Direction E — the Geographic Rule Engine

Date: 2026-06-05 Status: design, signed off (operator + DeepSeek 3-turn consult) Provenance: Japan cheap-probe (scripts/diag-japan-parse.ts) + DeepSeek consult (.agents/skills/deepseek-consult/session-notes-2026-06-05-japan-resolver-architecture.md) Sibling: Direction D / epic #239 — that one is the parser (anchor-based parsingaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates.); this one is the resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. (what happens after the parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates.).

Where this came from

We went looking at Japan to learn what the non-Latin story costs us, and the probe split the problem in two so cleanly that the whole design falls out of it. Run a real Japanese address through the current v0.7.2 universal parser and you see two failures that are nothing alike:

• (A) The modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is out-of-distribution. Romaji in Western order parsesaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. almost correctly (1-1-1 Ginza, Chuo-ku, Tokyo 104-0061 → housenumber / streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. / localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. / postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one., all roughly right). Native kanji scatters the labelscomponent tagOne of the 33 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag.. This is a coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present. problem — trainable — and in the meantime the postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. is a clean regex hit in _every sample, so coordinate-first resolution carries Japan exactly the way it carried out-of-distribution Dutch to 94.9%. The coarse win needs no parser change at all.
• (B) The schema cannot represent the address. There is no tag for 丁目/番/号 — chōmechōmeIn Japanese addressing, a district-level subdivision in the block-based chōme / banchi / gō numbering scheme, which uses area-and-block numbers instead of street names./banchibanchiIn Japanese addressing, a block within a chōme. The middle level of the chōme / banchi / gō hierarchy./gōgōIn Japanese addressing, a building or lot number within a banchi — the finest level of the chōme / banchi / gō block-based hierarchy., the district/block/building coordinates that most of Japan uses instead of streetsstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.. The modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' jams them into house_number/street/postcode because those are the only slots that exist. A perfectly trained modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' still couldn't emit a correct structure. TrainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. can never fix this; it's a vocabularyvocabularyThe fixed set of tokens a tokenizer can produce. Mailwoman's SentencePiece vocabulary is tens of thousands of subword pieces, with byte fallback for anything outside it. gap.

That distinction is the foundation of everything below. (A) lives in the resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon.. (B) is the one thing that has to change in the parser — and we're going to change it the smallest way that works.

The thesis

The machinery after the parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. is not one monolithic resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon., and it is not a Matryoshka of nested resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. objects either (orthogonal concerns — grid morphology vs postal schema vs alias lists — don't fit a single inheritance tree, and adding a citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. shouldn't mean a code deploy). It's a flat Geographic Rule Engine:

A convention table keyed by Who's-On-First polygon id, deep-merged up the ancestor chainparent chainThe sequence of administrative parents above a place — Springfield → Sangamon County → Illinois → United States. Used to check the geographic coherence of a parse. (Sapporo's row inherits Hokkaidō's inherits Japan's), that dispatches a short ordered sequence of named strategy functions. The data carries the routing and parametersparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values.; the strategies are a small registry of code primitives. "Plug in and out as needed" means a row in a table plus a pick from existing primitives — new code only for a novel spatial system.

Two things fall out for free:

• The dispatch chicken-and-egg dissolves. To pick Sapporo's rules you must know you're in Sapporo — but postcode → WOF locality → ancestors already hands you the entire admin chain in the coarse pass we run today. Convention selection rides on it at zero cost.
• The two audiences are one pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. at two depths. Our need (postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. → regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality./localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy., for trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. gold and coarse geocode) and a user's need (a specific address down to a building) are the same machine stopped at different depths, not two systemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed.. The 2025 Digital Address code is just an early-exit pure-lookup strategy at the front.

This also retires an older open question. Direction D imagined a coarse-placercoarse-placerA lightweight int8 country classifier (~0.79 MB) that predicts which of a set of target countries an address belongs to, feeding a soft prior into resolver disambiguation. that decides which localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' to reach for (#244/#245). With the rule engine, one universal parser plus a geography-keyed convention table replaces the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' router for every Latin localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. and every localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. the postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. can anchor. We mostly don't need to reach for a different modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.'.

The one foundational decision

The seductive idea is to throw out the semantic tags and have the parser emit positional levels (L0…Ln) that the convention names per localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.. It's elegant and would serve every future countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head.. It is also the highest-regret move on the board: it detonates four working localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for., forces a from-scratch retrain plus a resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon.-contract rewrite, with NL's 94.9% to re-earn from zero — and it breaks on partial queries, where the level index of a bare "Berlin" is undefined until you've already resolved the countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head.. We are not doing that, and PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 3 below exists only to record the tripwire that would ever reopen it.

The additive cut that buys ~80% of the value with zero regression risk: keep every existing semantic tag, and add exactly one new multi-valued tag — locator[] — for sub-localitydependent localityA sub-locality (neighbourhood or borough) hierarchically inside a larger locality — e.g. Brooklyn within New York City. Provides finer geographic specification below the primary locality. spatial unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. that aren't a Western streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. + number (chōmechōmeIn Japanese addressing, a district-level subdivision in the block-based chōme / banchi / gō numbering scheme, which uses area-and-block numbers instead of street names./banchibanchiIn Japanese addressing, a block within a chōme. The middle level of the chōme / banchi / gō hierarchy./gōgōIn Japanese addressing, a building or lot number within a banchi — the finest level of the chōme / banchi / gō block-based hierarchy., grid coordinates, block numbers). StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-based countries never emit it; block-based countries never emit street; the convention maps locator[] per localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.. DE/FR/GB/NL never reference it, so they cannot regress.

Make it less special: Japan, South Korea, Taiwan together

Building only for Japan would bake Japanese quirks into something we'd call "general." So the design is validated against three CJK block-style systemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed. at once — Japan (chōmechōmeIn Japanese addressing, a district-level subdivision in the block-based chōme / banchi / gō numbering scheme, which uses area-and-block numbers instead of street names./banchibanchiIn Japanese addressing, a block within a chōme. The middle level of the chōme / banchi / gō hierarchy./gōgōIn Japanese addressing, a building or lot number within a banchi — the finest level of the chōme / banchi / gō block-based hierarchy.), South Korea (도로명 road-name and 지번 lot systemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed.), Taiwan (段/巷/弄/號). If the convention table + locator[] + the strategy registry express all three without per-countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head. code beyond data, the abstraction has earned the word "general."

Update (2026-06-05) — the CJK coarse build is a name-match, not PIP. PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary).-1 implementation surfaced a structural finding (confirmed across JP/KR/TW): WOFWOF (Who's On First). An open-source gazetteer of places maintained by Mapzen/whosonfirst. Mailwoman builds a custom SQLite database from WOF GeoJSON repos, extended with postcode data, importance scores, and coincident-role relations. admin geometry is point-based at the municipalitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. level — there are no municipalitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. polygons — so the European point-in-polygon coordinate-first build is inapplicable to CJK (it gives ~25% JP coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present.). CJK resolves by authoritative name-match instead: postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. → national-postal-authority municipalitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. NAME (JP = KENALL) + GeoNamesGeoNamesA free global gazetteer combining administrative, postal, and POI data across 200+ countries. Supplements Who's On First for postcode centroids and places where WOF has gaps. point → cross-placetypeplacetypeThe Who's On First hierarchical classification of places: planet → continent → country → region → county → locality → neighbourhood. The resolver uses placetype to rank candidates — an exact locality match outranks a county-level match. match (locality+county+localadmin+borough) → WOFWOF (Who's On First). An open-source gazetteer of places maintained by Mapzen/whosonfirst. Mailwoman builds a custom SQLite database from WOF GeoJSON repos, extended with postcode data, importance scores, and coincident-role relations. id, written to the _same postcode_locality table so the same postcode_area_resolution strategy consumes it. This is the convention engine earning its keep exactly as designed — a different build for a different data shape, feeding one unchanged resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon.. Japan: 94.9% built, 93.9% resolved (independent gold). PIP-containment is replaced by name-agreement as the CJK metric. Full write-up: docs/articles/evals/2026-06-05-cjk-arena.md.

Sequencing (reversible-first)

PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 — resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon.-side, additive, reversible, no parser change. This is largely reshaping what we already have — addressing-conventions.json is a primitive convention table, and coordinate-first is postcode_area_resolution — into the general form, then adding rows for JP/KR/TW. Delivers coarse localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. geocodinggeocodingThe process of converting an address into geographic coordinates (latitude and longitude). Mailwoman geocodes in a multi-tier cascade: exact address-point match → street interpolation → locality centroid. Each tier is progressively coarser but more widely available. off the postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. alone, the same trick that got Dutch-OOD NL to 94.9%. Pure resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. work, the part of the stack that has never bitten us.

PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2 — parser-side, additive, gated. Teach the parser to emit locator[] from a handful of JP/KR/TW examples. Existing tags untouched, modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' rollback-able. This inherits every retrain fragility we've hit (the German end-of-string collapse, the PR3 negative), so it's gated on per-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.-F1 + resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. utility exactly like every other retrain, and PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 must prove the engine first.

PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 3 — deferred, decision-only. Revisit L0…Ln only if onboarding more systemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed. (India, China, …) proves the additive scheme insufficient — by then with real data and a mature engine, as a migration rather than a leap.

The coarse/fine split maps onto a safe/risky split: PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 is cheap and near-certain; building-level precisionprecisionOf the spans the model labeled as a given tag, the fraction it got right. High precision means few false positives. Paired with recall to compute F1. (PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2+) is a separate, harder bet, and we may live on PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 for a good while.

The datasets, as strategy primitives

Heterogeneous authoritative sources sit behind the uniform strategy interface; the resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. never cares which source answered. Built from source, never a prebuilt dump — same discipline as every other table we ship.

Dataset	Primitive	TiertierInternal versioning of which label classes the model emits. Tier 1 is the coarse components (country, region, locality, postcode); Tier 2 adds venue, street, house_number; Tier 3 (future) would add attention, po_box, and POI venue subtyping. Historically called 'Stage 1/2/3' before the runtime-pipeline naming made that ambiguous.
KEN_ALL (JP postal authority: 7-digit ↔ prefecture/municipalitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy./district)	postcode_area_resolution (+ block_area_lookup once block polygons are georeferenced)	coarse (+ fine)
JIGYOSYO (JP large-org / PO-box codes)	org_code_lookup	fine
Digital Address 2025 (JP 7-char code → registered building)	digital_code_lookup (early-exit)	fine
도로명주소 / 지번 (KR)	postcode_area_resolution + named_linear_feature_with_offset	coarse + fine
TW postal + 段/巷/弄/號	postcode_area_resolution + block_area_lookup	coarse + fine
WOFWOF (Who's On First). An open-source gazetteer of places maintained by Mapzen/whosonfirst. Mailwoman builds a custom SQLite database from WOF GeoJSON repos, extended with postcode data, importance scores, and coincident-role relations. admin polygons (already shipped)	ancestor walk for convention selection	all

Seed strategy registry: postcode_area_resolution, block_area_lookup, intersection_resolver, named_linear_feature_with_offset, alias_normalisation, transliterate_reorder (pre-parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates.), digital_code_lookup, org_code_lookup, fallback_fuzzy_name_match.

Pre-registered decision rules

• PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 promotion: a localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. ships when postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.-route PIP-containment clears ~85% on a real OA sample, with no regression to DE/FR/GB/NL (the existing coordinate-first numbers are byte-stable).
• PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2 gate: the locator[] fine-tune promotes only on per-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.-F1 + resolverresolverThe component that converts parsed address components (locality, region, postcode) into coordinates by looking them up in the gazetteer. The resolver ranks candidates by name match, population, and proximity, and returns the best-matching place with its centroid or polygon. utility, and never if it drops an existing localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. >2pp. Default to not promoting; live on PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1's coarse win.
• The locator[] shape, the first-ship scope, and the convention-asset format are open forks to settle at PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 kickoff (plain ordered slot vs typed; JP-only branch vs general refactor; WOFWOF (Who's On First). An open-source gazetteer of places maintained by Mapzen/whosonfirst. Mailwoman builds a custom SQLite database from WOF GeoJSON repos, extended with postcode data, importance scores, and coincident-role relations.-id-keyed read-only sqlite asset like postcode-locality-intl.db).
• PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 3 tripwire (the only thing that reopens L0…Ln): the additive locator[] scheme fails to express a newly-onboarded system as data and forces per-countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head. parser-output post-processing in ≥2 localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for..