What this deliberately does not do

The refusals are load-bearing. Most of them are settled in decisions.md, which wins over brief.md wherever the two disagree, and none should be reopened without evidence and a name.

We catalogue that a bootleg exists. We never point at one and never host one

decisions.md §6. These are three different legal arguments wearing one word.

Cataloguing that a recording exists, its date, its cast — those are facts, and under Feist v. Rural nobody owns them, regardless of whether the underlying recording was authorised. Linking is narrower ground than people assume: Perfect 10 v. Amazon concerned thumbnails and does not generalise cleanly to pointing at infringing audio. Hosting is the Internet Archive's Great 78 Project, sued for up to $621M and settled in September 2025 on confidential terms.

There is a separate trap. ProCD v. Zeidenberg holds that a site's terms of use can bind contractually even over facts copyright does not protect, so "it's only data" is not a complete answer to a scraping question.

data-sources.md found no documented case of a musical-theatre-bootleg cataloguing site being sued or shut down. That is an absence recorded as a finding, and it is not evidence of safety.

Why bother at all: broadway-data holds over twelve thousand productions and commercial recordings exist for a small minority. If only commercial releases count as documentation, the honest answer to "what got recorded" is almost nothing, and overwhelmingly the hits. Historical bootlegs are often the only surviving document of a production, and sometimes of a work-state — the 1986 Old Globe Into the Woods predates substantial revision, and that tape documents material later cut.

decisions.md §11 confirms this holds identically for scripts, scores and pro shots. Catalogue that the thing exists. Never point at it, never host it.

We do not host anything

No hosted PDFs, no scores or scripts for download, no private tier. Editions are catalogued and linked out to, like everything else. Availability links out per release, per territory.

flows.md establishes why the ambitious version of that is closed rather than merely unbuilt. Spotify's developer policy bans using its metadata as a standalone product and bans replicating a core user experience, so a JustWatch-style grid comparing availability across services is a terms-of-service wall, not a sequencing question. Deep-linking someone into Spotify is sanctioned and is what we do.

The honest fifth state in that waterfall is nothing — this recording exists, here is what it documents, and there is currently no way to hear it. It should not be styled as a failure. seed/europeana-operetta.json makes the scale concrete: 654 of 766 harvested recordings carry In Copyright rights statements.

/Volumes/Desiree/broadway-data is read-only, every time

It is the owner's separate live project. Read it, never write to it, ask first. decisions.md §8 sets the relationship: two datasets, two centres, joined at the production. Neither absorbs the other. Corrections travel as proposals through contrib/, described in docs/contributing.md.

The scope gate: a Broadway or West End production admits the work

decisions.md §12. Once a work is in, all of its recordings are in, from any country and in any language. Chicago played Broadway, so its Korean, German, Hungarian and Japanese cast recordings are in, and they are among the most interesting things about it. Elisabeth has played Vienna, Budapest and Tokyo, has never played Broadway or the West End, and is out.

The gate is on the work, not on each production, so tours and foreign stagings of an admitted work are in. Professional stagings only. A DPLA-facilitated batch of 2,428 playbill pages turned out to be substantially community, civic and university theatre — "performed by the Players Theatre of Columbus" — and filing a Lee University Hello, Dolly! beside the 1964 Merrick production makes the production tier useless.

Two words that look decisive and are not. Repertory is usually professional — the American Repertory Theater, the RSC, most British reps — so a filter treating it as a marker of amateur status would throw out much of what this is for. Operatic society runs the other way in British usage and usually means amateur. Neither decides anything. The company gets looked up, not pattern-matched.

What it costs, stated: Elisabeth is the most performed German-language musical written and it is out, along with most of the Central European repertoire that decisions.md §4 brought into scope. Decision 4 is deferred rather than reversed. The model keeps city and language on Production even though almost every value will be New York or London for a long time, because hardcoding them away turns adding Vienna into a migration rather than an import.

The Show boundary is chronological, not generic

decisions.md §4. Operetta is in, opera is in, and the cutoff is a date rather than a genre. Nobody in the research could draw a genre line because there is not one. castalbumdb.com's maintainer wrote out his own dilemma — "If I include No, No, Nanette, why not The Vagabond King?" — and excluded operetta, which structurally removed the German, Austrian and Hungarian repertoire from the nearest predecessor to this project. German Wikipedia's Operette article counts Oklahoma! and The Sound of Music as sometimes-operetta, so the boundary is unstable from both directions.

Changing the axis dissolves Porgy and Bess, Sweeney Todd, The Most Happy Fella, Threepenny Opera and Weill generally, all at once. No border, nothing to police.

Where the date sits is not yet set.

Granularity stops at the recording

decisions.md §7. A chorus member differing on night two is not a distinction. If a separate recording exists there is a separate record; if it does not, there is nothing to model. No performance tier. A dated audience recording carries its own granularity.

No version tier

decisions.md §9, which rejected a proposal from the lead. Candide exists in six separately licensable versions, and the rights world maintains that taxonomy rigorously because money rides on it — Concord ships three non-interchangeable *Show Boat*s with a non-refundable-if-wrong warning. Replicating it here would mean inheriting an argument the Library of Congress and the Bernstein Office have not settled between themselves.

The cost is stated rather than hidden. A reader cannot ask which recordings document the 1974 version of Candide. They can ask what each recording calls itself and which work it belongs to. Accepted trade.

Recording types are positive subtypes, never a null

decisions.md §2. A recording is typed by what it documents, and the type decides whether a production parent applies at all. Cast recording and audience recording take a production; concept album, studio album, demo and workshop take none.

Bernstein's 1984 West Side Story is a studio album with no staging behind it — the recording was the event. The Les Misérables Complete Symphonic Recording is the same case despite drawing its cast from five productions across four cities, because it documents the work rather than any staging. Production is null, not many-valued.

Transcribe, then resolve

decisions.md §3, and the single most repeated failure in the research. The album is recorded as it describes itself on its own cover or spine, verbatim and never normalised. Separately, it resolves to a production. Search runs on the resolution; display honours the transcription.

Four systems collapsed the two jobs into one string. MusicBrainz has been arguing since 2017 about whether "Original Broadway Cast Recording" belongs in the title or the disambiguation field, with a May 2026 thread sitting at zero replies. castalbumdb.com used a free-text "Cast name" field. CastAlbums.org uses the text after a >. iTunes and Plex flatten the lot to "compilation", which is structurally true of every cast recording and semantically useless for all of them.

Library cataloguing formalised this split long ago — MARC transcribes the title statement as it appears on the item, and the authority-controlled heading is a separate field that does the finding.

Production is a tier, not an attribute

decisions.md §1, overturning brief.md. A production may have multiple recordings: Pearl Bailey's 1967 Hello, Dolly! came from the same Merrick production that opened with Carol Channing in 1964. Ovrtur validates the shape empirically — one 1981 production of Bring Back Birdie links three distinct recordings, two of them preview casts separated from opening night.

Video is modelled now and populated later

decisions.md §5. Pro-shots, NT Live, Shiki DVDs, Hamilton on Disney+, and NYPL's Theatre on Film and Tape archive. The media-type distinction goes into the model now because retrofitting it onto a populated audio-only tier is expensive and an empty field costs nothing. TOFT's holdings, access terms and scale were costed by nobody in this research, and for most productions video is the only documentation that exists.

Silence is not a prohibition, but a stated term is a term

decisions.md §14, settled by the owner on 15 August 2026 and confirmed on the 16th. Material we hold where no licence statement exists at all — neither granted nor refused, simply absent — may be held, may be shown, and may lead a page.

The live instance is 135 records harvested from eBay search-results pages, whose own licence_stated field reads "NO LICENCE STATEMENT OF ANY KIND… eBay states no terms for seller-uploaded images and the seller states none."

What this does not touch, and each of these remains in force:

• A stated restriction is still a restriction. The images carrying a Commons personality or costume value are shown blurred, because those rights belong to the person in the frame or the costume designer and were never the photographer's to license.
• A stated crop or integrity clause is still binding. web/cropping.py refuses by default, and the Carl Van Vechten estate's request to preserve the integrity of the image is honoured.
• Bootlegs are untouched. §6 stands.

The line is between silence and a statement. Silence is not a prohibition; a stated term is a term and we keep it.

It is recorded as a fact rather than laundered into a verdict. Every one of those records still carries its licence_stated text verbatim and publication_cleared: false. We are not asserting the images are cleared. We are deciding that the absence of a statement does not stop us showing them, and keeping the evidence so anyone — including a future us — can reach a different conclusion.

How it came to be written down is the part worth keeping. The decision was made on 15 August and not recorded. Two artefacts then began citing it to §11, which is about production tiers and says nothing about licences. The verify-images agent refused a teammate's go-ahead on the grounds that no teammate can authorise reversing an owner decision, then found that the decision it was being asked to reverse had no written source at all. A decision repeated from memory is a decision drifting.

Stated access terms are honoured without exception

Six theatre organisations exclude this class of agent by name. Where a source says no, that is the answer, and the fact that we asked gets written down.

Surface the confusion; do not resolve it silently

decisions.md §10, and the reason it belongs on a page about refusals: we refuse to quietly clean things up.

MusicBrainz carries a release-group titled "Chess (1984 London cast)" — 26 tracks with [dialogue] markers, no date, no country, no barcode. The London production opened in 1986; 1984 is the concept album's year. It is labelled with one thing's year and another thing's city and describes neither.

The wrong response is to file it correctly and move on. evidence-mining.md found a Chess — London vs Broadway thread where three posters each named a different recording as "the real London recording", none of them the actual 1986 West End cast, and one wrote "something I didn't realize for a long time is that the concept album doesn't directly reflect the London stage show." Every system surveyed either flattens that disagreement into one string or argues about it out of public view. Nobody shows the argument to the person who came to look something up.

What it does not mean: an excuse to leave data unresolved. Resolve what can be resolved and surface the residue with its reasoning attached.

Every factual claim carries a source

A claim that cannot be sourced is a hypothesis and gets marked [HYPOTHESIS] inline, where it sits. Absence of evidence is itself a finding — "I searched for X and found nothing" is worth writing down, and the silence does not get filled.

If a sentence starts "users often feel" or "the market generally", it is wrong. Name who said it and where.