@mailwoman/corpus 10.0.0 → 10.1.0

package/lib/fr/recipes/bare-street.ts

@@ -3,46 +3,50 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  *
- *   `fr-bare-street` slice recipe (#251) — the postcode-anchoring-imbalance change. BAN (and every
- *   other comprehensive FR source) is postcode-COMPLETE, so the model learned the French
- *   street→locality boundary as "the token after the 5-digit postcode," never as "comma + city." Strip
- *   the postcode and it leaks the street's proper-noun tokens into the following locality ("Rue René
- *   Cassin, Paris" → street="Rue Ren", locality="Cassin"). This recipe mints the MISSING distribution:
- *   the BARE comma form, NO postcode, real `(street, number, city)` tuples from BAN (Licence Ouverte —
- *   permissive; the model stays clean of ODbL, unlike the opt-in OSM rooftop slices).
- *
- *   Each tuple → `<n> <Rue/Avenue/…> <proper-noun name>, <City>` with the FR prefix split
- *   ({@link decomposeFrStreet}: "Rue" → street_prefix, the rest → street). Tuples whose street carries
- *   no recognized FR type word are skipped — the failing class is precisely the prefix-led street.
- *
- *   ⚠ Convention loss-mask: this recipe TEACHES FR `street_prefix`. The conventions loss-mask forbids it
- *   for FR and will `-inf` these gold labels (the v1.6.0 ~7M-loss blow-up). Disable that mask for any
- *   run including this slice.
+ *   Generate postcode-free French examples from BAN with separate street-prefix labels.
  */
 
 import { FR_VOIE_TYPES } from "@mailwoman/codex/fr"
 import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils"
 
-import { decomposeFrStreet } from "#fr/adapters/ban/street-decompose"
-import { alignAndWrite, readTuples, type CorpusRecipe, sliceSourceID } from "#recipes/scaffold"
+import { decomposeFrStreet } from "#fr/street-decompose"
+import { alignAndWrite, readTuples, type CorpusRecipe, recipeSourceID } from "#recipes/scaffold"
+import { defaultRecipeSource } from "#recipes/sources"
+import { SourceRegister } from "#registers"
+import { SurfaceOrigin } from "#types"
+
+/**
+ * Shared provenance for postcode-free BAN rows.
+ */
+/**
+ * Resolved once, because `source_id` includes the source as its prefix and a pair
+ * that disagreed would name a source no row of this output declares.
+ */
+const FR_BARE_STREET_SOURCE = defaultRecipeSource("synth-fr-bare-street")
+
+const FR_BARE_STREET_PROVENANCE = {
+	register: SourceRegister.BaseAdresseNationale,
+	surface: SurfaceOrigin.Composed,
+	// The BAN record behind each row exists.
+	// The tuple file has no `sourceID` to identify it.
+	// The extraction must emit that ID to populate this field (#2359).
+	baseSourceID: null,
+}
 
 /**
- * Canonical voie type (lowercase, accent-kept) → its most common written abbreviation, from the codex table's first
- * entry. Types with no attested abbreviation stay canonical in the abbreviated form.
+ * First abbreviation listed for each street type.
  */
 const FR_VOIE_ABBREV: Record<string, string> = Object.fromEntries(
 	Object.entries(FR_VOIE_TYPES).flatMap(([canonical, abbrevs]) => (abbrevs[0] ? [[canonical, abbrevs[0]]] : []))
 )
 
 /**
- * The order-cycle slot for the BARE-STREET-ONLY form (`«voie» «name»`, no number, no locality) — the absence
- * counterweight to the locality-terminated comma-free forms (see the cycle comment).
+ * Index of the street-only form.
  */
 const BARE_STREET_ONLY_FORM = 3
 
 /**
- * Slice 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.
+ * Register the French bare-street recipe.
  */
 export const frBareStreetRecipe: CorpusRecipe = {
 	name: "fr-bare-street",
@@ -50,7 +54,7 @@ export const frBareStreetRecipe: CorpusRecipe = {
 		"FR bare street+city, NO postcode (#251): comma / comma-free / abbreviated-voie surfaces over one label set",
 	mode: "tuples",
 	async run(opts, write) {
-		// Seeded for parity with the other recipes; unused beyond reproducibility (the tuples drive the content).
+		// Preserve the recipe's seeded initialization.
 		makeMulberry32(opts.seed)
 		let read = 0
 		let emitted = 0
@@ -70,11 +74,7 @@ export const frBareStreetRecipe: CorpusRecipe = {
 
 			const { prefix, street } = decomposeFrStreet(fullStreet)
 
-			// A no-prefix nom_voie ("La Ville Mois") is not the prefix-led class the numbered forms
-			// exercise — but as a BARE surface it is exactly the non-voie-led counterweight the v4.5.1
-			// probe showed missing ('Savile Row'-shaped spans still fell to the trailing-locality
-			// prior; the voie-led bare form protected only voie-led spans). Alternate rows emit the
-			// whole span as a bare street; the rest skip as before.
+			// Retain some multiword streets without a recognized type as counterexamples.
 			if (!prefix || !street) {
 				if (read % 2 === 0 && fullStreet.split(" ").length >= 2) {
 					const bare = {
@@ -82,8 +82,8 @@ export const frBareStreetRecipe: CorpusRecipe = {
 						components: { street: fullStreet } as Record<string, string>,
 						country: "FR",
 						locale: "fr-FR",
-						source: "synth-fr-bare-street",
-						source_id: sliceSourceID("synth-fr-bare-street", {
+						source: FR_BARE_STREET_SOURCE,
+						source_id: recipeSourceID(FR_BARE_STREET_SOURCE, {
 							street: fullStreet,
 							f: "bare-nonvoie",
 							v: String(read),
@@ -93,7 +93,7 @@ export const frBareStreetRecipe: CorpusRecipe = {
 							"Synthetic — fr-bare-street; (street, number, city) from BAN (Base Adresse Nationale, Licence Ouverte)",
 					}
 
-					if (alignAndWrite(write, bare, "fr-bare-street")) {
+					if (alignAndWrite(write, bare, "fr-bare-street", FR_BARE_STREET_PROVENANCE)) {
 						emitted++
 					} else {
 						skipped++
@@ -107,18 +107,7 @@ export const frBareStreetRecipe: CorpusRecipe = {
 				continue
 			}
 
-			// Four surfaces over the same tuple, cycled deterministically. The comma form was the
-			// original change; the COMMA-FREE form is the colloquial register users actually type
-			// ('12 rue de Rome Paris' — the street↔locality boundary with NO delimiter, the fr-fr
-			// panel's named loss); the ABBREVIATED form is the typeahead register the geocoder-tester
-			// FR slice attests at scale; and the BARE-STREET-ONLY form is the absence counterweight —
-			// without it, every delimiter-free surface in the mix ENDS in a locality, the model learns
-			// "trailing span = locality" as categorical, and bare street names across locales flip to
-			// locality wholesale (the v4.5.0 no-promote's measured erosion: 'Calle de Alcalá',
-			// 'Madison Square West', and COMER's fork all fell to that prior). Tags are identical
-			// where present; each component VALUE is the span as written (BIO alignment binds value
-			// to surface); the bare form carries NO number and NO locality because the surface has
-			// neither.
+			// Rotate comma-separated, compact, abbreviated and street-only variants.
 			const form = read % 4
 			const prefixSurface = form >= 2 ? (FR_VOIE_ABBREV[prefix.toLowerCase()] ?? prefix) : prefix
 
@@ -127,9 +116,7 @@ export const frBareStreetRecipe: CorpusRecipe = {
 					? { street_prefix: prefix, street }
 					: { house_number: number, street_prefix: prefixSurface, street, locality }
 
-			// When the tuple carries a WOF-attested neighbourhood, the comma slot renders the
-			// THREE-slot middle surface — the dependent-locality counterweight (the v4.5.1 erosion's
-			// untouched half: the two-slot comma-free endings squeezed the middle tag out).
+			// Include a neighborhood in the comma-separated form when available.
 			const hood = String(t.neighbourhood ?? "").trim()
 
 			if (form === 0 && hood) {
@@ -145,21 +132,21 @@ export const frBareStreetRecipe: CorpusRecipe = {
 						? `${prefix} ${street}`
 						: `${number} ${prefixSurface} ${street} ${locality}`
 
-			const source_id = sliceSourceID("synth-fr-bare-street", { ...components, f: String(form), v: String(read) })
+			const source_id = recipeSourceID(FR_BARE_STREET_SOURCE, { ...components, f: String(form), v: String(read) })
 
 			const canonical = {
 				raw,
 				components,
 				country: "FR",
 				locale: "fr-FR",
-				source: "synth-fr-bare-street",
+				source: FR_BARE_STREET_SOURCE,
 				source_id,
 				corpus_version: "0.9.4",
 				license:
 					"Synthetic — fr-bare-street; (street, number, city) from BAN (Base Adresse Nationale, Licence Ouverte)",
 			}
 
-			if (alignAndWrite(write, canonical, "fr-bare-street")) {
+			if (alignAndWrite(write, canonical, "fr-bare-street", FR_BARE_STREET_PROVENANCE)) {
 				emitted++
 			} else {
 				skipped++

package/lib/fr/recipes/fragment.ts

@@ -3,93 +3,58 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  *
- *   `fr-fragment` slice recipe (#727 T2) — the HOUSE-NUMBER-LICENCE change.
+ *   Generates French street fragments without a postcode from BAN tuples.
  *
- *   The measured problem (T1c, `2026-07-16-t1c-fragment-board-verdict.md`): the shipped model scores
- *   **0.925** on `<n> Rue X` and **0.215** on `Rue X`. Same streets, same model; the only difference
- *   is a leading number. It has learned that a house number LICENSES a street reading at all — strip
- *   it and a designator-led phrase parses as a locality, designator included:
- *
- *   ```
- *   "Rue Montmartre"        -> locality="Rue Montmartre"
- *   "Allee Poque"           -> locality="Allee Poque"
- *   ```
- *
- *   `Rue` can only mean street in French. The model is not mislabelling an ambiguous toponym; it is
- *   mislabelling `Rue`.
- *
- *   WHY THE EXISTING RECIPE DOESN'T COVER THIS. {@link frBareStreetRecipe} (#251) mints
- *   `<n> Rue <name>, <City>` — the bare COMMA form, no postcode. It targets postcode-anchoring
- *   imbalance, and every row it emits still carries a house number AND a locality. It cannot teach
- *   the class above, because it never shows the model a street standing alone.
- *
- *   WHAT THIS MINTS. Five street forms — the first three carry NO house number and NO locality, which
- *   is the whole point — plus the counter-distribution:
- *
- *   1. `bare-street`         "Rue Montmartre"
- *   2. `street-particle`     "Rue de la Paix"
- *   3. `date-name`           "Allee du 11 Novembre 1918"
- *   4. `street-housenumber`  "12 Rue Montmartre"     — the anchor, so the licence isn't UNLEARNED
- *   5. `alnum-housenumber`   "12 bis Rue Montmartre"
- *   6. `bare-locality`       "Mery-sur-Oise"         — NEGATIVE: a bare toponym that IS a locality
- *
- *   The **admin/street homonym** class (`Rue de Rome`) is deliberately NOT a separate form. Homonym
- *   streets are already in the register and fall into 1–3 naturally; there is nothing different to
- *   teach about them, because the lesson is identical — the designator makes it a street. The
- *   fragment board scores them separately (they need measuring, not minting), and the recipe would
- *   need a second pass over the commune set to label them, buying a `synth_method` string and no
- *   training signal.
- *
- *   FORM 7 IS NOT OPTIONAL. T1c's standing prediction: the board's `bare-locality` cell reads 0.980
- *   for the WRONG REASON — the model calls everything without a house number a locality, and on bare
- *   localities that is accidentally right. Teach bare streets alone and the model has every incentive
- *   to flip that default rather than learn the distinction, trading a 0.215 for a 0.980. The slice
- *   must show BOTH bare forms so the discriminating evidence is the designator, which is the only
- *   thing that actually distinguishes them. This is the same counter-distribution principle
- *   {@link noStreetRecipe} established after synth-street pushed the model into "decompose mode".
- *
- *   SPLIT. `--exclude-surfaces` takes the fragment board's reserved surface list
- *   (`mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt`). Every listed surface is
- *   skipped — source-disjoint by normalized street SURFACE, never by record row. Row-disjoint leaks
- *   the surface across the boundary and measures memorization of `Rue de Rivoli` while claiming
- *   generalization to unseen streets. The recipe REFUSES to run without the list rather than
- *   silently minting a contaminated slice.
- *
- *   MIX. The slice is ~145K rows off a 120K-tuple draw; the intended corpus mix is **5–10%**, set at
- *   assembly time by slice weight rather than by row count. Keep the cap: a slice that fixes fragments
- *   by degrading full addresses has moved the failure, not fixed it — which is what
- *   `street-housenumber` / `alnum-housenumber` on the fragment board and the global parity floor are
- *   there to catch. `date-name` is ~0.5% of the slice because BAN only holds ~1,418 date-name streets
- *   after filtering and the tuple extractor already takes every one; if that class needs more, the
- *   change is slice weight, not invented data.
- *
- *   ⚠ Convention loss-mask: like {@link frBareStreetRecipe}, this recipe TEACHES FR `street_prefix`.
- *   The conventions loss-mask forbids it for FR and will `-inf` these gold labels (the v1.6.0 ~7M-loss
- *   blow-up). Disable that mask for any run including this slice.
+ *   The output mixes designator-led streets with bare commune names as counterexamples.
+ *   Some streets include a house number.
+ *   `--exclude-surfaces` keeps streets on the fragment evaluation board out of training. A model trained on
+ *   these rows needs the French `street_prefix` loss mask disabled.
  */
 
+import { titleCaseFR } from "@mailwoman/codex/fr"
+import { sample } from "@mailwoman/core/random"
 import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils"
 import { TextSpliterator } from "spliterator"
 
-import { decomposeFrStreet } from "#fr/adapters/ban/street-decompose"
-import { alignAndWrite, readTuples, type CorpusRecipe, sliceSourceID } from "#recipes/scaffold"
-import { pick } from "#synthesizers/utils"
+import { decomposeFrStreet } from "#fr/street-decompose"
+import { alignAndWrite, readTuples, type CorpusRecipe, recipeSourceID } from "#recipes/scaffold"
+import { defaultRecipeSource } from "#recipes/sources"
+import { SourceRegister } from "#registers"
+import { SurfaceOrigin } from "#types"
 
 /**
- * House numbers, weighted toward the small values that dominate real BAN rows.
+ * The provenance for every row.
+ * Street and commune names both come from BAN.
+ */
+/**
+ * Resolved once, because `source_id` includes the source as its prefix and a pair
+ * that disagreed would name a source no row of this output declares.
  */
+const FR_FRAGMENT_SOURCE = defaultRecipeSource("synth-fr-fragment")
+
+const FR_FRAGMENT_PROVENANCE = {
+	register: SourceRegister.BaseAdresseNationale,
+	surface: SurfaceOrigin.Composed,
+	// The BAN record behind each row exists.
+	// The tuple file has no `sourceID` to identify it.
+	// The extraction must emit that ID to populate this field (#2359).
+	baseSourceID: null,
+}
 
+/**
+ * House numbers to sample from.
+ * Repeated entries weight the draw toward small values.
+ */
 const HOUSE_NUMBERS = [
 	1, 1, 2, 2, 3, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 16, 18, 20, 21, 24, 27, 30, 33, 42, 57, 68, 84, 102, 115, 140,
 ]
 
 /**
- * FR alphanumeric house-number forms. `bis`/`ter` are separated; a bare letter is suffixed.
+ * French repetition and letter suffixes for house numbers.
  */
 const ALNUM_SUFFIXES = ["bis", "ter", "A", "B"]
 
-// Recipe-fidelity: this accent-STRIPPING fold is the surface key every committed fr-fragment slice and the fragment
-// board's reserved list were built with — not the diacritic-KEEPING `foldNOSurface` the Norwegian slices share.
+// The reserved-surface list is stored in this folded form: accents stripped, lowercase, single spaces.
 const norm = (value: string): string =>
 	value
 		.normalize("NFD")
@@ -99,69 +64,26 @@ const norm = (value: string): string =>
 		.trim()
 
 /**
- * French commune convention: capitalize each element, leave the joining particles lowercase. `saint-jean-de-luz` →
- * `Saint-Jean-de-Luz`, not `Saint-Jean-De-Luz`.
- *
- * Needed because BAN's sliced DBs keep only `locality_base` — normalized, lowercase, accent-stripped. Emitting that
- * verbatim would teach the counter-distribution that a lowercase accent-stripped string is a locality, which is not a
- * fact about French and would not match the fragment board (which reconstructs the same casing). The accents are gone
- * from the source and cannot be recovered here; the casing can.
- */
-const FR_LOWER = new Set([
-	"le",
-	"la",
-	"les",
-	"de",
-	"du",
-	"des",
-	"d",
-	"l",
-	"sur",
-	"sous",
-	"en",
-	"aux",
-	"au",
-	"et",
-	"lez",
-])
-
-export function frTitleCase(value: string): string {
-	const cap = (token: string, first: boolean): string =>
-		!first && FR_LOWER.has(token) ? token : token.charAt(0).toUpperCase() + token.slice(1)
-
-	return value
-		.split(" ")
-		.map((word, wordIndex) =>
-			word
-				.split("-")
-				.map((bit, bitIndex) => cap(bit, wordIndex === 0 && bitIndex === 0))
-				.join("-")
-		)
-		.join(" ")
-}
-
-/**
- * Does the street name carry a particle? Decides the particle vs bare classification.
+ * Matches common French particles in a street name.
  */
 const PARTICLE = /\b(de la|de l'|du|des|de|d'|le|la|les)\b/i
 
 /**
- * Does the street name carry date material (a year, or a day + French month)?
+ * Matches a year or a French day-and-month phrase in a street name.
  */
 const DATEISH =
 	/\b(1[0-9]|20)\d{2}\b|\b\d{1,2}\s+(janvier|f[ée]vrier|mars|avril|mai|juin|juillet|ao[ûu]t|septembre|octobre|novembre|d[ée]cembre)\b/i
 
-// A quarter of house numbers carry an alphanumeric suffix (`8 bis`, `12B`).
+// The share of numbered rows whose house number has a suffix.
 const ALNUM_HOUSE_NUMBER_SHARE = 0.25
 
 /**
- * Slice 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.
+ * The recipe registered with the corpus builder.
  */
 export const frFragmentRecipe: CorpusRecipe = {
 	name: "fr-fragment",
 	description:
-		"FR street fragments with NO house number (#727 T2): the house-number-licence change — bare/particle/date-name/homonym + the bare-locality counter",
+		"FR street fragments with NO house number (#727 T2): the house-number-license change — bare/particle/date-name/homonym + the bare-locality counter",
 	mode: "tuples",
 	options: [
 		{
@@ -184,7 +106,7 @@ export const frFragmentRecipe: CorpusRecipe = {
 		if (!excludePath) {
 			throw new Error(
 				"fr-fragment: --exclude-surfaces is REQUIRED. Pass the fragment board's reserved list " +
-					"(mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt) or this slice trains on its own eval set. " +
+					"(mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt) or this recipe trains on its own eval set. " +
 					"Source-disjoint by street SURFACE is the split discipline; there is no safe default."
 			)
 		}
@@ -192,10 +114,8 @@ export const frFragmentRecipe: CorpusRecipe = {
 		const excluded = new Set<string>()
 
 		for await (const line of TextSpliterator.fromAsync(excludePath)) {
-			const trimmed = line.trim()
-
-			if (trimmed && !trimmed.startsWith("#")) {
-				excluded.add(trimmed)
+			if (!line.startsWith("#")) {
+				excluded.add(line)
 			}
 		}
 
@@ -204,8 +124,7 @@ export const frFragmentRecipe: CorpusRecipe = {
 		const hnProb = opts.hnProb ?? 0.35
 		const bareLocalityProb = opts.bareProb ?? 0.25
 
-		// The locality pool is harvested from the tuples themselves — every BAN row carries its
-		// commune, so the counter-distribution needs no second source.
+		// Commune names collected here feed the bare-locality counterexamples.
 		const localities = new Set<string>()
 
 		let read = 0
@@ -228,7 +147,6 @@ export const frFragmentRecipe: CorpusRecipe = {
 				continue
 			}
 
-			// THE SPLIT. A surface on the fragment board never enters training.
 			if (excluded.has(norm(fullStreet))) {
 				contaminated++
 
@@ -237,8 +155,7 @@ export const frFragmentRecipe: CorpusRecipe = {
 
 			const { prefix, street } = decomposeFrStreet(fullStreet)
 
-			// The failing class is the designator-led street. A no-prefix nom_voie ("La Ville Mois")
-			// is a different problem and would muddy the signal.
+			// The recipe keeps only streets that start with a designator such as "rue".
 			if (!prefix || !street) {
 				skipped++
 
@@ -251,9 +168,9 @@ export const frFragmentRecipe: CorpusRecipe = {
 			let klass = DATEISH.test(street) ? "date-name" : PARTICLE.test(street) ? "street-particle" : "bare-street"
 
 			if (carriesNumber) {
-				const number = pick(HOUSE_NUMBERS, random)
+				const number = sample(HOUSE_NUMBERS, random)
 				const alnum = random() < ALNUM_HOUSE_NUMBER_SHARE
-				const suffix = pick(ALNUM_SUFFIXES, random)
+				const suffix = sample(ALNUM_SUFFIXES, random)
 
 				const houseNumber = alnum
 					? suffix === "bis" || suffix === "ter"
@@ -266,7 +183,7 @@ export const frFragmentRecipe: CorpusRecipe = {
 				klass = alnum ? "alnum-housenumber" : "street-housenumber"
 			}
 
-			const sourceID = sliceSourceID("synth-fr-fragment", { ...components, v: String(read) })
+			const sourceID = recipeSourceID(FR_FRAGMENT_SOURCE, { ...components, v: String(read) })
 
 			if (
 				alignAndWrite(
@@ -276,12 +193,13 @@ export const frFragmentRecipe: CorpusRecipe = {
 						components,
 						country: "FR",
 						locale: "fr-FR",
-						source: "synth-fr-fragment",
+						source: FR_FRAGMENT_SOURCE,
 						source_id: sourceID,
 						corpus_version: "0.9.4",
 						license: "Synthetic — fr-fragment; (street, commune) from BAN (Base Adresse Nationale, Licence Ouverte)",
 					},
-					`fr-fragment:${klass}`
+					`fr-fragment:${klass}`,
+					FR_FRAGMENT_PROVENANCE
 				)
 			) {
 				emitted++
@@ -290,18 +208,16 @@ export const frFragmentRecipe: CorpusRecipe = {
 			}
 		}
 
-		// MARK: counter-distribution — bare localities
-		// Minted last so the locality pool is complete. Without these the model can satisfy every row
-		// above by flipping its default from "bare => locality" to "bare => street", which trades one
-		// broken prior for another and would show up as bare-locality collapsing on the board.
+		// MARK: Bare-locality counterexamples
+
+		// The count makes bare localities `bareLocalityProb` of all emitted rows.
 		const pool = [...localities].toSorted()
 		const wanted = Math.round((emitted / Math.max(1, 1 - bareLocalityProb)) * bareLocalityProb)
 
 		for (let i = 0; i < wanted && pool.length; i++) {
-			// BAN gives `locality_base` normalized; restore the casing the fragment board also
-			// reconstructs, so train and eval show the model the same shape of French.
-			const name = frTitleCase(pick(pool, random))
-			const sourceID = sliceSourceID("synth-fr-fragment", { locality: name, v: `neg-${i}` })
+			// The fragment evaluation board uses title-cased commune names.
+			const name = titleCaseFR(sample(pool, random))
+			const sourceID = recipeSourceID(FR_FRAGMENT_SOURCE, { locality: name, v: `neg-${i}` })
 
 			if (
 				alignAndWrite(
@@ -311,12 +227,13 @@ export const frFragmentRecipe: CorpusRecipe = {
 						components: { locality: name },
 						country: "FR",
 						locale: "fr-FR",
-						source: "synth-fr-fragment",
+						source: FR_FRAGMENT_SOURCE,
 						source_id: sourceID,
 						corpus_version: "0.9.4",
 						license: "Synthetic — fr-fragment counter-distribution; commune from BAN (Licence Ouverte)",
 					},
-					"fr-fragment:bare-locality"
+					"fr-fragment:bare-locality",
+					FR_FRAGMENT_PROVENANCE
 				)
 			) {
 				emitted++