@mailwoman/corpus 8.5.0 → 9.0.0

package/src/shard-recipes/sub-venue.ts

@@ -0,0 +1,982 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `sub-venue` shard recipe (#35 step 4) — teach the `unit` tag the venue-INTERIOR shapes it was
+ *   never taught, so the modifier+designator class wins at the shipped `venueStructureBiasScale` of
+ *   6.0 instead of needing ~11 nats. `docs/engineering/sub-venue-corpus-task.mdx` is the spec; the
+ *   vocabulary is `corpus/data/sub-venue-lexicon.json` (v0.2.0) and the curation ledger is
+ *   `corpus/src/tools/sub-venue-promotions.ts`. The READ half — promotions, identifier
+ *   distributions, name pools — is `sub-venue-sources.ts`; this file renders lines and emits rows.
+ *
+ *   ── WHY THIS SYNTHESIZES RATHER THAN HARVESTS ────────────────────────────────────────────────────
+ *   Wave 1's lesson, and the reason the spec's "get real data first" instruction is honoured in a
+ *   shape it did not anticipate: the attested SURFACE STRINGS are thin. 87 GB features attest
+ *   `terminal`, 29 attest `wing`, 4 attest `concourse` (3 of which are a street called CONCOURSE WAY).
+ *   You cannot train a tag on 29 strings. What the five extracts DO carry at volume is the three
+ *   things a generator needs — 45,000+ real venue names across four countries, a per-region
+ *   identifier DISTRIBUTION measured over 2,868 gate/terminal/campus refs, and the confound
+ *   population that becomes the negatives. So the bulk is `designator × per-region-identifier ×
+ *   modifier` sampled per locale, and the attested strings ride along as seasoning
+ *   ({@link ATTESTED_FRACTION}) rather than as the corpus.
+ *
+ *   ── THE PER-REGION IDENTIFIER RULE, AND WHY IT IS NOT COSMETIC ───────────────────────────────────
+ *   `Gate A12` is a rendering, not a string anyone wrote down: all but 13 of Great Britain's 658
+ *   `aeroway=gate` features are unnamed and carry only a `ref`. The lexicon therefore ships a
+ *   distribution, and it differs by country far more than the shared English vocabulary suggests —
+ *   GB gates are 71% bare digits and JP 89%, FR and DE are ~60% letter-digit (`A37`, `B05`), and ES
+ *   gives a THIRD of its gates a range (`B18-B20`), which no other country does at that rate. A
+ *   generator that samples Great Britain's shape into a Spanish line produces a plausible string that
+ *   is wrong about Spain, so every leg samples its own region.
+ *
+ *   ── ONLY PROMOTED (designator, locale) PAIRS PRODUCE POSITIVES ───────────────────────────────────
+ *   A promotion names a designator, a phrase AND a locale, because the same token is a designator in
+ *   one language and a disaster in another: `hall` is 0-of-3,273 in Great Britain and 35-of-40 in
+ *   France; `wing` is 23-of-29 in Great Britain and 4-of-3,358 in the United States. A REJECTED pair
+ *   generates NEGATIVES in that locale instead — en-US `wing` rows are Red Wing, not units.
+ *
+ *   `shape: "identifier-required"` is honoured as the ledger's docstring demands: de-DE `halle` is
+ *   emitted only as `Halle <identifier>`, never bare and never after a modifier, because its 168-hit
+ *   confound includes the CITY Halle (Saale) and only the identifier-bearing shape separates them.
+ *   {@link buildSubVenueForm} enforces it and `sub-venue.test.ts` pins it.
+ *
+ *   ── LABELS ───────────────────────────────────────────────────────────────────────────────────────
+ *   Sub-venue is `unit`; the container is `venue`. That is the spec's wording and it invents nothing:
+ *   `block` / `sub_block` exist in the `ComponentTag` union but are JP-char-model-only and outside
+ *   `ACTIVE_TAGS` (STAGE3), so they are not reachable from a Latin shard.
+ *
+ *   ── WHAT IS DELIBERATELY NOT HERE ────────────────────────────────────────────────────────────────
+ *   1. **A modifier+designator form outside English.** `VENUE_STRUCTURE_MODIFIERS` is an English
+ *        list, and the extracts say the localized modifier surfaces do not exist to copy: `aile` in
+ *        France is 0 hits, `ala` in Spain 0, `flügel` in Germany 0. Generating `Terminal Sud` would
+ *        be inventing a vocabulary with no confound board behind it, which is the exact failure the
+ *        promotion ledger exists to prevent. Non-English legs get designator+identifier only.
+ *   2. **A ja-JP leg.** `ターミナル` is promoted (1,213 real of 1,215) and it is the one non-Latin
+ *        surface the task named, but Japanese addresses train through `build_jp_shard.py` against the
+ *        `stage3-jp` 47-label head, where the interior tags are `block`/`sub_block`/`building_number`
+ *        — a different model, a different label set, and a different builder. A katakana `unit` row
+ *        in this (Latin) feed would in any case be dropped by `country_weights`, which carries no
+ *        `JP` key. The JP extract is harvested and ready; the leg belongs to the JP shard.
+ */
+
+import type { ComponentTag } from "@mailwoman/core/types"
+import { dataRootPath } from "@mailwoman/core/utils"
+
+import { alignRow } from "../align.ts"
+import type { LocaleBaseTuple } from "../synthesize-german.ts"
+import type { SubVenueLexiconTable } from "../tools/sub-venue-lexicon.ts"
+import { makeMulberry32, shardSourceID, type ShardRecipe } from "./scaffold.ts"
+import {
+	buildIdentifierModel,
+	buildStreetNegatives,
+	defaultLexiconPath,
+	EMPTY_NAME_POOLS,
+	type IdentifierModel,
+	loadContextTuples,
+	type LegPools,
+	mergeNamePools,
+	type PoolQuery,
+	type PromotedSurface,
+	promotedSurfacesFor,
+	readExtractPools,
+	readPOIPools,
+	readSubVenueLexicon,
+	rejectedPhrasesFor,
+	sampleIdentifier,
+	type StreetNegatives,
+	titleCase,
+} from "./sub-venue-sources.ts"
+
+export * from "./sub-venue-sources.ts"
+
+/* oxlint-disable sister-software/no-unnamed-threshold -- the bare decimals in the register and
+   template samplers are weighted-sampler cutoffs, not thresholds: a `const r = random()` followed by
+   a cascade of `r < 0.45` branches IS the output distribution, and reading the cascade top-to-bottom
+   is how you see it. Genuine thresholds are named constants above. */
+
+//#region Plan
+
+/**
+ * One locale's leg of the shard.
+ *
+ * `positiveShare` / `negativeShare` are relative weights, normalized at run time — they do not have to sum to 1.
+ */
+export interface SubVenueLeg {
+	locale: string
+	country: string
+	/**
+	 * ISO 3166-1 alpha-2 key into the lexicon's `identifierShapes` AND the extract filename. The two axes are the same
+	 * axis: a distribution is measured in a region's own extract.
+	 */
+	region: string
+	/**
+	 * Extract filename under `--extracts-dir`. Absent = no OSM extract for this leg (en-US), which then draws its venue
+	 * and confound pools from `poi.db` instead.
+	 */
+	extract?: string
+	/**
+	 * May this leg use the English `<modifier> <designator>` grammar? See the module docstring's exclusion 1.
+	 */
+	english: boolean
+	positiveShare: number
+	negativeShare: number
+	/**
+	 * Ca-ES only — keep context tuples whose postcode starts with one of these. Catalan-language territories by postal
+	 * prefix (07 Illes Balears, 08 Barcelona, 17 Girona, 25 Lleida, 43 Tarragona) rather than by a REGION string, whose
+	 * spelling in the OA export is not something to guess at.
+	 */
+	postcodePrefixes?: readonly string[]
+}
+
+/**
+ * The legs, and the numbers behind the shares.
+ *
+ * En-GB and en-US carry the most because they are where the eval board lives (28 of the 30 confound rows are GB or US
+ * addresses) and because the English shipped vocabulary is the only one with a modifier grammar — the target class.
+ * fr-FR, de-DE and es-ES exist because the ledger promoted surfaces there (169, 19+32 and 190 real hits respectively)
+ * and a shard that skipped them would leave every non-English promotion untrained. ca-ES is small on purpose: its
+ * promotion is 15 hits and its line differs from es-ES only in the Catalan street vocabulary the postal-prefix filter
+ * selects for.
+ *
+ * The negative shares invert that ordering where the confound mass does. en-US carries the largest negative share
+ * because its confound population is the largest measured anywhere in the ledger — 3,354 `wing` (Red Wing boots 676,
+ * chicken wings 759), 2,330 `pier` (Pier 1 Imports, which IS the designator+identifier shape), 27,081 `hall`.
+ */
+export const SUBVENUE_LEGS: readonly SubVenueLeg[] = [
+	{
+		locale: "en-GB",
+		country: "GB",
+		region: "GB",
+		extract: "great-britain.jsonl",
+		english: true,
+		positiveShare: 0.25,
+		negativeShare: 0.3,
+	},
+	{ locale: "en-US", country: "US", region: "GB", english: true, positiveShare: 0.25, negativeShare: 0.35 },
+	{
+		locale: "fr-FR",
+		country: "FR",
+		region: "FR",
+		extract: "france.jsonl",
+		english: false,
+		positiveShare: 0.16,
+		negativeShare: 0.2,
+	},
+	{
+		locale: "de-DE",
+		country: "DE",
+		region: "DE",
+		extract: "germany.jsonl",
+		english: false,
+		positiveShare: 0.14,
+		negativeShare: 0.1,
+	},
+	{
+		locale: "es-ES",
+		country: "ES",
+		region: "ES",
+		extract: "spain.jsonl",
+		english: false,
+		positiveShare: 0.15,
+		negativeShare: 0.05,
+	},
+	{
+		locale: "ca-ES",
+		country: "ES",
+		region: "ES",
+		extract: "spain.jsonl",
+		english: false,
+		positiveShare: 0.05,
+		negativeShare: 0,
+		postcodePrefixes: ["07", "08", "17", "25", "43"],
+	},
+]
+
+/**
+ * En-US has no OSM extract, so its identifier distribution has to be borrowed. GB is the borrow, and the leg's `region`
+ * says so literally rather than in a comment: the two English-speaking aviation systems number their gates the same way
+ * (GB 71% bare digit) and poi.db — the only US source in reach — carries names, not refs, so it cannot supply a
+ * distribution of its own. Recorded here because it is the one place a leg's `region` is not its own country.
+ */
+export const US_IDENTIFIER_REGION_BORROWED_FROM = "GB"
+
+//#endregion
+
+//#region Tunables
+
+/**
+ * The row count this shard is built at, and the arithmetic behind it. `--count` overrides; this is the number to use
+ * absent a reason.
+ *
+ * The training sampler (`corpus-python/src/mailwoman_train/data_loader.py`, `_raw_row_stream`) draws SOURCES from a
+ * multinomial over `source_weights` and yields the next row from that source's iterator. Two consequences set the
+ * size:
+ *
+ * 1. A source's share of an epoch is `weight / Σweights`, independent of how many rows it has.
+ * 2. **A source that exhausts is DELETED from the multinomial** — there is no cycling. Under-size the shard and its
+ *    nominal dose is fiction for the rest of the epoch.
+ *
+ * Measured against the shipped `v4.1.0-gb-venue-l1e4-2k` weight table: 33 sources summing to 144.5. At the dose the B11
+ * GB-venue exercise settled on for a hard rare class — 12.0, the value `synth-fr-bare-street`, `synth-si-bare-village`,
+ * `synth-cz-pcfirst-preposition`, `synth-fr-fragment` and `synth-no-fragment` all carry — the share is `12 / 156.5 =
+ * 7.67%`, and `train_rows_per_epoch` is 1,000,000. So the epoch draws **76,677 rows** from this shard, and anything
+ * smaller runs dry mid-epoch. 120,000 clears that with room for a config that drops a source or raises the dose. (For
+ * contrast: `synth-fr-bare-street` is 10,803 rows at dose 12.0, so it exhausts 14% into its own nominal share every
+ * epoch — a precedent for the dose, not for the size.)
+ */
+export const RECOMMENDED_ROW_COUNT = 120_000
+
+/**
+ * Share of emitted rows that are NEGATIVES (the confound classes, carrying no `unit`).
+ *
+ * The spec's instruction is structural: "include the confound shapes as NEGATIVES in the same shard, or the model
+ * learns the surface rather than the structure". 0.3 is the dose the `no-fragment` recipe settled on for its own
+ * counter-distribution and there is no measurement here that beats it; `--negative-fraction` moves it.
+ */
+const DEFAULT_NEGATIVE_FRACTION = 0.3
+
+/**
+ * Share of POSITIVES whose sub-venue string is a REAL name lifted verbatim out of an extract rather than synthesized —
+ * `Terminal 2 D`, `Pier 1`, `Terminal 1 Flugsteig B`. The seasoning, per the module docstring. Kept small because the
+ * attested pool is small: after promotion + shape filtering it is 13–47 strings per leg, and a larger share would just
+ * repeat them.
+ */
+const ATTESTED_FRACTION = 0.1
+
+/**
+ * Within the SYNTHESIZED positives of an English leg: the split between the two proposal shapes.
+ *
+ * Modifier-heavy on purpose. Designator+identifier proposes at 0.85 confidence and already wins the decode at the
+ * shipped 6.0; modifier+designator proposes at 0.6 and needs 5.87–10.65. The failing class is the one to teach, and the
+ * passing one is here to not regress (`Concourse B` / `Terminal 5` / `Gate 12` / `Wing B` must stay correct).
+ */
+const ENGLISH_MODIFIER_FORM_FRACTION = 0.6
+
+//#endregion
+
+//#region Board reservation
+
+/**
+ * Surfaces reserved by `mailwoman/eval-harness/fixtures/venue-structure-confounds.jsonl` — the 30-row board this shard
+ * has to hold. A row containing any of these is DROPPED and counted in `contaminated`.
+ *
+ * The `--exclude-surfaces` precedent from `fr-fragment` / `no-fragment`, applied by hand rather than by file because
+ * the board lives in `mailwoman/` and `@mailwoman/corpus` cannot reach across that workspace boundary at run time. Keep
+ * it in sync when the board grows; a shard that trains on its own eval set measures memorization.
+ *
+ * Note what this costs and why it is still right: reserving `east gate` / `west gate` removes the two GB surfaces the
+ * board uses for its `modifier-designator-street` class, so the shard teaches that class from the OTHER real ones its
+ * sources carry (`North Gate`, `South Gate`, `East Hall`, `West Hall`, `Lower Hall`, `East Campus`, …). The class is
+ * taught; the board's own strings are not.
+ */
+export const BOARD_RESERVED_SURFACES: readonly string[] = [
+	// gb-street-gate
+	"briggate",
+	"kirkgate",
+	"castlegate",
+	"micklegate",
+	"fishergate",
+	"gallowgate",
+	"cowgate",
+	"canongate",
+	"westgate",
+	"northgate",
+	// gate-house-venue
+	"gate house",
+	"gatehouse",
+	// terminal-estate
+	"terminal industrial estate",
+	"terminal house",
+	"ocean terminal",
+	"terminal warehouse",
+	// wing-name
+	"wing yip",
+	"wing lee",
+	"bletchley park",
+	// designator-as-street
+	"campus drive",
+	"arcade avenue",
+	"concourse village",
+	"building society place",
+	"enclosure road",
+	// modifier-designator-street
+	"east gate",
+	"west gate",
+	"west wickham",
+]
+
+/**
+ * Does this row's text collide with a board-reserved surface?
+ */
+export function isBoardReserved(raw: string): boolean {
+	const low = raw.toLowerCase()
+
+	return BOARD_RESERVED_SURFACES.some((surface) => low.includes(surface))
+}
+
+//#endregion
+
+//#region Rendering
+
+/**
+ * One labelled piece of the line. Pieces inside a group are space-joined; groups are joined by the register's
+ * separator.
+ */
+interface Piece {
+	text: string
+	tag?: ComponentTag
+}
+
+type Group = Piece[]
+
+/**
+ * Surface register. Every eval in this repo gets a lowercase leg because lowercase is the register users type — Google
+ * Maps taught them — so every shard has to carry one.
+ */
+const Register = {
+	Canonical: "canonical",
+	CommaFree: "comma-free",
+	Lower: "lower",
+	Upper: "upper",
+} as const
+
+type Register = (typeof Register)[keyof typeof Register]
+
+function sampleRegister(random: () => number): Register {
+	const r = random()
+
+	if (r < 0.45) return Register.Canonical
+
+	if (r < 0.65) return Register.CommaFree
+
+	if (r < 0.9) return Register.Lower
+
+	return Register.Upper
+}
+
+/**
+ * Join groups into `raw` + `components`, applying the register to BOTH so alignment still finds every value.
+ */
+export function renderGroups(
+	groups: Group[],
+	register: Register
+): { raw: string; components: Partial<Record<ComponentTag, string>> } {
+	const fold = (text: string): string => {
+		if (register === Register.Lower) return text.toLowerCase()
+
+		if (register === Register.Upper) return text.toUpperCase()
+
+		return text
+	}
+
+	const separator = register === Register.CommaFree ? " " : ", "
+
+	const raw = groups
+		.map((group) => group.map((piece) => fold(piece.text)).join(" "))
+		.filter(Boolean)
+		.join(separator)
+
+	const components: Partial<Record<ComponentTag, string>> = {}
+
+	for (const group of groups) {
+		for (const piece of group) {
+			if (piece.tag && !components[piece.tag]) {
+				components[piece.tag] = fold(piece.text)
+			}
+		}
+	}
+
+	return { raw, components }
+}
+
+/**
+ * The street + tail groups for a country, in that country's own order.
+ *
+ * DE/ES/FR put the postcode before the locality and DE/ES put the house number after the street; GB and US keep the
+ * anglophone order and US carries a region. These are the same orders `synthesize-german.ts` renders, restated here
+ * because this recipe assembles its groups piece-by-piece (it has to, to place a sub-venue group in front of them).
+ */
+export function addressGroups(country: string, tuple: LocaleBaseTuple, withStreet: boolean): Group[] {
+	const groups: Group[] = []
+	const houseNumber = tuple.house_number?.trim()
+	const street = tuple.street.trim()
+
+	if (withStreet && street) {
+		const streetPiece: Piece = { text: street, tag: "street" }
+		const numberPiece: Piece | null = houseNumber ? { text: houseNumber, tag: "house_number" } : null
+
+		if (!numberPiece) {
+			groups.push([streetPiece])
+		} else if (country === "DE" || country === "ES") {
+			groups.push([streetPiece, numberPiece])
+		} else {
+			groups.push([numberPiece, streetPiece])
+		}
+	}
+
+	const locality: Piece = { text: tuple.locality.trim(), tag: "locality" }
+	const postcode = tuple.postcode?.trim()
+	const region = tuple.region?.trim()
+
+	if (country === "US") {
+		groups.push([locality])
+		const tail: Group = []
+
+		if (region) {
+			tail.push({ text: region, tag: "region" })
+		}
+
+		if (postcode) {
+			tail.push({ text: postcode, tag: "postcode" })
+		}
+
+		if (tail.length) {
+			groups.push(tail)
+		}
+	} else if (country === "GB") {
+		groups.push([locality])
+
+		if (postcode) {
+			groups.push([{ text: postcode, tag: "postcode" }])
+		}
+	} else {
+		// FR / DE / ES — postcode then locality, one group.
+		const tail: Group = []
+
+		if (postcode) {
+			tail.push({ text: postcode, tag: "postcode" })
+		}
+
+		tail.push(locality)
+		groups.push(tail)
+	}
+
+	return groups
+}
+
+//#endregion
+
+//#region Positive forms
+
+/**
+ * A sub-venue string plus how it was made, for the composition report.
+ */
+export interface SubVenueForm {
+	text: string
+	form: "designator-identifier" | "modifier-designator" | "attested"
+	designatorID: string
+}
+
+/**
+ * Build one sub-venue string for a leg.
+ *
+ * The `identifier-required` guard is here and not at the call site on purpose: it is the one rule in this file that a
+ * refactor must not be able to route around. A promotion carrying `shape: "identifier-required"` can only ever leave
+ * this function as `<Phrase> <identifier>`, and returns `null` rather than a bare or modified form.
+ */
+export function buildSubVenueForm(
+	leg: SubVenueLeg,
+	promoted: readonly PromotedSurface[],
+	model: IdentifierModel,
+	modifiers: readonly string[],
+	attested: readonly string[],
+	random: () => number
+): SubVenueForm | null {
+	if (!promoted.length) return null
+
+	if (attested.length && random() < ATTESTED_FRACTION) {
+		const text = attested[Math.floor(random() * attested.length)]!
+
+		return { text, form: "attested", designatorID: "attested" }
+	}
+
+	const modifierCandidates = leg.english ? promoted.filter((p) => p.modifierEligible && !p.identifierRequired) : []
+	const useModifier = modifierCandidates.length > 0 && random() < ENGLISH_MODIFIER_FORM_FRACTION
+
+	if (useModifier) {
+		const promotedSurface = modifierCandidates[Math.floor(random() * modifierCandidates.length)]!
+		const modifier = modifiers[Math.floor(random() * modifiers.length)]!
+
+		return {
+			text: `${titleCase(modifier)} ${promotedSurface.surface}`,
+			form: "modifier-designator",
+			designatorID: promotedSurface.designatorID,
+		}
+	}
+
+	const promotedSurface = promoted[Math.floor(random() * promoted.length)]!
+	const identifier = sampleIdentifier(model, promotedSurface.designatorID, random)
+
+	if (!identifier) return null
+
+	return {
+		text: `${promotedSurface.surface} ${identifier}`,
+		form: "designator-identifier",
+		designatorID: promotedSurface.designatorID,
+	}
+}
+
+/**
+ * Alias kept for readers of the arc's earlier drafts.
+ *
+ * @deprecated Use {@link buildSubVenueForm}.
+ */
+export const buildPositiveForms = buildSubVenueForm
+
+//#endregion
+
+//#region Negatives
+
+/**
+ * Negative classes, named so the composition report can count them and a failure can be attributed.
+ */
+export const NegativeClass = {
+	/**
+	 * A locale-REJECTED surface in the venue slot: Red Wing Shoes, Village Hall, Porte de Champerret.
+	 */
+	RejectedVenue: "rejected-venue",
+	/**
+	 * The designator inside a longer proper name, whole string tagged `venue`: Lochaline Ferry Terminal.
+	 */
+	LongerName: "longer-name",
+	/**
+	 * A real street whose name carries a designator token: Pier Road, Egg Hall, Orchard Gate.
+	 */
+	DesignatorStreet: "designator-street",
+	/**
+	 * A real street of the `<modifier> <designator>` shape — the class that would otherwise be read as a sub-venue.
+	 */
+	ModifierDesignatorStreet: "modifier-designator-street",
+	/**
+	 * A GB single-token `-gate` street: Eastgate, Southgate, Moorgate, Stonegate.
+	 */
+	GateSuffixStreet: "gate-suffix-street",
+	/**
+	 * A PROMOTED phrase outside the shape its promotion covers — `Halle Rosengarten`, `PHOENIX Halle`. The other half of
+	 * an `identifier-required` ruling; see `LegPools.unpromotedShapes`.
+	 */
+	UnpromotedShape: "unpromoted-shape",
+} as const
+
+export type NegativeClass = (typeof NegativeClass)[keyof typeof NegativeClass]
+
+/**
+ * Which negative classes this leg's pools can actually produce. A class with no source is ABSENT rather than
+ * substituted — the report then says so, and a reader can tell a missing class from an unsampled one.
+ */
+function availableNegativeClasses(pools: LegPools, streets: StreetNegatives): NegativeClass[] {
+	const available: NegativeClass[] = []
+
+	const sourced: ReadonlyArray<readonly [NegativeClass, number]> = [
+		[NegativeClass.RejectedVenue, pools.rejectedVenues.length],
+		[NegativeClass.LongerName, pools.longerNames.length],
+		[NegativeClass.UnpromotedShape, pools.unpromotedShapes.length],
+		[NegativeClass.DesignatorStreet, streets.designator.length],
+		[NegativeClass.ModifierDesignatorStreet, streets.modifierDesignator.length],
+		[NegativeClass.GateSuffixStreet, streets.gateSuffix.length],
+	]
+
+	for (const [negativeClass, size] of sourced) {
+		if (size > 0) {
+			available.push(negativeClass)
+		}
+	}
+
+	return available
+}
+
+//#endregion
+
+//#region Recipe
+
+/**
+ * Per-leg composition tallies the build prints and the report quotes.
+ */
+export interface SubVenueLegStats {
+	locale: string
+	positives: number
+	negatives: number
+	byForm: Record<string, number>
+	byDesignator: Record<string, number>
+	byNegativeClass: Record<string, number>
+	byRegister: Record<string, number>
+	poolSizes: Record<string, number>
+}
+
+const LICENSE =
+	"Synthetic — OpenStreetMap venue + sub-venue names (ODbL, © OpenStreetMap contributors) and Overture Places names " +
+	"(CDLA-Permissive-2.0) over OpenAddresses / HM Land Registry Price Paid Data address skeletons"
+
+const CORPUS_VERSION = "0.16.0"
+
+const bump = (record: Record<string, number>, key: string): void => {
+	record[key] = (record[key] ?? 0) + 1
+}
+
+/**
+ * Everything the emit loop needs that does not vary per row.
+ */
+interface EmitContext {
+	write: (line: string) => void
+	source: string
+	random: () => number
+	modifiers: readonly string[]
+	designatorPhrases: readonly string[]
+	counters: { emitted: number; skipped: number; contaminated: number }
+}
+
+/**
+ * Render one row, drop it if it collides with the eval board, align it, write it.
+ *
+ * A free function rather than a closure over the leg loop: a closure there is both a lint error (`no-loop-func`) and a
+ * real hazard, since it would capture the loop's mutable counters.
+ */
+function emitRow(
+	context: EmitContext,
+	leg: SubVenueLeg,
+	stats: SubVenueLegStats,
+	groups: Group[],
+	register: Register,
+	synthMethod: string,
+	disambiguator: Record<string, string>
+): boolean {
+	const { raw, components } = renderGroups(groups, register)
+
+	if (isBoardReserved(raw)) {
+		context.counters.contaminated++
+
+		return false
+	}
+
+	const aligned = alignRow({
+		raw,
+		components,
+		country: leg.country,
+		locale: leg.locale,
+		source: context.source,
+		source_id: shardSourceID(context.source, { ...components, ...disambiguator }),
+		corpus_version: CORPUS_VERSION,
+		license: LICENSE,
+	})
+
+	if (aligned.kind !== "labeled" || !aligned.row) {
+		context.counters.skipped++
+
+		return false
+	}
+
+	context.write(JSON.stringify({ ...aligned.row, synth_method: synthMethod, synth_base_id: null }) + "\n")
+
+	context.counters.emitted++
+	bump(stats.byRegister, register)
+
+	return true
+}
+
+/**
+ * Emit one leg's POSITIVE rows: `<sub-venue> unit`, a real `venue`, and the leg's own address skeleton.
+ */
+function emitPositives(
+	context: EmitContext,
+	leg: SubVenueLeg,
+	pools: LegPools,
+	promoted: readonly PromotedSurface[],
+	model: IdentifierModel,
+	stats: SubVenueLegStats,
+	target: number
+): void {
+	const { random } = context
+	let produced = 0
+	let guard = 0
+
+	while (produced < target && guard++ < target * 8) {
+		if (!pools.context.length || !pools.venues.length) break
+		const form = buildSubVenueForm(leg, promoted, model, context.modifiers, pools.attested, random)
+
+		if (!form) {
+			context.counters.skipped++
+
+			continue
+		}
+
+		const tuple = pools.context[Math.floor(random() * pools.context.length)]!
+		const venue = pools.venues[Math.floor(random() * pools.venues.length)]!
+
+		// A venue name that CONTAINS the sub-venue string (or vice versa) makes the two spans
+		// unresolvable — alignment claims the longer one and quarantines the other — and the row would
+		// teach an overlap that never occurs on a real envelope. Redraw instead.
+		const lowVenue = venue.toLowerCase()
+		const lowForm = form.text.toLowerCase()
+
+		if (lowVenue.includes(lowForm) || lowForm.includes(lowVenue)) continue
+
+		const register = sampleRegister(random)
+		const subGroup: Group = [{ text: form.text, tag: "unit" }]
+		const venueGroup: Group = [{ text: venue, tag: "venue" }]
+		const r = random()
+		// A quarter of rows carry no street: an airport terminal's address usually does not have one.
+		const body = addressGroups(leg.country, tuple, r >= 0.25)
+		// Both orders occur on real signage and mail — "Terminal 5, Heathrow" and "Heathrow, Terminal 5".
+		const groups = r < 0.55 ? [subGroup, venueGroup, ...body] : [venueGroup, subGroup, ...body]
+
+		const ok = emitRow(context, leg, stats, groups, register, `sub-venue:${form.form}`, {
+			leg: leg.locale,
+			form: form.form,
+			v: String(produced),
+		})
+
+		if (!ok) continue
+
+		produced++
+
+		stats.positives++
+		bump(stats.byForm, form.form)
+		bump(stats.byDesignator, form.designatorID)
+	}
+}
+
+/**
+ * Emit one leg's NEGATIVE rows — the confound classes, none of which carries a `unit`.
+ */
+function emitNegatives(
+	context: EmitContext,
+	leg: SubVenueLeg,
+	pools: LegPools,
+	stats: SubVenueLegStats,
+	target: number
+): void {
+	const { random } = context
+	const streets = buildStreetNegatives(pools.context, context.designatorPhrases, context.modifiers, leg.country)
+	const available = availableNegativeClasses(pools, streets)
+	let produced = 0
+	let guard = 0
+
+	while (produced < target && guard++ < target * 8) {
+		if (!available.length || !pools.context.length) break
+		const negativeClass = available[Math.floor(random() * available.length)]!
+		const register = sampleRegister(random)
+		let groups: Group[]
+
+		if (negativeClass === NegativeClass.DesignatorStreet) {
+			groups = addressGroups(leg.country, pickTuple(streets.designator, random), true)
+		} else if (negativeClass === NegativeClass.ModifierDesignatorStreet) {
+			groups = addressGroups(leg.country, pickTuple(streets.modifierDesignator, random), true)
+		} else if (negativeClass === NegativeClass.GateSuffixStreet) {
+			groups = addressGroups(leg.country, pickTuple(streets.gateSuffix, random), true)
+		} else {
+			const pool =
+				negativeClass === NegativeClass.RejectedVenue
+					? pools.rejectedVenues
+					: negativeClass === NegativeClass.LongerName
+						? pools.longerNames
+						: pools.unpromotedShapes
+
+			const name = pool[Math.floor(random() * pool.length)]!
+			const tuple = pools.context[Math.floor(random() * pools.context.length)]!
+
+			groups = [[{ text: name, tag: "venue" }], ...addressGroups(leg.country, tuple, random() < 0.75)]
+		}
+
+		const ok = emitRow(context, leg, stats, groups, register, `sub-venue-negative:${negativeClass}`, {
+			leg: leg.locale,
+			negative: negativeClass,
+			v: String(produced),
+		})
+
+		if (!ok) continue
+
+		produced++
+
+		stats.negatives++
+		bump(stats.byNegativeClass, negativeClass)
+	}
+}
+
+function pickTuple(pool: readonly LocaleBaseTuple[], random: () => number): LocaleBaseTuple {
+	return pool[Math.floor(random() * pool.length)]!
+}
+
+/**
+ * Split `total` across `shares` (which need not sum to 1), largest-remainder so the parts sum exactly.
+ */
+export function allocate(total: number, shares: readonly number[]): number[] {
+	const sum = shares.reduce((a, b) => a + b, 0)
+
+	if (sum <= 0) return shares.map(() => 0)
+	const exact = shares.map((s) => (total * s) / sum)
+	const floored = exact.map((v) => Math.floor(v))
+	let remainder = total - floored.reduce((a, b) => a + b, 0)
+	const order = exact.map((v, i) => ({ i, frac: v - Math.floor(v) })).toSorted((a, b) => b.frac - a.frac)
+
+	for (const entry of order) {
+		if (remainder <= 0) break
+		floored[entry.i]! += 1
+
+		remainder--
+	}
+
+	return floored
+}
+
+/**
+ * Build one leg's pools: the name pools from its extract and/or poi.db, plus its address context.
+ */
+function buildLegPools(
+	leg: SubVenueLeg,
+	query: PoolQuery,
+	contextByCountry: ReadonlyMap<string, LocaleBaseTuple[]>,
+	paths: { extractsDir: string; poiDb: string }
+): LegPools {
+	const extractPools = leg.extract ? readExtractPools(`${paths.extractsDir}/${leg.extract}`, query) : EMPTY_NAME_POOLS
+
+	// poi.db holds four countries; only these two legs are inside it.
+	const poiPools =
+		leg.country === "US" || leg.country === "FR" ? readPOIPools(paths.poiDb, leg.country, query) : EMPTY_NAME_POOLS
+
+	const names = mergeNamePools(extractPools, poiPools)
+	let context = contextByCountry.get(leg.country) ?? []
+
+	if (leg.postcodePrefixes) {
+		const prefixes = leg.postcodePrefixes
+		const filtered = context.filter((t) => prefixes.some((p) => (t.postcode ?? "").startsWith(p)))
+
+		// A leg that filters itself empty is a build-time fact worth failing on, not a silent fallback
+		// to the parent locale's rows under a different `locale` stamp.
+		if (!filtered.length) {
+			throw new Error(
+				`${leg.locale}: no context tuples matched postcode prefixes ${prefixes.join(",")} among ${context.length} ${leg.country} rows`
+			)
+		}
+
+		context = filtered
+	}
+
+	return { context, ...names }
+}
+
+function emptyStats(leg: SubVenueLeg, pools: LegPools, promotedCount: number): SubVenueLegStats {
+	return {
+		locale: leg.locale,
+		positives: 0,
+		negatives: 0,
+		byForm: {},
+		byDesignator: {},
+		byNegativeClass: {},
+		byRegister: {},
+		poolSizes: {
+			promoted: promotedCount,
+			context: pools.context.length,
+			venues: pools.venues.length,
+			attested: pools.attested.length,
+			rejectedVenues: pools.rejectedVenues.length,
+			longerNames: pools.longerNames.length,
+			unpromotedShapes: pools.unpromotedShapes.length,
+		},
+	}
+}
+
+/**
+ * Shard recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const subVenueRecipe: ShardRecipe = {
+	name: "sub-venue",
+	description:
+		"Venue-interior structure (#35): <sub-venue> unit + <venue> lines per promoted (designator, locale) pair, with the rejection ledger's confounds as negatives",
+	mode: "generate",
+	options: [
+		{ flag: "--lexicon <path>", description: "sub-venue lexicon JSON (default: the committed corpus/data one)" },
+		{
+			flag: "--extracts-dir <dir>",
+			description: "OSM sub-venue extract JSONLs (default: $MAILWOMAN_DATA_ROOT/sub-venue/extracts)",
+		},
+		{
+			flag: "--poi-db <path>",
+			description: "poi.db for the en-US / fr-FR pools (default: $MAILWOMAN_DATA_ROOT/poi/poi.db)",
+		},
+		{ flag: "--sub-venue-tuples <path>", description: "GB/US/FR address-context tuples JSONL" },
+		{
+			flag: "--negative-fraction <n>",
+			description: `share of rows that are confound negatives (default ${DEFAULT_NEGATIVE_FRACTION})`,
+		},
+	],
+	async run(opts, write) {
+		if (opts.count == null) throw new Error("sub-venue recipe requires --count <N>")
+		const count = opts.count
+		const negativeFraction = opts.negativeFraction ?? DEFAULT_NEGATIVE_FRACTION
+		const extractsDir = opts.extractsDir ?? dataRootPath("sub-venue", "extracts")
+		const poiDb = opts.poiDb ?? dataRootPath("poi", "poi.db")
+		const tuplesPath = opts.subVenueTuples ?? dataRootPath("corpus", "intermediate", "house-venue-tuples-v3.jsonl")
+		const lexicon: SubVenueLexiconTable = readSubVenueLexicon(opts.lexicon ?? defaultLexiconPath())
+
+		const context: EmitContext = {
+			write,
+			source: opts.sourceName ?? "synth-sub-venue",
+			random: makeMulberry32(opts.seed),
+			modifiers: lexicon.modifiers.filter((m) => m.shipped).map((m) => m.id),
+			designatorPhrases: lexicon.designators.filter((d) => d.tier === "subvenue").map((d) => d.id),
+			counters: { emitted: 0, skipped: 0, contaminated: 0 },
+		}
+
+		console.error(`  reading context tuples: ${tuplesPath}`)
+
+		const contextByCountry = await loadContextTuples(tuplesPath, opts.seed)
+		const legPools = new Map<string, LegPools>()
+		const legStats = new Map<string, SubVenueLegStats>()
+		const legPromoted = new Map<string, PromotedSurface[]>()
+
+		for (const leg of SUBVENUE_LEGS) {
+			const promoted = promotedSurfacesFor(leg.locale, lexicon)
+
+			const query: PoolQuery = {
+				promoted,
+				rejectedPhrases: rejectedPhrasesFor(leg.locale),
+				designatorPhrases: context.designatorPhrases,
+				modifiers: context.modifiers,
+				english: leg.english,
+			}
+
+			const pools = buildLegPools(leg, query, contextByCountry, { extractsDir, poiDb })
+
+			legPromoted.set(leg.locale, promoted)
+			legPools.set(leg.locale, pools)
+			legStats.set(leg.locale, emptyStats(leg, pools, promoted.length))
+		}
+
+		const positiveTotal = Math.round(count * (1 - negativeFraction))
+
+		const positiveQuota = allocate(
+			positiveTotal,
+			SUBVENUE_LEGS.map((l) => l.positiveShare)
+		)
+
+		const negativeQuota = allocate(
+			count - positiveTotal,
+			SUBVENUE_LEGS.map((l) => l.negativeShare)
+		)
+
+		for (const [index, leg] of SUBVENUE_LEGS.entries()) {
+			const pools = legPools.get(leg.locale)!
+			const stats = legStats.get(leg.locale)!
+			const promoted = legPromoted.get(leg.locale)!
+			const model = buildIdentifierModel(lexicon, leg.region)
+
+			emitPositives(context, leg, pools, promoted, model, stats, positiveQuota[index]!)
+			emitNegatives(context, leg, pools, stats, negativeQuota[index]!)
+		}
+
+		for (const stats of legStats.values()) {
+			console.error(
+				`  ${stats.locale}: +${stats.positives} positives / -${stats.negatives} negatives ` +
+					`forms=${JSON.stringify(stats.byForm)} designators=${JSON.stringify(stats.byDesignator)} ` +
+					`negclasses=${JSON.stringify(stats.byNegativeClass)} registers=${JSON.stringify(stats.byRegister)} ` +
+					`pools=${JSON.stringify(stats.poolSizes)}`
+			)
+		}
+
+		return { read: count, ...context.counters }
+	},
+}
+
+//#endregion