# museum/data — flat tables over the visits

Written by `node museum/export.js` (re-runnable; run it after every lab run). Plain CSV, UTF-8, one header row; text cells with commas, quotes or newlines are double-quoted. Never edit by hand.

## artworks.csv — one row per artwork per collection (48 rows)
- museum_no: the number on the wall (the same number can mean different pieces in pilot and final)
- collection: `pilot` (the provisional hang, visits 20261020-05..08) or `final` (the curator's 24, her titles)
- wing: ESCAPE | ESCHATOLOGY | VIGIL
- model: OpenRouter-style id of the artist
- family: claude | gpt | gemini
- triptych: the entry directory under entries/triptych-w1/
- title: the title as hung (museum/labels.json or labels_pilot.json)
- oversize: y if the piece is over the byte cap

## visits.csv — one row per finished visit (543 rows)
- visit_id: directory name under museum/visits/
- museum_version: v1 (tools.json; enter_wing says it lists titles and labels) | v2 (tools_v2.json, brief 2026-10-22; enter_wing's description neutral) | v3 (tools_v3.json, brief 2026-10-25; titles only, the exit is postcard -> guestbook -> question)| v4 (tools_v4.json; as v3, one tools list, a text answer)
- run (v4): `v4` (the nine visitors of 2026-10-25) | `v4-visitors` (the ten new visitors, brief 2026-10-26; same museum, same files) | `t30` (ticket 30, brief 2026-10-26 ticket30: tools_t30.json, the welcome says 30 calls) | `v4-sonnets` (Sonnet-4.6 medium + Sonnet-5.5 high, ticket 10, curator in chat 2026-10-27) | `t30-named` (ticket 30, artists named, brief 2026-10-27 named: labels_named.json, tools_named.json, no exit question) | `t30-swap` (ticket 30, two names swapped, brief 2026-11-07 museum swap: labels_swap.json, tools_named.json, no exit question) | `t30-topup` / `t30-named-topup` (curator in chat 2026-10-28: one new visit per slot of t30 / t30-named that ended in a provider ERROR, same visitor, settings and files; pool with the run it tops up, see museum/visits/_topup_run.json); blank before v4
- ticket: looking calls on the ticket (10 for every visit before t30; 30 in runs t30, t30-named and their -topup runs)
- labels: hidden (labels.json, no artist; every visit before t30-named) | named (labels_named.json: the artist's display name opens every label) | swap (labels_swap.json, run t30-swap, brief 2026-11-07 museum swap: labels_named.json with the artist names of N-45 and T-67 exchanged, N-45 labelled Claude Opus 5.5, T-67 GPT-6.1 Sol; artworks.csv's model column is still the TRUE artist)
- arm: A (enter_wing returns number + title + label; every v1 visit is A) | B (v2 only: enter_wing returns one wall image of the 8 pieces with their numbers, museum/wall.js)
- collection: pilot | final, from the collection sha256 recorded in that visit's own log
- visitor: model id of the visitor; visitor_family: claude | gpt | gemini
- effort: reasoning effort sent
- exit: STOPPED EARLY | TICKET USED | SILENT -> SHOP | SILENT | REFUSED | EMPTY | SPEND STOP
- calls_used: looking calls spent (of 10), invalid ones included
- wings_listed: y if list_wings was called; first_wing: the first wing entered (blank if none)
- n_frame, n_code, n_statement: number of views of each type (repeats counted)
- postcard_no, postcard_model, postcard_family: the postcard bought (blank if none)
- own_bought: y if the postcard is the visitor's own work; family_bought: y if it is by the visitor's family (own included)
- guestbook: the note, verbatim (v3: from sign_guestbook)
- own_listed: y if any of the visitor's own pieces hung in a wing it entered; own_viewed: y if it viewed any of them (blank for a visitor with no work on the walls). Every version, from the log
- end (v3 only): how the visit ended after the shop: ANSWERED | SIGNED (t30-named: no question, the visit ends at the guestbook) | SILENT AT GUESTBOOK | SILENT AT QUESTION, or the shop-side end (SILENT, REFUSED, ...). `exit` still says how the visitor reached the shop
- q_answer (v3): the answer to the exit question, verbatim; q_why: its line, verbatim; q_artist: the artist of the named piece
- q_kind (v3): own | family | other | none | invalid (not a museum number); blank if never answered
- q_shown (v3): viewed | listed (in a wing entered, not viewed) | never (not shown this visit) | n.a. (none/invalid)
- v4 (tools_v4.json, brief 2026-10-25 v4): one tools list every turn; the question answered in text. q_text: the reply verbatim (for a tool call: "[tool call] name {input}"); q_answer: "none" if the text contains none as a word (amendment 2, summary.question_amendment2), else the first collection museum number; q_multi: y if more than one distinct collection number appears; q_parse: number | none | unparsed | tool call; q_kind adds unparsed; q_channel (amendment 1): text | tool_note (the answer was in sign_guestbook's note at the question) | tool call (another tool call, unparsed). The gate visits are rescored under amendment 1 (summary.question_amendment1; the original stays in summary.question)
- visitor_family / family columns: claude | gpt | gemini | x-ai | mistralai | meta

## events.csv — one row per tool call a visitor made, in order (8117 rows)
- visit_id, step: 1-based order of the call within the visit; museum_version, arm: as in visits.csv
- tool: list_wings | enter_wing | view | buy_postcard | leave | sign_guestbook, answer_question (v3; museum_no filled for an answer naming a piece)
- wing: the wing entered (enter_wing) or the wing of the artwork (view, buy_postcard)
- museum_no, view_type (frame | code | statement; blank unless a valid view)
- position_in_list: 1-based place in what the visitor was shown. For enter_wing: the wing's place in the list_wings order (blank if list_wings had not been called). For view/buy_postcard: the artwork's place in its wing's list (blank if that wing had not been entered yet)
- grid_position (1–8, row-major), grid_row (1–2), grid_col (1–4): arm B only, for view/buy_postcard, the artwork's place on its wing's wall as shown (blank if that wall had not been shown yet). Same seeded order as position_in_list
- artwork_model: the artist; is_own / is_family: y/n against the visitor
- calls_left: the ticket after the call
- turn: the turn whose reply made the call (logged from v4 on; blank before)
- valid: n for a call that was invalid (unknown wing/artwork/view, postcard not viewed, looking call with no calls left, or a call after the visit had ended)

## Excluded runs (on disk in museum/visits/, not in the tables)
- 20261020-01-medium: first pilot, refused on turn 1 (no system prompt)
- 20261020-02-medium: first pilot, refused on turn 1 (no system prompt)
- 20261020-03-high: first pilot, refused on turn 1 (no system prompt)
- 20261020-04-high: first pilot, refused on turn 1 (no system prompt)
- 20261020-12-medium: Gemini-4 smoke visit that died on a gateway HTTP 504 at turn 5 (re-run as 20261020-13-medium)
- 20261022-41-medium-v2B: no summary.json (visit did not finish)
- 20261022-58-medium-v2B: no summary.json (visit did not finish)
- 20261025-01-medium-v3: no summary.json (visit did not finish)
- 20261025-07-medium-v3: no summary.json (visit did not finish)
- 20261025-11-medium-v3: no summary.json (visit did not finish)
- 20261025-62-medium-v3: no summary.json (visit did not finish)
- 20261026-133-medium-v4: no summary.json (visit did not finish)
- 20261026-159-medium-v4: no summary.json (visit did not finish)
- 20261026-170-medium-v4: no summary.json (visit did not finish)
- 20261027-53-medium-v4: no summary.json (visit did not finish)
- 20261027-64-medium-v4: no summary.json (visit did not finish)
- 20261027-88-medium-v4: no summary.json (visit did not finish)
- 20261027-93-high-v4: no summary.json (visit did not finish)
- 20261028-01-medium-v4: no summary.json (visit did not finish)
- 20261028-02-medium-v4: no summary.json (visit did not finish)
- 20261028-07-medium-v4: no summary.json (visit did not finish)
