Melba Liston — centennial data extraction

What jazz musicians said about Melba Liston (13 January 1926 – 23 April 1999; trombonist, arranger, composer) — and what she said about them — pulled from the Linked Jazz 2026 corpus for her 2026 centennial.

This repo is the data layer only. The web page comes later; nothing here renders anything.

It is a port of the Coltrane / Davis build (~/git/linked-jazz-coltrane-davis, published as ~/git/centennial-2026): same corpus, same conventions, same passes. Where it departs from that build, it is because of one fact — see the next section.

Source: ~/git/ch-jazz-mashup/linked_jazz.sqlite — 1,347 oral-history interview transcripts from four archives (Tulane/Hogan, Hamilton/Fillius, Smithsonian/NEA Jazz Masters, Rutgers/IJS). The DB is opened read-only everywhere and is never modified.

The one thing that makes her different

Coltrane and Davis were never interviewed. Liston was. The Smithsonian Jazz Oral History Program recorded her on 4–5 December 1996 at her home in Los Angeles; the interviewer was the trumpeter Clora Bryant, a contemporary and friend. So where the earlier build could only look one way — what other people said about the subject — this one looks three ways:

direction	where	size
others → her	24 other interviews	49 blocks about her, 20 people who actually spoke
her → others	her own interview	106 people, 66 with a line in her own words
her → herself	her own interview	173 lines, 53 timeline facts
(and Bryant → her, to her face)	her own interview	83 lines

The first row is the whole of what the earlier build had, and for Liston it is small: she is named in 24 interviews against Coltrane's 299. The centennial page will lean on the other three.

Two properties of her interview shape everything downstream:

• She had a stroke in 1985. Eleven years later she has memory loss and difficulty speaking. Her turns are short — many are a bare "Yeah." — and Bryant, who knows her life, often tells it and lets her confirm.
• Square brackets are the transcriber. Where she could not answer, the transcriber substituted a summary: "[Liston cannot recall what happened next.]". Those are never her words. No quote attributed to her anywhere in this repo contains one (checked programmatically, see verify_all.py).

Layout

liston/
  profile.json               identity, links, community, headline stats
  documents.json             the interviews that name her (+ text_hit_only_documents)
  --- others -> her ---
  quotes.json                every block that may be about her, with context
  relationships.json         upstream LLM relations to her (19; weak -- see witnesses.json)
  connections.json           network edges to every other person (123)
  cooccurrence.json          entities named in the same block (her own interview excluded)
  nicknames.json             raw text sweep for Melba / Liston -- the recall audit
  eval_input/ eval_output/   chunks to and from the block-level judgement
  enriched.json              ★ every block: window + judgement + provenance
  best_quotes.json           ★ the verbatim-checked display shortlist (21)
  false_positives.json       blocks judged not about her (1)
  recovered_*.json           pronoun follow-ons and answers; recovered_best.json ★ (2)
  witness_input/output.json  bundles to and from the person-level judgement
  witnesses.json             ★ one row per witness: relation, caption, lead quote, caution
  --- her -> others, her -> herself ---
  own_interview.json         her whole interview, block by block, roles resolved
  own_interview.txt          the same as reading text (what the judges read)
  own_interview_entities.json  places, bands, tunes, venues named in it (raw NER)
  her_relationships.json     upstream LLM relations FROM her (115; 59% wrong -- see below)
  her_words_input/ output/   chunks to and from that judgement
  her_words.json             ★ one record per person she spoke of, judged
  own_voice_output/          raw reading-pass output
  own_voice.json             ★ self (173) · bryant (83) · timeline (53)
  --- records ---
  discography.json           ★ 80 releases with her credit on each
  discography_personnel.json ★ 582 people on those records + the two-way interview join
  --- everyone ---
  people.json                ★ one record per person across all four layers (487)
shared/
  images.json                ★ portrait per person: Commons URLs, licence, attribution
  cover_art.json             album sleeves from the Cover Art Archive (38)
  communities.json           all 58 network communities; 2 is hers
  class_of_1926.json         her beside Coltrane and Davis, the other 1926 centennials
  *_summary.json             stats for each pass
img/
  p/<QID>.jpg                314 portraits (200px; hers also @2x)
  a/<MBID>.jpg               38 sleeves
  CREDITS.json               licence + attribution for every file
extract/                     transcript-derived layers (stdlib only) + the five SPEC.md files
discography/                 Wikipedia / Wikidata / Commons layers; cache/ and raw/
build_all.sh                 every deterministic step, in order

★ = what a page should read. Everything else is working material or audit trail. DATA_MODEL.md describes the starred files field by field.

How the judgement was done

Every decision that needs reading — is this block about her, what did she say about this person, which line leads — was made by Claude Opus 5.5 subagents, each working to a written spec in extract/. The earlier build used a mix of subagents, Gemini and a local Qwen for the same jobs; here it is one model throughout. Twelve agent runs in all:

pass	spec	agents	input → output
block evaluation	EVAL_SPEC.md	3	118 blocks in windows → eval_output/
recovered blocks	RECOVERED_SPEC.md	1	22 pronoun/answer blocks → recovered_output.json
her words	HER_WORDS_SPEC.md	4	117 detected "people" → her_words_output/
own voice	OWN_VOICE_SPEC.md	2	the whole transcript → own_voice_output/
witnesses	WITNESS_SPEC.md	1	25 bundles → witness_output.json
album credits	(in the prompt)	1	7 uncertain credits → raw/credit_review_output.json

The her-words and own-voice readers each read the entire transcript (1,057 blocks, ~125k characters) before judging anything, which is what the windowed approach of the earlier build could not do: an answer on page 40 can settle a question from page 9.

No judge's quote is trusted. Every merge script re-verifies every quote as an exact substring of one source block, spoken by the right voice, outside any transcriber bracket, and drops what fails. extract/verify_all.py then re-does the whole check independently against the database rather than the intermediate JSON: 698 quotes checked, 0 failures, and no quote had to be dropped at any merge step.

Merges of duplicate people (five of them) were recommended by the readers and are recorded with their reasons in build_her_words.py → MERGES; the absorbed records are kept under her_words.json → merged.

The numbers

interviews naming her (other than her own)	24 — 19 reconciled to her QID + 5 found only by text search
blocks about her in those interviews	49 of 50 evaluated (1 false positive)
…where the wider context changed the reading	44% of all 118 evaluated blocks
people who said something about her in their own words	20 (21 interviews; Slide Hampton twice)
display shortlist (best_quotes.json)	21 — 12 from other interviews, 9 Clora Bryant
recovered pronoun/answer blocks really about her	5 of 22 (23%)
people named in her own interview	117 detected → 106 real people after judgement
…with a relation established	84
…with a line in her own words	66 (156 verified quotes)
her lines about herself	173 (12 at the top grade)
Clora Bryant's lines about her	83 (8 at the top grade)
biographical timeline entries	53 (44 confirmed by Liston)
releases	80 — 72 with an article, 8 listed without one
…credited as player / arranger-conductor / composer	41 / 24 / 18
people credited on those records	582 (416 with a QID)
…who were interviewed in the corpus	86
…and said something about her	10
…whom she spoke of	36
people with a portrait	313 of 459 (68%); 19 of the 20 witnesses
network edges	123 (72 to reconciled people)

Eight people appear in both directions — they spoke of her and she spoke of them: Randy Weston, Slide Hampton, Gerald Wilson, Clora Bryant, Al Grey, Abbey Lincoln, Benny Powell, Vi Redd. Five of those are also on her records and in the network — all four layers: Weston, Hampton, Grey, Lincoln, Powell.

What came out of it

The upstream model was wrong about her interview more often than it was right. Of 113 relations it classified from her, the readers judged 67 wrong (59%). The cause is the stroke: she answers "Yeah.", so the model took what Bryant said for what Liston said. It also typed 16 non-people as people — six tune titles (Little Niles, Len Sirrah, Pam's Waltz, Ben Loves Lu…), a playground, an insurance company, "Saxophone player". her_relationships.json is kept as the audit trail; use her_words.json, and filter on basis:

basis	meaning	people
her_statement	she said it, in her own words	53
her_assent	Bryant said it, she agreed — the words are in bryant_quote	16
transcriber_summary	only a bracketed summary supports it	14
bryant_only	Bryant said it; Liston did not confirm	7
none	no relation established	16

The name-matching missed a fifth of the interviews that mention her. Only 40 blocks outside her own interview reconcile to her QID. A text search for a bare "Melba", "Liston, Melba" and "Melvin Liston" (extract/subject.py) found 10 more blocks in 5 interviews the person layer had not linked at all — Mary Lou Williams, Quincy Jones, Shirley Horn, Von Freeman, J.C. Higginbotham — and the evaluators confirmed all 10 as her. Williams's is one of the best passages in the set.

She deflects praise. Of Bryant's 83 remarks, Liston elaborates on 32, confirms 17, deflects 16 and denies 10. Bryant: "that is one of the most beautiful things you've ever done" — Liston: "I don't know." That pattern is in own_voice.json → bryant[].liston_response with the reply text alongside.

Her own top lines (own_voice.json → self, notable: 3):

"I just loved the instrument. I didn't need no inspiration."

"I do it all from my soul, and if it sounds like somebody, well it sounds like somebody. It's my inspiration. It's from my soul."

"I'm a woman. That's that."

"But blowing, I don't think about nothing but the notes."

"I'm sorry that I couldn't talk more."

What the witnesses said (witnesses.json, top of the list):

"She played really good, and she wrote music that weighs as much as this building." — Slide Hampton

"Then I realized that only Melba Liston could do this." — Randy Weston

"you have been the best, Melba. You've been a guiding light for a lot of people, including me" — Clora Bryant

Her discography is an arranger's. Her own Wikipedia article lists 63 records, filed as "sidewoman or guest". Sweeping the album articles that link to her found 17 more that her list omits — records she wrote for rather than played on (White Gardenia, For the First Time, Hub Cap, When I'm Alone I Cry), and nine where someone else recorded her tunes. subject_credit on every release says where the credit was found and what it is.

Read this before building the page

1. Every witness row carries a caution. All 25. Read them. Examples that would mislead if displayed bare: Shirley Horn's "a great jazz artist" is followed at once by "there are certain instruments I think a woman does not have the strength to play" — the row's stance is mixed. Abbey Lincoln's "It's a rare woman, like Melba Liston, who comes to the music for the music" sits inside a put-down of women musicians generally (stance_target: women_in_jazz). Curtis Fuller says Gerald Wilson was her husband — the judge flagged that as doubtful and nothing else in the corpus supports it.
2. spoke_themselves: false is not a witness. Four rows: Quincy Jones, Hank Jones, Paul Howard, and the Snooky Young / Gerald Wilson joint interview. In each her name is in the interviewer's mouth or the archive's front matter. In the Quincy Jones transcript every claim about her (that she wrote Reverie for his band, that she inspired the women of Diva) is the interviewer's, David Baker's — do not present it as Jones's testimony.
3. One witness is not the interviewee. In the Clark Terry interview the man talking about Liston is the co-interviewer, singer Joe Williams; Terry says nothing about her. The row is Williams's.
4. Bryant's facts are Bryant's. She supplies dates and events from memory and from notes, and Liston often cannot confirm or flatly denies them (recording with Basie; the Ellington influence). Timeline entries with stated_by: "bryant" and confirmed_by_liston: false are not her testimony.
5. documents.interviewee is wrong for two of her witnesses. The Teddy Edwards and Paul Howard transcripts both name the interviewer, Patricia Willard, as interviewee. witnesses.json → name is right; enriched.json → doc.interviewee_field is not. Several Smithsonian names are stored surname-first ("Akiyoshi Toshiko", "Horn Shirley").
6. Some of the best lines are not quotable. Quentin Jackson's "a great trombonist, too" and Higginbotham's "She can play, too, boy" sit in blocks that do not name her, outside any judged block, so the verbatim rule could not admit them. The evaluators recorded them in notes and summary.
7. Four people in her interview still need splitting by hand — see her_words.json → unmerged_leads. The main one: the detected "Minnie Hightower" folds a schoolmate with her mother, the teacher (Alma Hightower, who led the Melodic Dots). The record describes the mother; "Alma" is not in the transcript and was supplied by a reader.
8. Sensitive material exists. She describes a bandmate coming into her hotel room on the road (block 295771; her line "that's what girls have to go through" is in the next block), and Randy Weston's account of getting her to write again after the stroke includes "I insulted her… called her all kinds of names". Both are flagged in notes / caution, not filtered.
9. The discography is "what Wikipedia documents", not a sessionography. Three records her own article lists are not credited to her on the album's page (subject_role: "listed_unverified"): At Basin Street East, Mary Lou Williams, That Lovin' Feelin'. Eight more have no article at all. Release kind is mostly the generic album — her article has no studio/live split.
10. A composer credit is not a session. Nine releases carry her only on a track line (subject_role: "composer") — someone recorded Just Waiting or Melba's Blues. She was not necessarily in the studio.
11. Check review_flags before using a portrait large, and print attribution — 211 images require it. 49 are Linked Jazz thumbnails with no licence metadata (not_commons_licence_unverified).
12. Cover art is not free. 38 sleeves from the Cover Art Archive are in img/a/, copyrighted and supplied for identification. Using them was an explicit editorial call in the Coltrane / Davis build; it is carried over here on the same terms and should be re-decided before publishing.
13. Confidence is useless for ranking; notable is quota'd per chunk. Same as the earlier build. Rank by notable, then relation weight.
14. A Wikidata duplicate. The corpus authority file holds Quentin "Butter" Jackson = Q99676854, a bare item with no dates or article, beside the real Quentin Jackson Q720707. Witness rows resolve to Q720707; the roster still carries one credit under the stub. Worth merging upstream.

The page

docs/ is the site, laid out for GitHub Pages ("deploy from branch", folder /docs). It follows the design in the claude.ai/design project Melba Liston Centennial Page: hero, Wikipedia blurb and links, timeline, what people said, her own words, discography.

file	what
docs/index.html, style.css, app.js	the page: static, hand-written, no framework, no build step
docs/data.json	everything the page shows, built by site/build_site_data.py
docs/transcripts.json	the transcript text behind every quote, shown in a popup: her interview whole, each witness's interview as a window (10–16 turns either side) around the quoted turn, each with the archive's own URL. The page does not link to the transcript reader. Built by the same script, from linked_jazz.sqlite
docs/img/p/, docs/img/a/	only the portraits and sleeves the page uses, copied from img/
docs/img/liston.png, linked-jazz.png	the design's own two images (not generated)
site/lanes.json	the career bars above the timeline dots, hand-authored, each with its source
site/editorial.json	display decisions the data cannot make by rule: the tagline, and per-witness hide / second quote / context note, each with the caution it follows from

uv run python site/build_site_data.py          # after any change under liston/
cd docs && uv run python -m http.server 8765   # preview (data.json is fetched, so file:// will not do)

What the page does with the cautions above: a witness gets a card only if they spoke themselves and have a verbatim lead quote: 17 of the 24 people in witnesses.json. Quincy Jones, Hank Jones, Paul Howard and Snooky Young never speak of her themselves, Maxine Sullivan and Hayes Pillars have no quotable line, and Higginbotham is hidden by editorial.json. Timeline dots are the interview's own 53 entries; the 22 she gave no year for are placed between their dated neighbours and drawn grey. A line of hers that needs its question (stands_alone: false) is shown with the reading pass's one-sentence summary under it.

"Who was on the records"

Inside the Discography section. Counted only on the 67 releases she played on, arranged or led (a composer credit is not a session): 464 named musicians, 176 on more than one, 22 on eight or more. The page lists everyone on three or more plus anyone with words in either direction (123 people). Nine spoke of her (their card is the same vetted quote as in "What people said"); she spoke of 23 (her_words.json, only her_statement and her_assent; where she only agreed, the reading pass's summary is shown beside or in place of a quote). The six she recorded with most often (Clark Terry, Jerome Richardson, Jimmy Cleveland, Ernie Royal, Joe Newman, Phil Woods) are in neither group, though three of them have oral histories in the corpus. Choosing a person filters the release grid to the records they share. Portraits flagged not_commons_licence_unverified or picked_from_category_not_curated are not used (the latter gave Oliver Nelson a photograph of a car).

Sections of "In her own words"

Every pull quote in own_voice.json (173 hers, 83 Bryant's) carries a category: one of ten page sections, listed in own_voice.json → categories. The pass is extract/CATEGORY_SPEC.md: one Opus subagent designed the scheme from the quotes, two sorted all 256 independently, a fourth settled the one quote they split on and reviewed the 32 others either had doubts about, overruling the scheme's rules on two. The two sorters agreed on 255 of 256, but that measures how tightly the scheme's 26 boundary rules are written, not how obvious the sections are. The older multi-label theme is still on each item.

Rebuilding

./build_all.sh        # every deterministic step, then the verification gate

It is safe to re-run: the Wikipedia/Wikidata/Commons responses are cached under discography/cache/, images already on disk are skipped, and the judgement outputs are read from the tree. The script's comments mark where each judgement pass reads its input; if one of those inputs changes, that pass has to be re-run (a subagent, the spec, the input file) before the merge step after it.

What was not ported

build_site_data.py and build_site.py — the page bundle and the page — are presentation and were left out on purpose. Three scripts they depended on are replaced rather than ported: reclassify_knows_of.py and gen_one_liners.py (local Qwen) and classify_recovered_gemini.py (Gemini) are now the witness and recovered passes above. shared/coltrane_davis.json, the two-subject bridge, has no equivalent with one subject; shared/class_of_1926.json is the nearest thing — Liston beside the other two 1926 centennials, including the one piece of direct evidence that build could not have: asked whether John Coltrane was in Dizzy Gillespie's band when she joined, she said yes.