diff --git a/wip/274-method.md b/wip/274-method.md
new file mode 100644
index 0000000..2926679
--- /dev/null
+++ b/wip/274-method.md
@@ -0,0 +1,215 @@
+---
+title: 274 - how the roster was built
+categories:
+  - Songwriting notes
+wiki:
+  publish: false
+---
+
+# 274: how the roster was built
+
+This is the provenance document for the 川端橋–右左府 roster and the interactive
+transect page. The roster itself lives in [274 reference tables](274-dev.md); the song
+page is [274](274.md).
+
+Nothing here is authoritative about the road. It is a record of where each number came
+from, which numbers are derived rather than measured, and which were corrected by hand
+against reality.
+
+## What exists
+
+| File | What it is |
+|---|---|
+| `274-dev.md` | The roster. 168 entries, 37.27–107.83 km, 26 kinds, 95 distinct names, 118 with readings. Source of truth. |
+| `274-lookup.csv` | The same rows, 16 columns, generated from the Markdown. |
+| `274-lookup.geojson` | The same rows as points, plus the route line and the district bands. |
+| `274-transect.html` | Self-contained interactive page. Opens from disk; also published as an Artifact. |
+| `274-route.json` | 1,715 route points with their kilometre. |
+| `274-districts.json` | 23 大字 bands with entry and exit kilometres. |
+| `274-transect-build.py` | Re-reads the roster and rewrites the page's data block. |
+| `274-romaji.py` | Fills the Romaji column from the Kana column. |
+
+Flow is one-way: **`274-dev.md` → everything else.** Editing a generated file is pointless;
+it will be overwritten on the next build.
+
+## Scope
+
+The range is **川端橋 (37.27 km) to 右左府 (106.22 km)**, chosen by the author, not derived.
+
+- West end at the bridge rather than the settlement: 川端's substance lies upstream, west
+  of the bridge, and is deliberately outside the frame. The JR station was excluded for
+  the same reason and later re-added.
+- East end at 右左府 — 発電所, 変電所 and the transmission line 右左府線 — after which the
+  road continues east out of the piece.
+- Everything beyond is out of scope: 千呂露橋, 大颱橋, the whole of 日勝峠, the Tokachi
+  descent. Full-route material stays in the KB at `~/infra/kb/inbox/surreal/md/`.
+
+Two entries sit past the nominal east end (滝の沢川 107.29, ルウケシオマナイ川 107.83)
+because they fell inside the computation window; they are not part of the intended range.
+
+## Sources
+
+**北海道開発局 個別施設計画, 令和8年3月** — the national facility registers, three PDFs:
+橋梁長寿命化修繕計画, 道路トンネル個別施設計画, 大型の構造物個別施設計画. Gives every bridge,
+tunnel and shed with length, width, type, build year, inspection year and 判定区分. No
+coordinates and no kilometrage. Text extracted with pypdf; columns reconstructed by
+anchoring on the inspection year and grade at the right and the bureau/office at the left.
+**Not checked against the original page images.**
+
+**OpenStreetMap** (ODbL, retrieved 2026-09-08/09) — the route geometry, and most points of
+interest. Route relation `r6651380`. Reading tags `name:ja-Hira` and `name:ja-Latn` where
+they exist, which was 42 of the original 170 rows.
+
+**国土数値情報 河川データセット (W05)** via the Geoshape repository at 国立情報学研究所 —
+watercourse names and geometry for the 石狩川, 鵡川 and 沙流川 systems. All 561 named rivers
+were fetched, not only the 「その他」 group: named rivers carry their own 河川コード and are
+served separately, which is the mistake that caused the first pass to miss 夕張川, 鵡川,
+沙流川, 穂別川, サヌシュッペ川, both ウシャップ rivers and ホルカクルキ川.
+
+**国土地理院 逆ジオコーディング** — `mreversegeocoder.gsi.go.jp`. Coordinate to municipality
+code and 大字. Joined to GSI's `muni.js` for the municipality name including the 郡.
+
+**日本郵便 郵便番号データ** (Hokkaido file, CP932) — kana readings for municipality and
+町域, folded from half-width katakana to hiragana.
+
+**route01.com r274p2–p4** and **hgf03030.a.la9.jp** — two drivers' 2019 trip reports. The
+only source for structure *order*, which the registers do not carry. Also the source for
+several details no dataset has: the painted-out hexagon on the 道道610 sign, the bear
+statue at the drive-in, the 合目 markers restarting on the Shimizu side.
+
+**Google Maps** — via the author, for everything OSM lacks. See below.
+
+## Method
+
+**Kilometre datum.** A node graph is built from the ways of the OSM route relation.
+Dijkstra from a seed node gives distance along the road for every node. The relation is
+broken across the divided-carriageway stretch in 札幌, so the seed is the west end of the
+largest connected component, near 上野幌 — **not the official start**. Comparison against
+KP values in a 帯広開発建設部 disaster report suggests official KP ≈ these figures + 14 km,
+unverified. One kilopost read on Street View would settle it. Everything downstream —
+ordering, district bands, offsets — is internally consistent regardless.
+
+**Point features** are projected to their nearest route node; the offset column is the
+straight-line distance from feature to road.
+
+**Structures** have no coordinates in the registers, so their positions are derived two
+ways. First, OSM tags each bridge and tunnel as its own way, giving a position and a
+geometric length; aligning that ordered sequence against the registers' ordered lengths by
+Needleman–Wunsch transfers the positions. 穂高トンネル lands within 9 m. Second, where a
+bridge is named after the watercourse it crosses, the geometric crossing point gives an
+independent position that overrides the alignment — this fixed 草木橋, ポン久留喜橋,
+オロロップ橋, モトツ橋, 福山大橋 and 新川橋.
+
+The alignment is reliable in the 60–97 km core, where tunnels anchor it, and unreliable at
+both ends where small bridges cluster and OSM ways do not correspond one-to-one. 唐沢橋
+(9.1 m) and 福本沢川橋 (3.0 m) have no position at all — below the size OSM splits out.
+
+**Watercourses** are positioned at closest approach, which equals the crossing for streams
+the road crosses and is arbitrary for streams it runs beside. サヌシュッペ川 is tangent at
+two reaches 5 km apart for this reason; 沙流川 is placed where the road joins its valley
+at 右左府, not at its nearest point, which falls past the range end.
+
+**District bands** come from reverse-geocoding the road itself every 250 m — 287 samples
+producing 23 bands with entry and exit kilometres. This is the only method here that can
+say where the road enters and leaves a district; point data cannot, because representative
+points sit at settlement centres.
+
+**Name extraction** strips generic tails (橋, 川, トンネル, 覆道, 郵便局, 発電所, 町, 丁目),
+strips the Ainu qualifiers パンケ/ペンケ/ポン, and aliases variants together (十三哩→十三里,
+滝ノ上→滝の上, 浪ノ沢→浪の沢, 能見→熊見, クルキ→久留喜). Mechanical and meant to be corrected.
+
+**Romaji** is generated from kana, never from kanji, in wāpuro style — おう→ou, ゅう→yuu,
+っ doubles, ー repeats, ん→n' before a vowel or y. The trailing generic is split off and
+lowercased. Kana is the source of truth; where kana is blank, romaji stays blank.
+
+## Human in the loop
+
+The machine-derived roster was wrong in ways no amount of cross-referencing would have
+caught. Every item below came from the author reading signs, satellite imagery and Street
+View, or from knowledge not in any dataset.
+
+**Corrections to derived data**
+
+- **十三哩橋 / 十三里橋** — the register writes 哩, the kanji for *mile*; the plate on the
+  bridge reads 里. Confirmed on Street View. As a place name it is read **とみさと**. The
+  origin is the 13-mile post from 追分駅, so the name runs mile → ri → hamlet, and the
+  register is the only place the mile survives.
+- **ホルカクルキ川** — I claimed it had no counterpart in W05. It does, 河川コード
+  8101030169; I had fetched only 12 of 561 named rivers. The 北海道 plate on the bank
+  confirms the spelling and gives the Ainu original **ホロカ・クルヒ (horka・kurihi)**,
+  "to turn back" and "its shadow" — *horka* marks a stream running counter to its parent.
+  route01's ホロカクルキ is a slip.
+- **モトツ橋** — I had it at 86.16 km on the パンケモトツ沢川 crossing. It is at 87.12,
+  the sheltered span between モトツトンネル and 福山トンネル. The identical 62 m lengths of
+  bridge and shelter are the same deck, not a coincidence. Street View and satellite,
+  July 2025: the two tunnels now read as continuous.
+- **下福山橋 and オロロップ橋** — order swapped by the author to 77.05 / 77.55, matching
+  route01 against my length-alignment.
+- **滝ノ上信号場** — demoted from passenger station in 2024. Not in any dataset consulted.
+- **登川** — the active community now falls under 楓 jurisdiction; 登川 is largely the
+  administrative remnant of Yubari's first settlements. This explains why 登川神社, 登川橋
+  and 登川郵便局 all geocode to 夕張市楓 while the district called 登川 begins further east.
+- **穂別ダム神社** belongs to 穂別長和 — the dam and reservoir are in that district and only
+  the shrine fell the wrong side of an old boundary. **稲里神社**, the proper shrine for
+  穂別稲里, is too far inland to reach from R274. Neither fact is derivable from the
+  geocoded address, which reports only where each thing physically sits.
+
+**Found by hand, absent from OpenStreetMap**
+
+滝下八幡神社 · 滝下発電所 · 登川神社 · 豊進神社 · 穂別ダム神社 · 福山神社 · 名石駐車場 ·
+日高神社 · 日高発電所 · 右左府発電所
+
+Ten entries, all located from Google Maps pins resolved to coordinates, then positioned
+against the route and geocoded like any other row. OSM carried two of the eight shrines.
+A shrine 18 m from the carriageway (穂別ダム神社) was missing, so the shrine count is a
+floor rather than a total, and parking and rest areas are effectively unmapped along the
+whole route — 名石駐車場 came through only because it has a Google place listing.
+
+**Judgement calls**
+
+Readings were filled by hand and are not all guaranteed; the plaques will settle the
+doubtful ones. Rivers are being skipped. Shops, restaurants, bus stops, temples, fuel
+stations, campsites and hotels were removed deliberately. 名称不明, ベリベツ川 and
+デチン沢川 were excluded with reasons recorded. A vacant Seicomart was rejected as an entry
+and read instead as a relocation to the shop 859 m away.
+
+## Errors found and fixed
+
+Recorded because they indicate where this kind of pipeline fails.
+
+- **Duplicate-named features shared one set of coordinates.** A name-keyed lookup meant
+  both セイコーマート pointed at the same spot 50 km apart; likewise three ENEOS and paired
+  bus stops. 12 coordinate cells were wrong.
+- **穂高トンネル is 1,991 m long and was built in 1991.** The parser took the first
+  year-shaped token and grabbed the length. Fixed by anchoring build year and age from the
+  row end. The only such collision on the route.
+- **W05 coverage** — see ホルカクルキ川 above. Assuming the 「その他」 group was the whole
+  dataset produced a confident false negative.
+- **Over-stripping** — 登川 reduced to 登, and のぼりかわ romanised as "Nobori kawa", because
+  a trailing 川 looks generic. Both are toponyms where it is integral.
+- **A table column mismatch** in the page: a header change applied while its matching cell
+  insert silently did not, shifting every column right of Kana by one.
+- **A join key type mismatch** left the 大字 column empty in a committed CSV.
+
+## Limits
+
+- Distances are **not official kiloposts**.
+- Structure positions are derived, not surveyed. Two have none at all.
+- Register figures are not checked against the original page images.
+- Shrine and parking coverage is a floor, not a total.
+- The roster records where each object physically sits. It cannot express which community
+  an object belongs to — that distinction lives in the prose.
+- The underlying 全国道路施設点検データベース restricts commercial use and copying of the
+  dataset. Facility names and build years are facts; wholesale redistribution of the tables
+  is not.
+
+## Regenerating
+
+```sh
+python3 wip/274-romaji.py --check   # dry run: report romaji differences
+python3 wip/274-romaji.py           # fill the Romaji column from Kana
+python3 wip/274-transect-build.py   # rewrite the page's data block
+```
+
+Both are stdlib-only and idempotent. The CSV, the GeoJSON and the published Artifact are
+regenerated from the same Markdown by hand.