@mailwoman/corpus 10.3.0 → 10.5.0

package/lib/it/adapters/anac/adapter.ts

@@ -29,7 +29,7 @@
  * public bodies, and the one role whose parties can be sole traders publishes no address.
  * {@linkcode readANACParty} still applies a name check and refuses a party whose name takes the
  * `Surname, Firstname` form, because a later edition could add one. On the 2025 edition the check
- * refuses no party: 0 of the 18,740 address-bearing buyer and payer parties take that form, while 56
+ * refuses no party: 0 of the 18,740 buyer and payer parties that have an address take that form, while 56
  * hold a comma inside an organization name — `PARCO NAZIONALE DEL CILENTO, VALLO DI DIANO E
  * ALBURNI` and `SETTORE PATRIMONIO, AMBIENTE E LAVORI PUBBLICI` among them. The check therefore
  * requires one word on each side of the comma rather than any comma.
@@ -304,8 +304,9 @@ const KILOMETRE_POINT = /[\s,]+km\.?[\s,]*\d+(?:[.,]\d+)?$/iu
 /**
  * A name of the form `Surname, Firstname`: one comma, one word on each side, letters only.
  *
- * An organization name in this release holds a comma often and holds a word on each side of it rarely:
- * `PARCO NAZIONALE DEL CILENTO, VALLO DI DIANO E ALBURNI` has eight words after the comma.
+ * An organization name in this release can hold a comma, but a name with a single word on
+ * each side of the comma is rare: `PARCO NAZIONALE DEL CILENTO, VALLO DI DIANO E ALBURNI`
+ * has eight words after the comma.
  * The single word on each side is what the inverted personal form has
  * and what a department or authority name does not.
  */

package/lib/jp/adapters/wof-admin/adapter.ts

@@ -13,8 +13,8 @@
  *   1. **Reversed ordering** — region → county → locality → block → house_number "東京都世田谷区南町1-2-3" not
  *        "1-2-3 Minamicho, Setagaya, Tokyo"
  *   2. **No street names** — most JP addresses use a block/sub-block grid system. The "neighborhood"
- *        placetype (丁目) is the closest analog to a street but is actually a grid cell.
- *   3. **Postcode-first convention** — addresses are often prefixed with `〒NNN-nnnn`.
+ *        placetype (丁目) is the closest analog to a street but is a grid cell.
+ *   3. **Postcode-first convention** — addresses can be prefixed with `〒NNN-nnnn`.
  *
  *   Schema mapping to ComponentTags (subset of Stage 3 + JP-specific Phase 6 tags):
  *

package/lib/kr/adapters/juso/label-rows.ts

@@ -152,7 +152,7 @@ function joinNumber(main: string, sub: string): string {
 
 export interface ReadJusoOptions {
 	/**
-	 * Stop after this many rows per 시도, for a fixture or a smoke run.
+	 * Stop after this number of rows per 시도, for a fixture or a smoke run.
 	 * Omit for the whole register.
 	 */
 	maxRowsPerRegion?: number

package/lib/kr/adapters/localdata/key-index.ts

@@ -142,7 +142,8 @@ export function aliasKeyIndex(index: KeyIndex): void {
 }
 
 /**
- * How many tokens from `start` form one of the region's 시군구 — 2 for `수원시 장안구`, 1 for `종로구`, 0 for none.
+ * The number of tokens from `start` that form one of the region's 시군구 —
+ * 2 for `수원시 장안구`, 1 for `종로구`, 0 for none.
  *
  * Two-token first, because `수원시` by itself is also listed and taking it would
  * leave `장안구` to be read as a road.

package/lib/lu/adapters/bd-adresses/adapter.ts

@@ -33,7 +33,7 @@
  * Measured over the 179,622 rows of the 2026-09-28 edition with a generic-word probe rather than a
  * language classifier: 133,449 values open with a French street generic such as `Rue`, `Avenue` or
  * `Montée`, 18,629 close with a Luxembourgish one such as `-strooss`, `-wee` or `-gaass`, and the
- * remaining 27,544 match neither probe and include plainly Luxembourgish forms such as
+ * remaining 27,544 match neither probe and include Luxembourgish forms such as
  * `Am Stengber`, `An Hiesel` and `Um Reebou`.
  * No column marks a row's language, so `locale` reads `und-LU`.
  * `und` is BCP-47's undetermined language subtag, and the repository already stores it in

package/lib/nl/adapters/kadaster/adapter.ts

@@ -360,8 +360,8 @@ function documentChunks(documentPath: PathBuilderLike): AsyncIterable<Uint8Array
 /**
  * Every document under `inputPath`, where `inputPath` is one GML file or a directory holding them.
  *
- * A directory is read because the download is one national file today and the
- * service document could list more than one entry tomorrow.
+ * A directory is read because the download is one national file now and a later
+ * service document could list more than one entry.
  * A directory holding no document raises rather than yielding zero rows, so a wrong
  * path reports itself instead of reading as an empty publisher.
  */

package/lib/nl/tools/fetch/kadaster.ts

@@ -54,12 +54,12 @@ export const NL_KADASTER_FEED_URL = "https://service.pdok.nl/kadaster/ad/atom/ad
  * The adapter's `inputPath` states this name.
  *
  * The publisher's own name for the file.
- * The adapter reads a `.gml.gz` by inflating it, so the suffix is load-bearing.
+ * The adapter reads a `.gml.gz` by inflating it, so the suffix is required.
  */
 export const NL_KADASTER_DOWNLOAD_FILENAME = "addresses.gml.gz"
 
 /**
- * The attribution recorded in the manifest.
+ * The manifest records this attribution.
  *
  * CC0 1.0 reserves no act and owes no attribution clause.
  * This attribution is a credit the repository chooses, rather than a condition of the grant.

package/lib/parquet/shuffle.ts

@@ -10,8 +10,8 @@ import { shuffleWith } from "@mailwoman/core/random"
 /**
  * Rows held in memory while shuffling, when the caller names no window.
  *
- * Five 50,000-row groups span more than one output row-group and reach `wof-admin`'s
- * whole 13-country set, while a window below `ROW_GROUP_SIZE` buys no wider mix.
+ * Five 50,000-row groups span more than one output row-group and reach `wof-admin`'s whole
+ * 13-country set, while a window below `ROW_GROUP_SIZE` draws from no more than one row-group.
  */
 export const DEFAULT_SHUFFLE_WINDOW = 250_000
 

package/lib/parquet/span-sql.ts

@@ -7,7 +7,7 @@
  *   A parquet row stores component offsets in `span_starts` and `span_ends` arrays.
  *   A third array, `span_tags`, identifies each component. The text comes from slicing `raw` at those offsets.
  *   A query that groups or filters by a
- *   component performs that slice in SQL instead of decoding every row in the process.
+ *   component takes that substring in SQL instead of decoding every row in the process.
  *
  *   `holdoutComponents` in `#tools/overlay/split-slice` performs the same read in JavaScript.
  *   The SQL and JavaScript operations use different string indexing. The difference is described below.
@@ -24,8 +24,8 @@
  * `alignRow` records a span in UTF-16 code units.
  *
  * DuckDB's `substr` counts characters.
- * The two agree for the Basic Multilingual Plane and diverge by one unit per astral character,
- * so a row carrying one slices late by the number of astral characters before the span.
+ * The two agree for the Basic Multilingual Plane and diverge by one unit per astral character, so in a
+ * row that contains one, the substring starts late by the number of astral characters before the span.
  *
  * In the first 40 train files of `v0.6.0-register-surface`, 6,982 of 40,000,000
  * rows (0.0175%) contain an astral character.
@@ -34,15 +34,17 @@
  * `𐍀𐍂𐍉𐍆𐌹𐌳𐌰𐌹𐌽𐍃, RHODE ISLAND` read its region as `ND`, taken from `ISLAND`.
  *
  * So this is sound for an aggregate over millions of rows and unsound for a decision about one row.
- * A caller that acts per row narrows with this and confirms with `holdoutComponents`, which
- * slices with JavaScript string semantics and therefore reads the offsets as they were written.
+ * A caller that acts per row narrows with this and confirms with `holdoutComponents`.
+ *
+ * That function indexes with JavaScript string semantics and therefore reads
+ * the offsets as they were written.
  */
 export function componentAtSpanSQL(tag: string, rawColumn = "raw"): string {
 	return componentAtIndexSQL(`list_position(span_tags, '${tag}')`, rawColumn)
 }
 
 /**
- * The same slice against a span index the caller already computed.
+ * The same substring against a span index the caller already computed.
  *
  * A query can hoist `list_position(span_tags, …)` into a subquery column and read it by name.
  * A second evaluation in the projection would run the search twice per row.

package/lib/parquet/streams.ts

@@ -63,7 +63,7 @@ export interface ParquetRowStreamOptions {
 	 */
 	columns?: readonly string[]
 	/**
-	 * Stop after this many rows.
+	 * Stop after this number of rows.
 	 */
 	limit?: number
 }

package/lib/parquet/writers.ts

@@ -5,7 +5,7 @@
  * @file Writes Parquet files and the manifest that describes their contents.
  *
  *   This module exposes two writers. {@linkcode writeParquetFile} writes one file from rows already in memory.
- *   {@linkcode writeParquetSplits} streams per-split row iterables into as many files as the row cap requires. It
+ *   {@linkcode writeParquetSplits} streams per-split row iterables into the number of files the row cap requires. It
  *   returns the manifest. Both writers create their parent directory, as `@mailwoman/core/fs` writers do
  *   encode rather than repeat at each call site.
  *
@@ -212,7 +212,7 @@ export interface WriteParquetSplitsOptions {
 	 * A descriptor is appended only after its `.parquet` file closes, so each recorded file is complete.
 	 *
 	 * A resumed run rereads each descriptor's size.
-	 * It discards that many rows from the front of each split's iterable and opens the next file index.
+	 * It discards that number of rows from the front of each split's iterable and opens the next file index.
 	 *
 	 * The caller therefore has to supply the same iterables in the same order.
 	 * For `buildCorpus` that means the same labeled files and the same shuffle seed.

package/lib/pl/tools/fetch/emuia.ts

@@ -155,7 +155,7 @@ export interface HarvestEMUiAOptions {
 	 */
 	pageSize?: number
 	/**
-	 * Stop once the manifest holds this many pages, for a probe rather than a full harvest.
+	 * Stop once the manifest holds this number of pages, for a probe rather than a full harvest.
 	 */
 	maxPages?: number
 	/**
@@ -165,7 +165,7 @@ export interface HarvestEMUiAOptions {
 	 * The measurement costs about 2·log2(n) requests — 51 for a register of this size — and the
 	 * harvest does not need it: paging to the first empty page reads the type whole either way.
 	 *
-	 * It buys a denominator for the progress lines and a figure for the manifest,
+	 * It supplies a denominator for the progress lines and a figure for the manifest,
 	 * and it rests on the monotone-presence assumption this module's header states.
 	 */
 	measureCount?: boolean

package/lib/recipes/scaffold.ts

@@ -170,7 +170,7 @@ export interface OATupleFields {
 
 export interface ReadOATuplesOptions<T> {
 	/**
-	 * Stop after this many distinct tuples.
+	 * Stop after this number of distinct tuples.
 	 *
 	 * The `break` closes the reader and releases the archive.
 	 * GB-scale countrywide extracts need this behavior.
@@ -496,7 +496,7 @@ export interface RecipeOptions {
 	 */
 	venueNames?: PathBuilderLike
 	/**
-	 * `venue`: share of locality rows that also carry the tuple's region after the locality.
+	 * `venue`: share of locality rows that also print the tuple's region after the locality.
 	 *
 	 * @defaultValue `0.5`
 	 */

package/lib/recipes/sources.ts

@@ -285,7 +285,7 @@ export const CARRIED_SOURCES: ReadonlyArray<string> = [
 	"osm-bd",
 	// One id per French overseas jurisdiction for the BAN department files,
 	// written by `corpus run ban --source-name`.
-	// Metropolitan rows keep `ban`, so its weight does not decide how often a territory is drawn.
+	// Metropolitan rows keep `ban`, so its weight does not decide the rate at which a territory is drawn.
 	"ban-gp",
 	"ban-mq",
 	"ban-gf",
@@ -371,7 +371,7 @@ export function defaultRecipeSource(retired: string): string {
  *
  * Its `source-names.json` records the mapping applied per file with the row count
  * and the ordered-`source_id` digest verified on both sides.
- * `v0.7.0-de-holdout/corpus-v0.7.0-de-holdout` is assembled from those bytes over 766 slices.
+ * `v0.7.0-de-holdout/corpus-v0.7.0-de-holdout` is assembled from those bytes over 766 recipe-output files.
  *
  * A recipe writing the retired spelling from here on would produce an output
  * disagreeing with the corpus it joins.
@@ -391,7 +391,7 @@ export const WRITE_CURRENT_SOURCE_NAMES = true
  * throw on every row in the corpus.
  *
  * Callers of those two tables resolve the key through this, so a lookup answers under
- * either spelling and a genuinely unknown source still refuses.
+ * either spelling and a source unknown under both spellings still refuses.
  *
  * The value is returned unchanged for a source outside the table — an adapter id,
  * or an overlay source id — because those tables are keyed by whatever the producer emits.

package/lib/recipes/sub/venue/render.ts

@@ -31,7 +31,7 @@ export type Group = Piece[]
 /**
  * The case and punctuation style of a rendered row.
  *
- * Users often type addresses in lowercase, so every recipe output includes lowercase rows.
+ * Some users type addresses in lowercase, so every recipe output includes lowercase rows.
  */
 export const Register = {
 	Canonical: "canonical",

package/lib/recipes/sub/venue/sources.ts

@@ -330,7 +330,7 @@ const MIN_OWN_SHAPE_OBSERVATIONS = 20
 /**
  * Builds a region's identifier model from the lexicon's `identifierShapes`.
  *
- * The pooled fallback combines the gate, terminal and campus distributions.
+ * The pooled fallback combines the airport gate, terminal and campus distributions.
  */
 export function buildIdentifierModel(lexicon: SubVenueLexiconTable, region: string): IdentifierModel {
 	const byDesignator = new Map<string, ShapeBucket[]>()
@@ -515,8 +515,9 @@ const POI_CONFOUND_CATEGORIES: readonly string[] = [
 ]
 
 /**
- * Reads the venue and confound name pools for one country from the Overture place-names file,
- * which holds `name`, `category` and `country` for every country the release covers.
+ * Reads the venue and confound name pools for one country from the Overture place-names file.
+ *
+ * That file holds `name`, `category` and `country` for every country the release covers.
  *
  * The venue categories fill the `venues` pool.
  * Every row of a venue or confound category is tested against the rejected phrases

package/lib/recipes/sub/venue.ts

@@ -52,7 +52,8 @@ export interface SubVenueLeg {
 	/**
 	 * The ISO 3166-1 alpha-2 key into the lexicon's `identifierShapes`.
 	 *
-	 * Gate and terminal numbering differs by country, so each leg samples identifiers from its own region.
+	 * Airport gate and terminal numbering differs by country, so each leg samples
+	 * identifiers from its own region.
 	 */
 	region: string
 	/**
@@ -142,7 +143,7 @@ export const SUBVENUE_LEGS: readonly SubVenueLeg[] = [
  * The region whose identifier distribution the en-US leg uses.
  *
  * The en-US leg has no OSM extract.
- * The Overture names file carries venue names without refs.
+ * The Overture names file holds venue names without refs.
  * The recipe borrows the identifier distribution from GB.
  */
 export const US_IDENTIFIER_REGION_BORROWED_FROM = "GB"
@@ -514,7 +515,7 @@ function emitPositives(
 		const subGroup: Group = [{ text: form.text, tag: "unit" }]
 		const venueGroup: Group = [{ text: venue, tag: "venue" }]
 		const r = random()
-		// Some rows omit the street because venue addresses such as airport terminals often lack one.
+		// Some rows omit the street because venue addresses such as airport terminals can lack one.
 		const body = addressGroups(leg.country, tuple, r >= NO_STREET_SHARE)
 
 		// An empty body means that no address layout covers this country.

package/lib/recipes/venue.ts

@@ -5,11 +5,11 @@
  *
  *   `venue` recipe. A real venue name written as the whole query, and the same name followed by a
  *   locality drawn from the country's admin tuples. The rows teach the model that
- *   `Manchester Art Gallery` is a venue rather than a locality and a region, which the venue-head
- *   prior in `@mailwoman/neural` also asserts, so the model and the prior agree.
+ *   `Manchester Art Gallery` is a venue rather than a locality and a region. The venue-head
+ *   prior in `@mailwoman/neural` asserts the same reading, so the model and the prior agree.
  *
  *   Venue names come from the Overture place-names Parquet file the venue-head table is built
- *   from, which holds `name`, `category` and `country` for every place at confidence 0.85 or
+ *   from. That file holds `name`, `category` and `country` for every place at confidence 0.85 or
  *   higher. The file has no locality column, so the locality row joins a name to a tuple of the
  *   same country and asserts no fact about where the venue stands, so its surface is
  *   `SurfaceOrigin.Composed`. The bare row is the publisher's own string, so its surface is
@@ -25,7 +25,8 @@
 
 import { dataRootPath } from "@mailwoman/core/data-root"
 import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils"
-import { venueHeadLexicon } from "@mailwoman/poi-taxonomy/venue-heads"
+import { foldCaseWhitespace } from "@mailwoman/normalize/fold"
+import { isSpacelessScriptWord, venueHeadLexicon } from "@mailwoman/poi-taxonomy/venue-heads"
 import type { PathBuilderLike } from "path-ts"
 
 import { normalizeGauntletSurface, readGauntletInputs } from "#gauntlet-inputs"
@@ -59,7 +60,7 @@ export const DEFAULT_VENUE_NAMES: PathBuilderLike = dataRootPath(
 export const DEFAULT_NAMES_PER_COUNTRY = 20_000
 
 /**
- * Share of locality rows that also carry the region when `--venue-region-fraction` is absent.
+ * Share of locality rows that also print the region when `--venue-region-fraction` is absent.
  */
 export const DEFAULT_REGION_FRACTION = 0.5
 
@@ -105,13 +106,12 @@ export interface VenueHeadSuffixLookup {
 	suffix(word: string): number | null
 }
 
-const SPACELESS_SCRIPT_RE = /^[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\p{Script=Thai}]+$/u
-
 /**
  * Characters a spaceless-script name needs before it is written as a bare query.
  *
- * A two-character name such as `東京` is the length of a city name, and the head the table recognizes is
- * the name's last character, so a name shorter than this is a head with one character of stem.
+ * A two-character name such as `東京` is the length of a city name.
+ * The head the table recognizes is the name's last character, so a name shorter
+ * than this is a head with one character of stem.
  */
 export const MIN_SPACELESS_BARE_CHARACTERS = 3
 
@@ -135,7 +135,7 @@ export function isBareVenueName(name: string, heads?: VenueHeadSuffixLookup): bo
 	return (
 		Boolean(heads) &&
 		[...word].length >= MIN_SPACELESS_BARE_CHARACTERS &&
-		SPACELESS_SCRIPT_RE.test(word) &&
+		isSpacelessScriptWord(word) &&
 		heads!.suffix(word) !== null
 	)
 }
@@ -144,7 +144,7 @@ export function isBareVenueName(name: string, heads?: VenueHeadSuffixLookup): bo
  * The comparison key for a name against the country's localities.
  */
 export function venueNameKey(name: string): string {
-	return name.normalize("NFC").trim().toLowerCase().replaceAll(/\s+/gu, " ")
+	return foldCaseWhitespace(name.normalize("NFC"))
 }
 
 /**

package/lib/registers.ts

@@ -204,6 +204,13 @@ export const SourceRegister = {
 	 * HM Land Registry Price Paid Data, England and Wales.
 	 */
 	LandRegistryPricePaid: "gb-hm-land-registry-ppd",
+	/**
+	 * Ordnance Survey's OpenData products under the Open Government Licence v3.0.
+	 *
+	 * The `gb-os-open` adapter joins three of them per row.
+	 * Code-Point Open, OS Open Names and OS Open Roads share one publisher and one license.
+	 */
+	OrdnanceSurveyOpenData: "gb-os-opendata",
 	/**
 	 * Australia's Geocoded National Address File.
 	 */

package/lib/ro/adapters/retea-scolara/adapter.ts

@@ -16,7 +16,7 @@
  * The sheet splits the address into `Strada`, `Numar`, `Cod postal` and `Localitate unitate`, so the
  * adapter splits no street line. Measured over the 2025–2026 edition's 18,022 units:
  *
- * - `Strada` is the street name as the school writes it, usually without the street-type word:
+ * - `Strada` is the street name as the school writes it, in most rows without the street-type word:
  *   `REPUBLICII`, `PRINCIPALA`, `Piața 1848`. 198 values begin with `STR.` and 247 with `CALEA`.
  *   The value is copied unchanged, because a street type the publisher did not write is not evidence.
  * - `Numar` holds the house number. `FN` (`fără număr`, "without number") and `-` state that the premise
@@ -36,7 +36,7 @@
  *
  * ## Identity and license
  *
- * The row id is content-addressed over the aligned components. A school and its kindergarten often share
+ * The row id is content-addressed over the aligned components. A school and its kindergarten can share
  * one address under two names, so they are two rows.
  *
  * data.gov.ro publishes the dataset under `CC-BY-4.0`, the license its CKAN record names for the

package/lib/runner.ts

@@ -47,9 +47,9 @@ export interface RunnerProgress {
 export const DEFAULT_DEDUP_SLOTS_LOG2 = 27
 
 /**
- * How many distinct dedup keys a `Set`-backed run holds before it stops adding new ones.
+ * The number of distinct dedup keys a `Set`-backed run holds before it stops adding new ones.
  *
- * A V8 `Set` refuses a 16,777,216th entry outright, so this cap was a real bound rather than a preference.
+ * A V8 `Set` throws on a 16,777,216th entry, so this cap was a real bound rather than a preference.
  * It applies only to a run that opts out of the fingerprint table
  * through {@linkcode RunAdapterOptions.dedupMaxSize}.
  */
@@ -78,7 +78,7 @@ export interface RunAdapterOptions {
 	adapterOptions: AdapterOptions
 
 	/**
-	 * Hold dedup keys in a V8 `Set` capped at this many, rather than in a fingerprint table.
+	 * Hold dedup keys in a V8 `Set` capped at this number of keys, rather than in a fingerprint table.
 	 *
 	 * The cap is what the fingerprint table replaces.
 	 * A run that sets this reproduces the pre-2026-09-28 behavior.

package/lib/sk/tools/fetch/inspire.ts

@@ -237,7 +237,7 @@ export interface HarvestSKInspireTypeOptions {
 	 */
 	pageSize?: number
 	/**
-	 * Stop once the manifest holds this many pages, for a probe rather than a full harvest.
+	 * Stop once the manifest holds this number of pages, for a probe rather than a full harvest.
 	 */
 	maxPages?: number
 	signal?: AbortSignal
@@ -314,7 +314,7 @@ export interface FetchSKInspireOptions extends BaseFetchOptions {
 	 */
 	pageSize?: number
 	/**
-	 * Stop each type once its manifest holds this many pages, for a probe rather than a full harvest.
+	 * Stop each type once its manifest holds this number of pages, for a probe rather than a full harvest.
 	 */
 	maxPages?: number
 	signal?: AbortSignal
@@ -339,7 +339,7 @@ export function skInspireInputPath(outRoot: BaseFetchOptions["outRoot"]): PathBu
  * for `mailwoman corpus fetch sk-inspire`.
  *
  * Each type is reported separately and a type that fails does not stop the rest, because the
- * documents are read together by the adapter and a partial directory is still worth resuming.
+ * documents are read together by the adapter and a resumed run completes a partial directory.
  * Re-runnable per type: a complete harvest whose file still hashes to its manifest's digest
  * makes no request, and an interrupted one resumes at the page after the last that reached disk.
  */

package/lib/source-register/effective-manifest.ts

@@ -54,7 +54,8 @@ export function epochMixtureAuditPath(configPath: string): PathBuilder {
  */
 export const ExclusionReason = {
 	/**
-	 * The config's `source_weights` map does not name the source, so the sampler never considers it.
+	 * Neither the config's `source_weights` map nor its `source_reps` map lists the source,
+	 * so the sampler never considers it.
 	 */
 	UnweightedSource: "unweighted-source",
 
@@ -87,7 +88,8 @@ export interface EffectiveSourceRecord {
 	license: string
 
 	/**
-	 * The weight `source_weights` gives it, or `null` where the map omits it.
+	 * The weight `source_weights` gives it, the repetition count `source_reps` gives it,
+	 * or `null` where both maps omit it.
 	 */
 	weight: number | null
 
@@ -278,7 +280,7 @@ export interface EffectiveTrainingManifest {
 	 * Sources the epoch emitted that the corpus's frozen manifest does not name, with their emitted rows.
 	 *
 	 * `freezeTrainingManifest` runs inside `buildCorpus` and records the base build's sources.
-	 * An overlay is merged into the corpus afterwards by `overlay-manifest`,
+	 * A corpus overlay is merged into the corpus afterwards by `overlay-manifest`,
 	 * so its rows are in the corpus and outside that manifest.
 	 *
 	 * Measured over `v0.6.0-register-surface` on 2026-09-28: the frozen manifest names 11 sources
@@ -329,7 +331,7 @@ export interface EpochMixtureAudit {
 		/**
 		 * Per-source detail the audit computes from the corpus files rather than from the corpus manifest.
 		 *
-		 * `rows` is how many rows of that source the corpus holds.
+		 * `rows` is the number of rows of that source the corpus holds.
 		 * It is the only place a source outside the frozen manifest has a corpus row count, so it is what
 		 * lets a reader read a source's rows in the corpus beside the rows one epoch drew and emitted.
 		 */
@@ -348,7 +350,11 @@ export interface EffectiveConfigView {
 }
 
 /**
- * Read `data.source_weights` and `data.augment_exclude_sources` from a training config.
+ * Read `data.source_weights`, `data.source_reps` and `data.augment_exclude_sources` from a training config.
+ *
+ * `source_weights` and `source_reps` both admit a source to the sampler, one as a share
+ * and one as a repetition count, so an entry in either block counts as the source's weight here.
+ * A source in neither block is one the sampler never considers.
  *
  * A line scanner instead of a YAML parse.
  * `corpus audit` already reads these blocks this way.
@@ -368,7 +374,7 @@ export function readConfigView(text: string): EffectiveConfigView {
 	const sourceWeights: Record<string, number> = {}
 	const augmentExcludeSources: string[] = []
 
-	type ConfigField = "source_weights" | "augment_exclude_sources"
+	type ConfigField = "source_weights" | "source_reps" | "augment_exclude_sources"
 
 	let field: ConfigField | null = null
 	let fieldIndent = -1
@@ -385,7 +391,7 @@ export function readConfigView(text: string): EffectiveConfigView {
 		}
 
 		if (!field) {
-			const opener = /^([\t ]*)(source_weights|augment_exclude_sources):\s*(#.*)?$/.exec(raw)
+			const opener = /^([\t ]*)(source_weights|source_reps|augment_exclude_sources):\s*(#.*)?$/.exec(raw)
 
 			if (opener) {
 				field = opener[2] as ConfigField
@@ -399,7 +405,7 @@ export function readConfigView(text: string): EffectiveConfigView {
 			continue
 		}
 
-		if (field === "source_weights") {
+		if (field === "source_weights" || field === "source_reps") {
 			const entry = /^[\t ]+([\w-]+):\s*([\d.]+)/.exec(raw)
 
 			if (entry) {

package/lib/source-register/training-manifest.ts

@@ -15,6 +15,18 @@ import { stringifyJSON } from "@mailwoman/core/json"
 
 import { LicenseReviewState, type LicenseDecision, type SourceOperation } from "#source-register/types"
 
+/**
+ * The license label recorded for a source whose rows carry no label.
+ *
+ * A recipe can write rows with an empty `license` column.
+ * An empty string in the record would read as a source that carries no terms,
+ * and `summarizeLicense` would find no identifier in it and report no obligation.
+ *
+ * This label is unrecognized by design, so every reader reports the source's
+ * terms as unknown rather than as none.
+ */
+export const NO_LICENSE_RECORDED = "none-recorded-on-rows"
+
 /**
  * What one source contributed to a build.
  *
@@ -30,7 +42,22 @@ export interface TrainingSourceRecord {
 	 * A synthetic row counts under the source it was fanned from.
 	 */
 	rows: number
+	/**
+	 * The license label the source's rows carry.
+	 *
+	 * When the rows carry more than one label, this is the label on most of them
+	 * and {@linkcode licenseValues} lists every label with its rows.
+	 */
 	license: string
+	/**
+	 * Every license label the source's rows carry, with the rows under each, when there is more than one.
+	 *
+	 * A recipe stamps a description of its own output into the column, and one recipe can stamp
+	 * several, so a synthetic source reads several labels where an adapter's source reads one.
+	 * The record keeps the whole set because a reader that saw only {@linkcode license}
+	 * would take the majority label for the terms of every row.
+	 */
+	licenseValues?: Record<string, number>
 	/**
 	 * The register's decision for that license when the build ran, or `null`
 	 * when the register contains no decision for it.
@@ -97,18 +124,19 @@ export function freezeTrainingManifest(input: {
 	corpusVersion: string
 	builtAt: string
 	profile: string
-	rowsBySource: ReadonlyMap<string, { rows: number; license: string }>
+	rowsBySource: ReadonlyMap<string, { rows: number; license: string; licenseValues?: Record<string, number> }>
 	decisionsByLicense: ReadonlyMap<string, LicenseDecision>
 	refused: ReadonlyMap<string, readonly string[]>
 }): TrainingManifest {
 	const sources: TrainingSourceRecord[] = [...input.rowsBySource.entries()]
-		.map(([source, { rows, license }]): TrainingSourceRecord => {
+		.map(([source, { rows, license, licenseValues }]): TrainingSourceRecord => {
 			const decision = input.decisionsByLicense.get(license)
 
 			return {
 				source,
 				rows,
 				license,
+				...(licenseValues ? { licenseValues } : {}),
 				decision: decision
 					? {
 							licenseID: decision.licenseID,

package/lib/source-register/types.ts

@@ -351,8 +351,8 @@ export interface ElectedLicense {
 	 * Creative Commons permits that.
 	 *
 	 * Where that happens, {@linkcode ElectedLicense.spdx} records the one a build filters on
-	 * and this records what the publisher actually said, so a later reviewer reads
-	 * the conflict rather than one reading presented as settled.
+	 * and this records what the publisher said, so a later reviewer reads the conflict
+	 * rather than one reading presented as settled.
 	 *
 	 * The engineering policy elects the strictest grant for what a released corpus may include.
 	 * It does not establish which grant governs, and `electedBecause` states

package/lib/source-register.ts

@@ -41,6 +41,7 @@ export {
 export {
 	auditTrainingManifest,
 	freezeTrainingManifest,
+	NO_LICENSE_RECORDED,
 	sourcesNotPermitting,
 	trainingManifestDigest,
 } from "#source-register/training-manifest"