# CityMETER · Hotel — data and method How `data/hotels.json` is built, what one row and one figure mean, and what the demo never claims. Every number below is a snapshot of release 0.3.10 (rebuilt 2026-11-08 from the same export of 2026-11-01 and the checked rows of the live `49_SITE_BOUNDARY_CHECK` tab: the figures of 0.3.1 unchanged, 24 stays under their brand's listed name, §2a; 463 stay sites, §7c); the page computes everything from the data at run time, so these are fixtures, not constants. ## วิธีนับ — อ่านตัวเลขในหน้าอย่างไร หน้าจอไม่มีคำเตือน (เจ้าของ 1 พ.ย. 2569: “ไม่มี caveats เลย”) ทุกบรรทัดที่มาท้ายหน้ามีลิงก์ “วิธีนับ” มาที่นี่ ขอบเขตของตัวเลขอยู่ตรงนี้ที่เดียว - **ห้อง** = จำนวนห้องที่แหล่งบันทึก (`rooms_selected`) รวมต่อพื้นที่ ไม่ใช่ห้องว่างต่อคืน อัตราการเข้าพัก หรือขนาดตลาด ที่พักที่ไม่ทราบจำนวนห้องไม่นับเป็น 0 ห้อง (แท่ง “ไม่ทราบ” ลายทแยง) - **แห่ง** = รายการที่พักหนึ่งรายการตามแหล่ง ยังไม่ได้ตรวจภาคสนามว่ายังเปิดอยู่ ชุดข้อมูลเป็น research master ไม่ใช่สำมะโนที่พักทั้งประเทศ - **สัดส่วนของจังหวัด** = ห้อง (หรือแห่ง) ในพื้นที่ ÷ ทั้งจังหวัดในตัวกรอง รวมที่พักที่แหล่งระบุได้แค่จังหวัด ซึ่งนับในยอดจังหวัดแต่ไม่อยู่ในสีของเขต/อำเภอใด - **ใกล้รถไฟฟ้า** = ห่างจุดอ้างอิงของสถานีที่เปิดแล้วไม่เกิน 500 ม. เป็นระยะตรง ไม่ใช่ระยะเดินหรือเวลาเดินทาง นับเฉพาะที่พักที่มีพิกัด (ตัวตั้งและตัวหารอยู่ในคำอธิบายของแต่ละแท่ง) - **ราคาที่บันทึก** = ต่ำสุด สูงสุด เฉลี่ยต่อคืนตามที่ CityMETER Hotel V3 บันทึก ไม่ทราบช่วงเวลา ไม่ใช่ราคาวันนี้ ADR หรือรายได้ แสดงรายโรงแรมเท่านั้น ตัวเลขจากคำตอบของ Google ที่ปุ่ม “ถาม Google AI” เปิดให้ไม่ถูกนำมาใส่ในหน้านี้ - **คู่เทียบ** = ที่พักประเภทเดียวกัน ห้องครึ่งเท่าถึงสองเท่า ในระยะตรง 2 กม. เรียงใกล้ก่อน เป็นรายชื่อเริ่มต้นสำหรับเทียบ ไม่ใช่ชุดคู่แข่งที่ยืนยันแล้ว - **เครือ** = จากแบรนด์ที่ชุดข้อมูลบันทึกพร้อมหลักฐาน ที่พักที่ไม่มีแบรนด์ไม่ได้แปลว่าเป็นโรงแรมอิสระ - **ชื่อที่บันทึก** = ชื่อที่พักตามแหล่ง (ทะเบียนที่พัก ททท. หรือ CityMETER Hotel V3) แสดงใต้ชื่อเมื่อที่พักเปลี่ยนแบรนด์หรือชื่อแล้ว ชื่อหัวข้อเป็นชื่อที่แบรนด์ใช้บนหน้าเว็บของแบรนด์ (ลิงก์ “หน้าแบรนด์” และวันที่ตรวจอยู่ใน “ที่มาและพิกัด”) ค้นด้วยชื่อใดก็เจอ ช่องค้นหาหาทุกคำที่พิมพ์ในชื่อ แบรนด์ ประเภท หรือพื้นที่ (เช่น “novotel karon”) - **อาคาร** = อาคารบนแผนที่ (CityMETER Building V4 หรือ OpenStreetMap ตามพื้นที่) ความสูง “สูงราว” เป็นค่าประมาณของชุดข้อมูล ไม่ใช่ผลรังวัด “ที่พักในอาคารนี้” = หมุดของที่พักอยู่ในขอบอาคาร - **อาคารใกล้หมุด** = เมื่อหมุดของที่พักไม่อยู่ในขอบอาคารใด มุมมองอาคาร 3 มิติเปิดที่อาคารบนแผนที่ที่ใกล้หมุดที่สุดภายใน 300 ม. (ไม่มีก็ขยายเป็น 800 ม.) โดยมีหมุดอยู่ในภาพด้วย เป็นการเลือกมุมกล้องเท่านั้น ไม่ได้ยืนยันว่าที่พักอยู่ในอาคารนั้น และไม่นับใน “ที่พักในอาคารนี้” ถ้าในระยะ 800 ม. ไม่มีอาคารบนแผนที่ ภาพแสดงพื้นที่รอบหมุด - **ทั้งรีสอร์ท N อาคาร** = อาคารบนแผนที่ที่อยู่ในขอบเขตไซต์ของที่พัก ทีมวิจัยวาดขอบเขตจากรูปแปลงที่ดินของกรมที่ดิน (LandsMaps) หรือกรอบพื้นที่ของที่พักใน OpenStreetMap ร่วมกับผังของที่พักและภาพถ่ายดาวเทียม (3 แห่งในกะรน ภูเก็ต จากรูปแปลง ตรวจ 3 พ.ย. 2569 · 12 แห่งจากกรอบ OpenStreetMap ที่เทียบผังแล้ว ตรวจ 6 พ.ย. 2569) หรือกรอบ OpenStreetMap ของที่พักที่ตรวจอัตโนมัติ ยังไม่ได้เทียบผัง (448 แห่ง ตรวจ 8 พ.ย. 2569: กรอบชื่อของที่พักเอง 267 แห่ง และกรอบที่ผ่านกฎชื่ออัตโนมัติ 181 แห่ง ทุกแห่งเทียบอัตโนมัติกับเว็บไซต์ แบรนด์ อาคาร และหมุดแล้ว หน้าเว็บอ่าน “กรอบ OpenStreetMap (ตรวจอัตโนมัติ)”) ข้อมูล OpenStreetMap © OpenStreetMap contributors ภายใต้ ODbL 1.0 กดที่พักแล้วแผนที่แสดงและจัดภาพทุกอาคารของไซต์พร้อมกัน ไซต์เป็นการจัดกลุ่มอาคารเพื่อดูบนแผนที่ ไม่ใช่ขอบเขตโฉนด เนื้อที่ หรือการยืนยันกรรมสิทธิ์ และไม่เผยแพร่รูปแปลง เลขโฉนด หรือกรอบพื้นที่ - **ภูมิประเทศ** = พื้นดินรอบหมุดของที่พัก (รัศมี 500 ม. ต้นไม้ 300 ม.) จาก SRTM 30 ม. (สำรวจปี 2543) และแผนที่ OSM ไม่ใช่การยืนยันวิว ราคา หรือความเสี่ยง - **สถานีที่กำลังสร้าง** = โครงการ ยังไม่ใช่บริการ ## 1. Sources | Input | What it is | Published here? | |---|---|---| | Research master | Google Sheets “CityMETER Hotel — Venue, Branch & Building Research Master 2026-10-24” (registry id `research.TH.hotel`, internal). Tabs read: `01_VENUES`, `13_BUILDING_LINKS`, `16_BRANCH_SPATIAL_CANDIDATES`, `17_BRANDS`, `19_AFFILIATIONS`, `23_TYPE_ASSIGNMENTS`, `37_IDENTITY_ALIASES`, and (0.2.0) `47_ROOM_RESOLUTION` (dates of the room re-checks written back to `01_VENUES`) and `48_LOGO_LIBRARY` (brand identities and logo hashes) | No. Only the SHA-256 of the .xlsx export (`meta.exportSha256`) and the derived columns in `data/hotels.json` | | CityMETER production base info | `https://landometer.com/api/v2/citymeter-v3/base/info/lv0` and `…/lv1/{uuid}` (77 provinces, 928 districts, production codes, Thai names, areaSqm), fetched by `scripts/fetch-base.py` | Receipts only (`data/boundary-receipts.json`: url, status, bytes, sha256, retrievedAt for 78 responses) | | District outlines | Display geometry per province (`data/frames/{CODE}.json`) and province frames (`data/provinces.geojson`), built by `scripts/build-boundaries.py` | Yes | | CityMETER · Transit | `montri-th/transit` `data/transit.json`, commit and SHA-256 in `meta.context.rail` (release `transit-data-2026-10-31-r1`, as of 2026-10-31): 195 operating and 79 under-construction stations referenced | Station id, name, line labels, status and reference point of the referenced stations | | CityMETER Shopping centers | `montri-th/shoppingcenter` `data/venues.json`, commit and SHA-256 in `meta.context.retail` (as of 2026-10-26): 584 venues referenced | Venue id, names, group and reference point of the referenced venues | The rows themselves come from the sources the research master lists (`sources[]`): CityMETER Hotel V3 (the hotel dataset of CityMETER production, 28,381 rows), the TAT accommodation register on the TAT Data Catalog (2,040 rows), official group and property pages, and provincial lists published on data.go.th. ## 2. One row, two measures **One row = one row of `01_VENUES`**: a source-listed accommodation record, as the sheet defines it — not a field-verified census of operating hotels (`operation_status = operation_not_revalidated` for 30,525 rows, `official_listing_not_field_verified` for 205). Nothing is merged, split or moved. The alias row of `37_IDENTITY_ALIASES` (an earlier record of the same current property, `accepted_same_current_post2024_property`) is kept out of every figure (`meta.aliases`), as the sheet's own rule says. **30,729 rows** remain. | Measure | Definition | Words on screen | Colour family | |---|---|---|---| | `rooms` (default) | Σ `rooms_selected` — the room count the sheet selected from its sources. Recorded stock, not rooms available per night, occupancy or demand | “ห้อง” | LI `li.supply` (capacity) | | `count` | accommodation records | “แห่ง” | LDS `count` | A record with unknown rooms (blank `rooms_selected`, **79 rows** after the research team's room re-check of 1 Nov 2026 — `47_ROOM_RESOLUTION`, 258 tasks, 178 written back; raw zeros the sheet left unselected stay unknown) adds nothing to a room total and is counted apart (“ไม่ทราบจำนวนห้อง”) — never as 0 rooms. A district whose records all have unknown rooms is `no_data` under the rooms measure (hatch), never a measured 0. Known rooms: **1,033,157 rooms in 30,650 records**. No ratio to area or population is drawn. Recorded rooms per km² or per person would read as a density or a market measure the data does not support. **Recorded prices** (`min/max/avg_price_raw_thb`, 28,381 CityMETER Hotel V3 rows, `price_period = unknown`) are shown per hotel as “ราคาที่บันทึก” (the unknown period is the “วันที่ราคา · ไม่ระบุในแหล่ง” field of the hotel's source details and a line of “วิธีนับ”), in the table and in the comparison set — never summed, averaged across hotels or called a rate, an ADR or revenue (owner decision 2026-11-01: publish per hotel, including the recorded prices). Prices are never taken from a search engine's AI answer (owner question 2026-11-02): it states no stay date, room, occupancy or tax basis, gives different figures from one ask to the next (Sri Panwa the same morning: 8,999–20,382 บาท per night in an AI Overview, 19,000–24,000 for its cheapest zone and a THB 17,159 map pin in AI Mode) and its sources are third-party pages. Dated prices come only from `46_PRICE_OBSERVATIONS` of the research sheet — each quote with its publisher, stay dates, room, occupancy, tax basis, status, source URL and capture hash (59 quotes for 13 hotels, checked 2026-10-25, not yet published) — rebuilt like every other field. Star ratings are not published (64 rows only, `source_recorded_not_certification_verified`; a star would read as a rating). ## 2a. Names: recorded and current (0.3.4) Owner question 2026-11-02: “ทำไม novotel กะรน ใช้ logo Holiday Inn?”. The research sheet links a record to the brand that operates the property now, with evidence, and keeps the record's own name: the TAT register still names the asset at 568 Patak as Novotel Phuket Karon Beach Resort and Spa, IHG lists it as Holiday Inn Resort Phuket Karon Beach (`19_AFFILIATIONS`, retrieved 2026-10-24), so the panel showed a Novotel name beside a Holiday Inn Resort logo. Rule (`scripts/build-data.py#current_name`, judged after every row so each brand has all its labels): a branded record whose Latin name (`name_en`, else a Latin `venue_name`) carries none of its brand's labels, and which that brand's accepted listing (a listing role; status official, resolved or membership; `brand_id` of the linked brand, or the same group when the listing has no brand id) names differently, is shown under the listed name (`data.currentNames`: name, link, retrieved on). A listing in capitals only reads in title case. The record's own names stay in the row: under the heading and in the hover card, lists and the popup as “ชื่อที่บันทึก”, in “ที่มาและพิกัด” beside the listing link, and in search. A Thai-only record name (no Latin to compare) or a brand that lists the property under several names keeps the record's name. Nothing else of the row changes: rooms, prices, district, point and brand are the record's. 24 rows at this build: 13 renamed or rebranded properties (Novotel Phuket Karon → Holiday Inn Resort Phuket Karon Beach, Novotel Bangkok Fenix Silom → Hilton Garden Inn Bangkok Silom, Pullman Bangkok Grande Sukhumvit → Hilton Bangkok Grande Asoke, Le Meridien Suvarnabhumi → Hilton Bangkok Suvarnabhumi Golf Resort & Spa, Majestic Grande → DoubleTree by Hilton Bangkok Ploenchit, The Tawana → voco Bangkok Surawong, Windsor Suites → Fairmont Bangkok Sukhumvit, Imperial Adamas → Phuket Marriott Resort and Spa, Nai Yang Beach, Wiang Indra → Kahavadi Chiang Rai, Hyatt Place Phuket Patong → Andaman Beach Hotel Phuket, Hilton Garden Inn Phuket Bangtao → Mamaka Phuket, Journey Hub → Mercure Phuket Patong Journeyhub, Tana Boutique → Mercure Samui Chaweng Tana) and 11 whose record spells the name differently, without the brand's words or with a broken character (Hop in → HOP INN Sa Kaeo, Emmiline → Emmaline Hotel Nan, Kimton → Kimpton Kitalay Samui, Le M?ridien → Le Méridien Chiang Rai Resort, Sheraton Grande Sukhumvit → … a Luxury Collection Hotel, Bangkok). Source names that carry HTML entities (`&`, 240 rows) read as the character on screen. ## 3. District membership (build time, never re-assigned on screen) The first method that applies (`row.admin`, `meta.adminMethods`): | Method | Rule | Rows | |---|---|---| | `source` | CityMETER Hotel V3 rows: the production DISTRICT the source API listed the row under (`…/hotel/info/beeSwarm/DISTRICT/{uuid}` → production code) | 28,380 | | `name` | the row's own district name is a production district of its province | 2,117 | | `point` | the row's selected point lies in a production district of its province (display outline) | 197 | | `province` | province only — in every total, in no district colour | 35 | | `conflict` | the point lies in another province than the row's own: placed nowhere, reported | 0 | A source district that disagrees with the row's point is kept (the production listing is the membership) and reported for research (`meta.checks.sourcePointDisagree` = 7, ids in `sourcePointDisagreeSample`). A source district in another province than the row's own province is not used: the row falls through to the name or point method of its own province (`meta.counts["source.provinceDisagree"]` = 1). Two district-name spellings that do not match production fall through to the point method. **860 of 928** districts hold at least one record; the other 68 are measured zeros of this dataset (EVID-05: every district is in the register's national scope). ## 4. Points The sheet's selected point (`latitude`/`longitude`) with its role (`coordinate_status` → `cols.coord`, four values in `vocab.coord`): 29,149 rows. Points outside Thailand's box are dropped (none in this export). Rows the sheet holds or quarantines (`held_identity_location_conflict`, `known_wrong_selected_point_quarantined_owner_point_pending`) carry no selected point and are not drawn. A row without a point is still counted in its district. ## 5. Context from the same registry (straight lines, build time) Distances are straight lines between the hotel's selected point and the other dataset's reference point — never a walking route, a travel time or a catchment. - **Rail** (`cols.rail`, `railM`): the nearest operating station of CityMETER · Transit within 2,000 m (1,472 rows); `rail2`/`rail2M`: the nearest station under construction within 2,000 m (973 rows; a project, not a service). Operating = on the operator roster of CityMETER · Transit (the Bangkok metropolitan rail network in this release). - **Retail** (`cols.shop`, `shopM`): the nearest countable, located CityMETER Shopping centers venue of the groups mall · community · outlet · component · hoststore with status `supply`, within 1,500 m (7,108 rows). The panel links to the same venue in CityMETER Shopping centers by its id. - **Same building (research)** (`inBuilding`): rows of `16_BRANCH_SPATIAL_CANDIDATES` qualified `spatial_candidate_requires_identity_review` — a branch of another CityMETER research dataset (restaurants, grocery, pharmacy, …) whose source point falls in the same research building mass as the hotel's point (93 hotels). Published as industry and public brand only; it is not a tenancy, floor or located-in claim. - **Meeting rooms** (`cols.meet`): the meeting-room count of the TAT accommodation register (2,115 rows). - **Groups** (`brands[]`, `chains[]`): the brand name the sheet selected with evidence (468 rows), joined to the brand register `17_BRANDS` and its group by id. 20 selected brand names (34 rows) have no group in the register yet; they read as “ไม่ระบุเครือ” and are listed for research (§9). A record without a brand is not proof of an independent hotel. ## 6. What the page computes - **First Aha** (`js/area-insight.js`): the leading province = the province with the most of the measure among rows placed at district level; up to three of its districts with the most of it, never cutting through a tie and never every district that holds the measure (a province with N ≤ 3 such districts shows its top N − 1); share = Σ focus ÷ the province's whole total in the selection (`homeValue`: placed rows and rows placed only at province level; since 0.3.1, so the sentence reads “35% ของห้องพักกรุงเทพมหานคร” without a qualifier). The district panel's share of its province uses the same denominator. Rail comparison: share of the measure within 500 m of an operating station among the focus rows with a point vs the rest of the province's placed rows with a point. Outside finding: the district(s) with the most of the measure outside that province. Snapshot (rooms): Bangkok · คลองเตย 20,679 · วัฒนา 17,774 · ราชเทวี 15,063 = 53,516 / 153,200 (35%; placed total 152,864 + 336 rooms placed only at province level); rail 67% (25,768/38,690) vs 42% (32,732/77,957); outside: บางละมุง ชลบุรี 62,561. Snapshot (records): Phuket · เมืองภูเก็ต 1,024 · กะทู้ 862 = 1,886 / 2,223 (85%). - **Station panel** (`buildStationProfile`): rows with a point within 500 m of the station's reference point (a row near two stations is listed at both). **Station ranking** (`stationsOf`): each row counted once, at its nearest operating station. - **Comparison set** (`compsetOf`): rows with a point within 2,000 m, of the same accommodation group, whose known rooms are within ½–2× the hotel's own (both known), nearest first, up to 10; plus every record within 2,000 m of any group. Reads every row of the dataset, so it does not move with the filters. A screening aid, not an STR competitive set. ## 7. Colour and categorical rules - Rooms → Location Intelligence Profile for LDS 0.9.7 `li.supply` (scaleVersion `671112f3…c497`); records → LDS 0.9.7 `count` (`4ff5c1fb…0980`). All 41 samples by rank among the measured nonzero values being read (owner standing rule), the same HEX on both themes; EVID-05 treatments from the base `dataState`. `npm run check:lds -- --lds --li ` compares every shipped value with both source files (124 checks). - Accommodation groups bind by stable id to `landometer-series-10-v8` with the series shape cue (`js/cues.js#GROUP_SERIES`): hotel 2 · resort 1 · apartment 3 · guesthouse 4 · bungalow 5 · homestay 10 · other 9 (no series 6 / star). - Symbol size (0.2.0, owner 2026-11-01): the symbol's area grows with the recorded rooms — radius = k(zoom) · √(rooms ÷ 100), rooms capped at 1,500, never below 1.6 px (`js/model.js#pinRadius`, `PIN_SCALE`); the colour stays the group's. Unknown rooms: the same shape as an open outline at the 50-room size (EVID-05 no_data — never a small filled mark). Quiet context dots: grey, about half size, at most 6 px of radius. ## 7a. Names, logos and terrain on the map (0.2.0) - **Which names show** (`js/labels.js`): from zoom 12.5, the focus rows whose point lies in the central ellipse of the visible map (36 % of its half-width and half-height), ranked by an importance score = the mean of the record's national percentile of recorded rooms and of recorded price (avg, else max, else min). A record without a recorded price ranks on its rooms alone and one without rooms on its price alone; the price is never guessed and the score is never shown. At most 9 labels on desktop and 5 on a phone, measured and placed so they never cover the map chrome, the district chips, the popup or each other. - **Label text**: a branded stay reads as two lines of text — the brand in bold, then the site: its name (English when there is one) without the brand words, a leading type word (โรงแรม …), the province name (Thai or English, กรุงเทพฯ) and generic endings (รีสอร์ท แอนด์ สปา, Hotel …); other stays read as one such short name. Never shorter than 3 characters; the full name is the accessible name and heads the card. Logos are not drawn in labels: at label size they cannot be read (owner 2026-11-01). - **Chain logos** (`data/brand-logos.json`, `assets/brand-logos/`, `scripts/build-logos.py`): the research team's chain-brand logo library (`Hotel_Chain_Logos_Web`, checked 2026-11-01) joined by `48_LOGO_LIBRARY`; 107 of 112 brands of this release. Every file's SHA-256 matches the library manifest and the tab; SVGs are checked for scripts, handlers and external references; rasters above 512 px or 64 KB are scaled uniformly to 384 px (WebP) after a transparent margin is trimmed — never recoloured, cropped into the artwork, traced or upscaled. Each logo sits on the library's recommended surface (theme-invariant backing), at a readable size in the hover card (≈ 40 px tall) and the hotel panel (≈ 44 px). Rights stay with each owner; showing a mark to identify a brand is not an endorsement or a licence. - **Terrain mode** (`js/shell/terrain.js`): at site scale (zoom ≥ 12.5, 3D view), when the visible ground spans at least 30 m (p95 − p5 of a 14 × 14 sample of the DEM at zoom 11; off below 20 m or zoom 12.2), the map shows 3D ground exaggerated × 1.5 (said in the legend) and a hillshade in basemap tones; the district fill switches off from zoom 12.5 (a hard switch). Since 0.3.0 (owner 2026-11-01: “อยากให้เห็นอาคารบนภูมิประเทศ”) the building or stay being looked at keeps terrain too: its site decides — p95 − p5 of a 12 × 12 sample at zoom 11 over ±1 km ≥ 30 m (Sri Panwa 31 m, above Patong 154 m; Bangkok's towers read 16–20 m) — and the building camera, the x-ray and the pins stand on the drawn ground. Terrain is switched on or off only while the camera is still (MapLibre 5.6.2 stops rendering when a terrain is set during a camera move). DEM: Terrain Tiles on AWS (Mapzen/Tilezen terrarium; SRTM 30 m, acquired 2000, over Thailand). - **Terrain facts of a stay** (card and panel): ground elevation at the point (DEM zoom 13); relief = max − min of the point and 24 points on rings of 250 m and 500 m; higher/lower than its surroundings when the point differs from the mean of the 500 m ring by ≥ 15 m; the nearest of the rings 0 · 150 · 300 · 600 · 1,200 m (straight line) that reaches an OSM ocean polygon of the loaded basemap tiles; tree cover = share of a 75 m grid within 300 m under OSM `wood` land cover, shown only when above 0 (absent land cover is not 0 % trees). Facts of the ground only — never a view, a price or a risk; unknown is left out. ## 7b. Buildings per area (0.3.0) Owner decision 2026-11-01 (point 3): “ใช้เกณฑ์ 15% … ได้”, “กรุงเทพฯ ใช้เขตเป็นหน่วย … ได้”, with the question first answered whether CityMETER keeps OpenStreetMap's multi-part (LOD2+) buildings. - **Unit**: every อปท. (CityMETER production `municipalityInfos`) that holds at least one stay with a point, by its production outline; in Bangkok the เขต (production `districtInfos`, outlines = the page's district frames). The production building tiles clip Bangkok by แขวง, so a เขต is read through its แขวง. - **Rule**: CityMETER Building V4 where its building count in the unit (production `building/info/lv1/{province}?version=V4`) is at least 1.15 × the OpenStreetMap count of the same unit; otherwise OpenStreetMap. OSM count = ways tagged `building=*` in the unit polygon (Overpass API `out count`, outer rings minus holes, polygons simplified to ≤ 200 vertices with a closed-ring Douglas–Peucker). The count in the unit's bounding box is an upper bound: where 1.15 × it is still ≤ the CityMETER count, or it is 0, it decides without a polygon count. A unit whose OSM count could not be read stays on OSM. Units where both sources hold fewer than 10 buildings are flagged `gap`. - **Result** (counts of 2026-11-01, `data/building-units/index.json`): 3,947 units hold a stay with a point (3,897 อปท. and the 50 เขต). CityMETER Building V4 is drawn in 3,900 of them, OpenStreetMap in 47 (stays with a point: 27,816 in CityMETER units, 1,206 in OSM units). Bangkok: 46 of 50 เขต on CityMETER; บางรัก, ป้อมปราบศัตรูพ่าย, สัมพันธวงศ์ and ยานนาวา stay on OSM. The OSM units are mostly old towns and islands OSM maps closely (เชียงคาน, เกาะช้าง, เกาะเต่า, ศาลาด่าน, ปาย, พระนครศรีอยุธยา). OSM counts: 3,330 units decided by the bounding-box upper bound, 437 with no OSM building in the box, 180 by polygon count; 28 units were counted on an Overpass mirror holding OSM of 2026-06-01, the rest on OSM of 2026-11-01 (`meta.osm.osmBaseDays`). 3 units hold fewer than 10 buildings in both sources (`gap`). Reported, not patched: a few CityMETER counts look low for a town (เทศบาลตำบลเชียงคาน 98, เทศบาลเมืองนครพนม 72). - **On the map** (`js/shell/building-units.js`, `js/massing.js`): inside a CityMETER unit the focus tier draws the unit's CityMETER Building V4 buildings — production tiles `building/detailMvt/14/{x}/{y}` read per อปท. (Bangkok: per แขวง), every building drawn once from the tile that holds its interior point — with the same rounded, slightly enlarged shape as OSM buildings; height = `height_m`, the dataset's model estimate (3 m when missing). OSM buildings whose interior point falls inside the unit are left out of the focus and hidden from the quiet city tier (a hidden OpenFreeMap feature's buildings outside the focus are redrawn as quiet context). Context outside the focus stays OSM. A CityMETER tile that fails or exceeds the request bound (40 per focus) leaves its part of the unit on OSM. - **Parts override (LOD2)**: inside a CityMETER unit, buildings OSM maps in parts (`building:part`; OpenMapTiles hides their outline in 3D, `hide_3d`) stay OSM, and the CityMETER buildings whose interior point stands in such an outline are left out. Basis (Observed fact, OSM at 2026-11-01 13:47 UTC and taginfo Geofabrik to 2026-10-31): OSM 1.77 million buildings in Thailand vs CityMETER V4 29.86 million; 2,175 `building:part` pieces; 2,236 buildings with a roof shape, 92 % of them with a CityMETER footprint (median height difference 0.5 m); 71 % of the parts lie inside a CityMETER footprint and 23 % more within 20 m; 231 multi-level buildings (141 with ≥ 5 m between levels, mostly central Bangkok) would lose their levels as CityMETER prisms. MapLibre fill-extrusion draws no roof shapes, so the levels are what the override keeps. - **Wording**: the attribution names “อาคาร: CityMETER Building V4” while any of its buildings is drawn; the building panel says “สูงราว” with the source in its title and in the provenance line (“หมุดในขอบอาคาร CityMETER Building V4”); the join of stays to buildings (a pin inside the outline: “ที่พักในอาคารนี้”) is the same for both sources (keys `osm:` / `cm:`). ## 7c. Stay sites (0.3.6–0.3.10) Owner request 2026-11-03: “ไม่ต้องทำเครื่องมือเลือกอาคาร แต่พี่เลือกโฉนดให้เลย มาสร้างขอบเขตอาคาร ร่วมกับ site layout ของ โรงแรม/รีสอท ทดลองกับสามแห่งนี้ก่อน”, after “ลองใช้พิกัดโรงแรม/รีสอท ไปค้นจาก landsmaps.dol.go.th … เพื่อไปดูว่ามีโฉนดใดบ้างเป็น ของโรงแรม รีสอทนั้น แล้วนำกลับมาจัดกลุ่มอาคารรวมเป็นพื้นที่รีสอทเดียวกัน” (with assessprice.treasury.go.th). - **Sources of the boundary (research, in the sheet)**: Department of Lands parcel outlines as LandsMaps draws them after a deed search (its parcel map layer, read from the display), the Treasury Department's appraisal map (D-Value) for the deed numbers it identifies at a point, the stay's own site layout and Esri World Imagery. LandsMaps gives no owner and hides part of a deed number; D-Value gives no outline and identified no parcel under the Phuket Orchid buildings or the Holiday Inn main block. Parcels were chosen where the layout's buildings stand; a building the layout shows outside the chosen parcels was added by its own outline. Result (`49_SITE_BOUNDARY_CHECK`, checked 2026-11-03, `boundary_source = dol_parcels+site_layout`): Holiday Inn Resort Phuket Karon Beach — 2 deeds and the main-block parcel, plus the long building on the east side, site ≈ 18,300 m² (layout: a guest photo of the resort's printed map on Tripadvisor); The Old Phuket Karon Beach Resort — 1 deed, site ≈ 13,200 m² (the stay's own site: Sino and Serene wings); Phuket Orchid Resort & Spa — 3 parcels, site ≈ 30,400 m² (Kata Group's resort map on a travel agency page). A site area is the area of the drawn polygon, never a title area. - **Published** (`data.siteGroups`, `meta.siteGroups`): per checked stay the interior points of the drawn buildings inside the boundary (`buildings_in_boundary`; CityMETER Building V4 in Karon), the count, the check date, the source code and the evidence page — 60 buildings: Holiday Inn 20, The Old Phuket 19, Phuket Orchid 21. The boundary itself, deed numbers, parcel data, the notes and the checker stay in the sheet (`meta.excludedFields`). A site building more than 600 m from the stay's point stops the build. The live tab is read through `--sites` (its header and checked rows, SHA-256 in `meta.siteGroups.file`) while the export predates it. - **On the map** (`js/shell/building-join.js#buildingsAtPoints`, `js/app.js#focusTarget`): opening the stay highlights, x-rays and frames (two-point camera, the stay's pin in the frame) every drawn building whose outline holds one of the points, every piece of each; a point no drawn outline holds is left out, never snapped to a neighbour. The massing square round the pin grows to hold the site (`siteReach`). The panel reads “ทั้งรีสอร์ท N อาคาร” (N = the sheet's count), the source details a “ไซต์” row (count · source · หน้าอ้างอิง · date) and the provenance line “ไซต์ รูปแปลงกรมที่ดิน + ผังที่พัก ”. - **Next sites** (owner 2026-11-03 “กรรมสิทธิ์ไม่ต้อง” and “ทำไมไม่ใช้โฉนด จากกรมที่ดินและธนารักษ์”): ownership is not proven, but the parcels still bound the site — deeds read on D-Value and LandsMaps, matched with the stay's layout (`dol_parcels+site_layout`); a site rests on the layout and imagery alone (`site_layout`, provenance “ไซต์ ผังที่พัก”) only where no parcel can be read and no OpenStreetMap area of the stay is found (next item). `scripts/site-candidates.mjs` lists the CityMETER Building V4 buildings the map draws round a stay (the map's own code, so every point it writes stands in a drawn building), turns the chosen buildings into the tab's cells (`buildings_in_boundary`, a 6 m envelope or the researcher's polygon as `boundary_geojson`) and checks a row's points (`check`). - **Third source — an OpenStreetMap area of the stay** (owner 2026-11-06 20:39 “approve”, to the question “เพิ่ม `osm_polygon+site_layout` เป็นแหล่งที่ 3 ของ `49_SITE_BOUNDARY_CHECK` หรือไม่”): the site follows a `tourism`, `leisure=resort` or named `landuse` way or relation of OpenStreetMap that holds the stay's pin, carries the stay's own name and holds no other stay's pin (skill `osm-area-site-join`, class `own`, within the size band of the dataset); the buildings are the drawn buildings whose interior point stands in the area (the rule of `site-candidates.mjs row --site`). The area is checked against the stay's own layout and imagery as for the other two sources — the layout page is `boundary_evidence_url`, the OSM element and the date of the OSM data read go in the notes; an area with no layout to check it is skipped as `no_layout`; a building cut by the area's edge is decided with the layout; where parcels are read for the stay they win over the area. A pin inside an area is not proof that the area is the stay's: the name and the layout must agree, and a neighbour's building put in is worse than one building left out. Provenance “ไซต์ กรอบ OpenStreetMap + ผังที่พัก ”. OpenStreetMap data is © OpenStreetMap contributors under ODbL 1.0: the points of such a site are derived from it (`meta.siteGroups.credits`; the “ไซต์” row links “© OpenStreetMap contributors” to https://www.openstreetmap.org/copyright and the map attribution names OpenStreetMap); the area itself stays in the sheet like any boundary. The first national read (Overpass API, 6 Nov 2026, the eight provinces read first; internal drafts in the CityMETER Project docs) found an area of the stay's own for 148 of the 7,702 stays read (1.9%), 130 of them holding two or more drawn buildings — most sites still need parcels and a layout. - **Result 0.3.7** (`49_SITE_BOUNDARY_CHECK`, checked 2026-11-06, `boundary_source = osm_polygon+site_layout`): eleven of the tab's stays whose own area was found (one OSM way each, OSM data of 2026-11-06 read from the main Overpass API server; element and date in the notes), each compared with its layout page on Esri World Imagery; a building the layout does not show as the stay's, or that may be a neighbour's, was left out and named in the notes. Pullman Phuket Karon Beach Resort 26 buildings · Mida Grande Resort Phuket 63 · ANA ANAN Resort & Villas Pattaya 6 · The Imperial Phukaew Hill Resort 26 · Sea Two Pool Villa 33 · Silavadee Pool Spa Resort 77 · Banyan Tree Krabi 33 · Navana Nature Escape 11 · The Gems Mining Pool Villas Pattaya 32 · Sea Sand Sun Resort and Villas 33 · Amari Vogue Krabi 14 — 354 buildings (with the three Karon parcel sites, 14 sites and 414 buildings). The evidence pages are the stays' own or their booking, travel or guest pages that show the layout (`boundary_evidence_url`). Frames that missed part of the stay on its layout (for example the wings of one Phuket resort outside its OSM area) are held, not written. Fair House Villas & Spa (Ko Samui, 14 buildings; resort map on a guest page) was checked the same way after the release and is published from 0.3.8. - **Fourth source — the same area checked automatically** (owner 2026-11-08 08:17, choosing “อัตโนมัติก่อน แล้วค่อยเทียบผังแห่งสำคัญ” over a layout check of every stay, 5–10 minutes each; and 07:45 “เพิ่มเลย”, to appending the stays the tab did not list): `boundary_source = osm_polygon_auto`. The OpenStreetMap area of the stay's own (class `own` as above) is read from the Geofabrik extract of Thailand (`thailand-261106.osm.pbf`, OSM data to 2026-11-06 21:21 UTC; the Overpass servers did not answer on 7–8 Nov) with the same filters, and every stay of the data set with a point is matched in one pass. The buildings are the drawn buildings whose interior point stands in the area, then by rule: a building that holds an OpenStreetMap lodging element named for another place (a hotel, guest house, hostel, apartment or resort node or area whose name is not the stay's; a villa or chalet of the stay is not cut), a place of worship, police, fire station, town hall, bus station, prison, school, kindergarten, college, university or hospital named for another place, a fuel point in a building under 1,000 m², an OSM building tagged religious, school or hospital, or an interior point in housing land named for another place or in education or religious land lying inside the area is cut; a building whose footprint lies mostly outside the area is cut and one mostly inside is added; the building the pin stands in is never cut; fewer than two buildings left is no site. No layout or imagery is compared: the evidence page is the OSM element (`https://www.openstreetmap.org/way|relation/`), `checked_by` is “Claude (ตรวจอัตโนมัติ)”, and the notes record the element, the OSM date, what was cut and why. A layout check later turns such a row into `osm_polygon+site_layout`; parcels still win. Provenance “ไซต์ กรอบ OpenStreetMap (ตรวจอัตโนมัติ) ”, the same ODbL credit. Stays whose frame a layout check had already found wrong or incomplete are held, not written by rule. - **Result 0.3.8** (checked 2026-11-08, `osm_polygon_auto`): 345 stays had an area of their own; 280 held two or more drawn buildings (27 one building, 18 none, 19 where the map draws OpenStreetMap buildings, 1 area larger than a site); the rule cut 1 building as another business (a guest house inside a hotel's area), 9 mostly outside the area and added 11 mostly inside; 11 of the 280 were already checked with their layout and 2 were held (Diamond Cliff Resort and Spa: its resort map shows two wings outside the area; Duangjitt Resort & Spa: its parcels await the owner) — 267 stays in 50 provinces, 3,722 buildings (Phuket 53 stays, Chon Buri 39, Surat Thani 37, Krabi 14, Chiang Mai 13; e.g. Phuket Graceland Resort and Spa 23, Rayavadee 55, Patong Bay Hill Resort 148). Eleven rows of the tab were filled in place and 256 appended (rank blank, `multi_building_signal = osm_own_frame`); the tab was exported and every written cell read back unchanged, and every checked row's points stand in drawn buildings (`site-candidates.mjs check`). With the parcel and layout-checked sites: 282 sites, 4,150 buildings. - **Review rules (0.3.10)** (owner 2026-11-08 09:45 “เทียบอัตโนมัติ และ พี่ทำกฎอัตโนมัติเลย / อันไหนพี่มั่นใจน้อยค่อยทำรายการมาให้ผมดู มารีวิวหลังจากที่ขึ้นเวบเสร็จแล้ว”): the areas the match left for a person (class `review`, 757 stays) are decided by rule, in order. An area outside the size band, or carrying a part word the stay lacks (phase, condo, wing, tower …): no. Not a lodging area (tourism hotel, motel, guest house, hostel, apartment, chalet, resort or camp site; leisure resort or beach resort): no — a residential area only when its name carries a lodging word and the names score high. Another stay's pin inside: no, unless every such pin is a record of the same place (the same name, or names that carry each other's distinctive words), and then each record takes the area. The pin a few metres outside: yes when no stay's pin is inside and the names match. No name: yes (low) for a lodging area that holds only this pin. Otherwise the names must match. Names match on their distinctive words in either script: a Thai word is romanised (RTGS) and compared with Latin words by sound class (ไดอาน่า = Diana, เดอะราชา = The Racha); the generic words of hotel names and place names (provinces, districts, subdistricts, towns, beaches and islands — Patong, หัวหิน, เกาะหมาก) do not count; a short word must be spelt alike, a part of a Thai compound counts only when it covers most of the word, and each name must find most of the other's words. Different numbers (Amonruk Hotel 2 / Amonruk Hotel) mean another branch; an area that stays of different places take goes to the one whose pin is inside, or to none. Claude read the pairs the first version accepted (323) and tightened the rules where a pair named two places (อาลาวี่ เฮ้าส์ / Navy House 33, Swan Beach Resort / สวอนเลค, Royal Grand Pattaya / Royal Green Park, Yao Yai Resort / Santhiya Koh Yao Yai Resort, an area named “Motel”); stays a layout round had already doubted stay held (Avani Ao Nang Cliff: the area's other name; Cross Pattaya Oceanphere: an estate; Veranda Chiang Mai: another stay's pin inside). The buildings are cut by the rules of the fourth source and the rows written as `osm_polygon_auto`, the rule and its evidence in the notes. - **Automatic comparison (0.3.10)** of every `osm_polygon_auto` site with evidence its frame did not come from: the area's website (`website`, `contact:website`, `url`; booking, listing and social sites ignored) and `brand`/`operator` tags against the stay's own link and brand (the same: up; different: down); buildings mostly across the frame's edge (a third of the site: down); floor area per recorded room (footprint × floors of 3.3 m; under 12 m² or over 1,500 m²: down); the pin more than 60 m from every site building (down). The tier starts from the name evidence (the stay's own name: high; a review rule: its tier), goes one step up on agreement and one down per disagreement; below low, or three disagreements, holds the site. Buildings just outside the frame are recorded, not scored: frames checked with their layout have as many such neighbours as one that missed two wings. The 267 sites of 0.3.8: high 257, medium 10, none held. The tier and its reasons are in the notes; the page shows the source as before. - **Result 0.3.10** (checked 2026-11-08, `osm_polygon_auto`): the rules accepted 243 of the 757 review stays (high 111, medium 90, low 42); 185 areas held two or more drawn buildings (21 one, 17 none, 20 where the map draws OpenStreetMap buildings or the area is larger than a site); the cut removed 1 building as another business and 7 mostly outside and added 5; Ibis Hua Hin was left with one building; the comparison held 3 (Anantara Chiang Mai, Venice Krabi Villa Resort, Apple A Day Resort: floor area per recorded room) and graded the rest high 83, medium 56, low 42 — 181 stays, 2,821 buildings (e.g. Diana Garden Resort 2, Ratilanna Riverside Spa Resort 6, Royal Twins Palace 5 for each of its two records). Six rows of the tab were filled in place and 175 appended; the tab's rows 1–732 were copied from the sheet and compared cell by cell with what was written and with the earlier export (16,836 cells, no difference). All sites: 463, 6,971 buildings (6,558 distinct points: two records of one place carry the same site). The less certain sites go to the owner as a list after the release. ## 8. Not published Addresses, sub-districts, phone numbers, raw coordinates, operator names, source record ids, researcher notes, quality flags, review statuses, star fields, building source ids, branch names of other datasets, the sheet's README text, the internal sheet or its export, private Drive/Sheet URLs, and of `49_SITE_BOUNDARY_CHECK` the site boundary, deed numbers, parcel data, notes and checker (`meta.excludedFields`). Source links are public https pages; CityMETER Hotel V3 rows link to CityMETER production (`?d=hotel`), never to its API. ## 9. Research follow-ups (reported, not patched in `data/`) - 7 rows whose selected point lies in another district than the one CityMETER Hotel V3 lists (`meta.checks`), and 1 row whose listed district lies in another province than the row's own. - 14 brands with no group in `17_BRANDS` (Design Hotels, Banyan Tree, Homm, Garrya, Kempinski, Eastin Grand, Prince Palace, LiT Bangkok, Cassia, YOTEL, GO Hotel, Kacha, Angsana, Meliá) — link them in the sheet, then rebuild. (0.2.0: six brand names that differed from `17_BRANDS` — Curio, DoubleTree, Four Points, MGallery, Renaissance, W — now join through the logo library identities of `48_LOGO_LIBRARY`.) Five brands have no accepted logo (Novotel Living, Eastin Grand: not found; Prince Palace, LiT Bangkok, Kacha: no library identity). - 267 rows with `star6_semantics_unresolved`; 79 rows with unknown rooms after the 47_ROOM_RESOLUTION re-check (80 tasks held, `selection_status = hold_*`). - `46_PRICE_OBSERVATIONS` (dated price quotes) is not used in this release; it is the path to seasonality. - 642 research building masses (Drive bundle `hotel-building-mass.geojson`) are not drawn in this release; the map draws OpenStreetMap or CityMETER Building V4 massing per area (§7b). The research team checks the site boundaries of the 300 most important resorts and pool villas in `49_SITE_BOUNDARY_CHECK` (owner 2026-11-01); a stay spread over several buildings is joined to one building until its boundary is in. The tab held the 300 stays (resort 282 · hotel 14 · bungalow 2 · other 2; 203 with a point); on 2026-11-08 it holds 731 rows (431 appended with an OpenStreetMap area): 463 `checked` (§7c: three Karon resorts from parcels on 2026-11-03, twelve stays from their OpenStreetMap area checked with the layout on 2026-11-06, 448 from their area checked automatically on 2026-11-08 — 267 own areas and 181 by the review rules), 268 `todo`. The owner reviews the less certain automatic sites after 0.3.10; a layout check turns a row into `osm_polygon+site_layout`. - 24 records whose name predates a rebrand or spells the brand differently read under their brand's listed name (§2a); the register names stay as the recorded names. Renaming the record itself is the research team's call. ## 10. Rebuild ```sh python3 -I scripts/fetch-base.py --out $BUILD/api python3 -I scripts/build-boundaries.py --pharmacy --api $BUILD/api --ogd --out . python3 -I scripts/build-receipts.py --api $BUILD/api --out . python3 -I scripts/build-data.py --sheet [--csv ] --api $BUILD/api \ --index data/area-index.json --frames data/frames --transit --shopping \ --logos [--sites <49_SITE_BOUNDARY_CHECK.csv>] --out . python3 -I scripts/build-logos.py --package --tab <48_LOGO_LIBRARY.csv> --data data/hotels.json --out . python3 scripts/build-pictogram-font.py python3 -I scripts/build-building-units.py --api $BUILD/api --geom --counts --out . npm test ``` Link changes are made in the research sheet itself, then re-exported and rebuilt — never patched in `data/`.