	Highlights + topics; notebook_eligibility_v1	retain	—
llm_summary	adaptation	Abstractive summary	Optional Ollama; abstain offline	retain	known limitation (model)
llm_action_items	adaptation	Actions / decisions / questions	Optional Ollama	retain	known limitation (model)
llm_custom_qa	adaptation	Ask the notebook	Grounded QA; refuse unsupported	retain	known limitation (model)
narrative_summary	adaptation	Longer narrative	Optional Ollama; unavailable_model offline	retain	known limitation (model)

Mandatory scrutiny

Deterministic highlights / summaries vs LLM equivalents — keep both, but do not present an LLM skip as if the deterministic brief were the same claim. Seed-bucket topics, TF-IDF keyphrases, and baseline wordclouds stay adaptations, not parity with TranscriptX.