Releases & capabilities

Two version series exist and used to be easy to confuse:

• npm version — what npm install gives you. All workspaces release in lockstep (e.g. mailwoman@4.1.0 pairs with @mailwoman/neural-weights-en-us@4.1.0).
• 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. series (v0.x.y) — internal 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.' lineage used by configs, the evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. ledger, and the docs. A 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. run only becomes an npm release if it passes its pre-registered gates.

Since 4.0.0 the npm version is the one that matters to consumers; the shipped weightsparameterA 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.' model-card.json records the exact 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. lineage, corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus., and tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses.. The per-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.' score ledger evals/scores-by-version.json (schema: eval-ledger.schema.json) was frozen at v4.4.0 and revived at the 2026-07-02 re-anchor (#885): the v5.0.0 row carries the full re-score, appends are automated (mailwoman eval ledger-append, invoked from the promotion gategated promotionA release discipline where a model ships to production only if it passes pre-registered metrics (e.g. 'DE locality ≥ 70%, no tag regresses > 2pp'). Failing models can be uploaded as experiments but not promoted to default.'s PASS output), and the 4.5–4.16 gap is documented in the row notes — those versions' headlines live in the matrix below + their model-card.json.

The matrix

npm	Date	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.' lineage	What it added	Per-tag truth
7.8.1 (current) · 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.' 6.6.3	2026-07-24	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.' bytes unchanged (v385). Two decode fixes: the GB pair-index re-cut with transitionBeta: 5 — a position-scoped ViterbiViterbi decodingA dynamic programming algorithm that finds the most likely sequence of hidden states (labels) given a sequence of observations (token emissions). Mailwoman uses Viterbi over a linear-chain CRF to produce globally coherent BIO label sequences from per-token model scores. transition bonus at pair-hit boundaries fixing the comma-free path-fusion class (emission biasadditive biasA logit adjustment added before softmax from a prior or external knowledge source, typically soft-capped so it influences but never overrides the model's prediction. The mechanism behind shallow fusion. won per-tokentokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words., the global path fused 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. anyway), with title-preposition suppression ({at, of}) keeping venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. names ("New Inn at Hoff") clean — and the GB 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. conventions row opening the designed snap-repair (#1275/#1290).	GB comma-free dependent_locality 55→63/69 (95.2% correct); GB 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. exact 26→83, clips 44→0; venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. floors and every other surface byte-identical (NZ, golden, presets). The v1 battery caught a 1-row venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. regression and the suppression healed it — pre-registration working as designed.	placetype-pair scorecard
7.8.0 · 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.' 6.6.2	2026-07-24	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.' bytes unchanged (v385). 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. two for the pair-prior pattern: @mailwoman/neural-weights-en-nz first publish (pair-index-nz.bin — LINZ census, 3,134 pairs, δ=10) + the identical-adjacent-segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. rule in @mailwoman/neural (NZ's repeated-name convention — "Mangawhai, Mangawhai" — read as registry evidence via the (x,x) identity pair; GB structurally inert, zero identity pairs).	NZ dependent_locality live: 246/246 emit + tag-correct as-written (100%), 244/246 comma-free, zero golden FPs, GB surfaces byte-exact. Same recipe as GB, zero 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. — register → δ-sweep → pre-registered battery.	placetype-pair scorecard
7.7.0 · 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.' 6.6.1	2026-07-24	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.' bytes unchanged (v385) — 6.6.1 is a decode-artifact bundle revision. pair-index-gb.bin re-cut at δ=10 (the Option-A sweep: v385's dead-classifier deficit is uniform, δ=10 saturates the board), and @mailwoman/neural gains the auto probe chain: segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. path on comma'd input (byte-identical to 7.6.0 behavior), the new anchored-adjacent path on comma-free input (parent window pinned to the post-town position, child adjacent, left-maximal).	GB dependent_locality goes live on the shipped 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.' — 69/69 emit + tag-correct on the golden board, 50/69 comma-free (94% correct), venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label.-confound floors 3.338% (comma'd) / 0.062% (comma-free), zero golden FPs, zero new invariance violations, non-GB 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. byte-identical. Zero 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. — the whole capability is registry evidence at decode time.	placetype-pair scorecard
7.6.0 · 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.' 6.6.0	2026-07-23	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.' unchanged (v385 / 6.6.0). Code + packaging: @mailwoman/neural-weights-en-gb first publish (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.-gb.bin + the 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.-pair index pair-index-gb.bin, δ=5.0 — inert until a dep-loc 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.' ships), the 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.-pair decode prior + invariance evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. suite, five production parser fixes (tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. byte-fallback offsets incl. per-character CJK run splitting, two word-grouping drop classes, two formatter slot-rendering classes), pair-index 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.-gate test, fileMD5 dedup. First release through the PR-based two-phasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). publish flow (prepare → auto-merged release PR → publish; the Production Integrity ruleset retired the direct release-it push).	Ships the en-GB soft-feed surface + the decode-prior machinery npm-side; GB 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. gain 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. anchoring out of the box.	placetype-pair scorecard
7.5.0	2026-07-20	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.' unchanged (v385 / 6.6.0). mailwoman doctor out-of-box diagnostic; POIpoint of interest (POI). A named place that is not strictly an address — landmark, transit stop, venue, amenity, or franchise. Mailwoman tags these as venue and resolves them through the gazetteer. sparse-category ring-reach fix; react browser-mode CI leg; dependabot group re-apply + adm-zip resolution; link-dev-weightsparameterA 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. repoint to the v385 int8.	Doctor + POIpoint of interest (POI). A named place that is not strictly an address — landmark, transit stop, venue, amenity, or franchise. Mailwoman tags these as venue and resolves them through the gazetteer. polish.	—
7.4.0	2026-07-20	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.' unchanged (v385 / 6.6.0). @mailwoman/react (explorers → headless hooks + Storybook/Vitest-browser harness), @mailwoman/fastify plugin, ghcr.io/sister-software/mailwoman Docker API image.	Three new integration surfaces: React, Fastify, Docker.	—
7.3.0	2026-07-20	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.' unchanged (v385 / 6.6.0). Full Overture POI taxonomy snapshot (~2k categories) + curated overlay; brand-search indexed fetch; regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-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. coherencecoherenceThe property of a parse whose resolved places form a consistent geographic hierarchy — the resolved locality really does sit inside the resolved region. (Montreal QC → Québec); MCP bin smoke leg; layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6.-contract reference.	POIpoint of interest (POI). A named place that is not strictly an address — landmark, transit stop, venue, amenity, or franchise. Mailwoman tags these as venue and resolves them through the gazetteer. search grows up; the Montreal-class namesake fix.	—
7.2.1	2026-07-19	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.' unchanged (v385 / 6.6.0). Hotfix: publish exports map derived at pack time (publishConfig.exports retired) — repairs the fully-broken 7.2.0 package resolution.	Install 7.2.1, never 7.2.0.	—
7.2.0	2026-07-19	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.' unchanged (v385 / 6.6.0). POIpoint of interest (POI). A named place that is not strictly an address — landmark, transit stop, venue, amenity, or franchise. Mailwoman tags these as venue and resolves them through the gazetteer. layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6. at full scale (13.68M rows, streamed builds) + docs live POIpoint of interest (POI). A named place that is not strictly an address — landmark, transit stop, venue, amenity, or franchise. Mailwoman tags these as venue and resolves them through the gazetteer. tester. ⚠ Broken package resolution (the publishConfig.exports removal shipped without its replacement) — superseded by 7.2.1 within hours.	Superseded — see 7.2.1.	—
7.1.0 · 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.' 6.6.0	2026-07-18	v3.8.5-latam-8k (v385) — fine-tune off v381 (6.5.0), one change: a dedicated overture-latam shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. (CA 299,809 + MX 156,421 Overture-Addresses rows, weightparameterA 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. 6.0) with +CA/+MX in country_weights. TokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses./lexicon/FSTFST (finite-state transducer). A compact automaton that reads an input sequence and emits an output sequence. Mailwoman encodes gazetteer names and street affixes as FSTs for fast prefix matching and prior injection without search overhead./architecture unchanged; weightsparameterA 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.-only.	The first genuinely-new localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. 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.. Mexican addresses go from mediocre to good — MX localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 74.0→82.7 (+8.7, native-order held-out board): 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.' reads Calle Zeverino León 749, 28809 Manzanillo and places Manzanillo correctly instead of guessing. Canada ships free (US-order; v381 already parsed it). Golden 2pp gate PASS (0 tags >2pp below shipped v381; the one nick is 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. −1.6, small-N). Gauntlet 33/33. The v6.0.0-shipped-baseline floors clear except fr.cedex_real (85.6 vs 85.8, −0.2pp, ~1 row, operator-adjudicated noise). BR deferred — Overture rows are 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.-implicit and dilute 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. emission (−5.3pp); the countryAppendFraction intl-render fix lands in v7.2. Overture CA/MX is CDLA-Permissive-2.0.	v7.1.0 CA/MX grade
7.0.0	2026-07-18	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.' unchanged — same v3.8.1-punct-fix-full (v381) weightsparameterA 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. as 6.5.0. This is a package-only major: no retrain, no tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses./lexicon/FSTFST (finite-state transducer). A compact automaton that reads an input sequence and emits an output sequence. Mailwoman encodes gazetteer names and street affixes as FSTs for fast prefix matching and prior injection without search overhead./architecture change	The rules-parser excision — the legacy deterministic rules parser and the @mailwoman/classifiers + core substrate (solver / classification / context / permutate / split) are deleted; the neural 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 now the sole parser. Also folds the #875 acronym-casing renames (16 public identifiers) and removes the deprecated sdk/cli + sdk/test shims. Consumers pinned to @mailwoman/classifiers should stay on @6.x (it works standalone; add zx as a dep). See the v7 migration guide. No capability change vs 6.5.0.	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.' identical to v6.5.0 grade
6.5.0	2026-07-17	v3.8.1-punct-fix-full (v381) — fine-tune off v310 with two folded changes, one variable each: the synth-no-fragment digit-ownership shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. + augment_punct_drop_prob 0.3 (the #1101 delimiter-free augmentation). TokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. v0.9.0-multisplice (73,143), lexicons, FSTFST (finite-state transducer). A compact automaton that reads an input sequence and emits an output sequence. Mailwoman encodes gazetteer names and street affixes as FSTs for fast prefix matching and prior injection without search overhead., architecture unchanged	The #1101 comma-free fix + digit ownership — the first 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 pass the full Gauntlet clean (shipped 6.4.0 fails it). A comma-free 1600 Pennsylvania Ave NW Washington DC absorbed the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. into the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. and lost the rooftoprooftopGeocoding precision at the building or parcel level — coordinates within a few metres — the highest tier of the geocode cascade. Sourced from address-point and situs data. (~2 km, address_point→admin); the punct-drop augmentation closes it (metamorphic INV[comma-drop] 8/10→9/10 held). The digit-ownership shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. moves the NO board bare-street-hn 0.693→0.733, bare-pc guard held 1.000. Golden gate PASS, 0 tags >2 pp below shipped, us.country_homograph +2.3. WeightsparameterA 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.-only. Known open: the FR no-anchor comma-free (Rue du Chevaleret Paris) stays a tracked #1101 xfail — a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. name-index (#30) closes it.	v6.5.0 punct+digit grade
6.4.0	2026-07-16	v3.1.0-fr-fragment (v310, #1143) — init_from v264 + the synth-fr-fragment BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. @ 12.0, one variable. TokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. v0.9.0-multisplice unchanged	The house-number licence — 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.' had learned that a digit licenses the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. reading, so Rue Montmartre with no number read as locality. The BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. fragment shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. teaches the designatordesignatorThe closed-vocabulary leading word of a secondary-address phrase — 'Apt', 'Suite', 'Floor', 'PO Box', 'Level' — paired with an identifier to form a complete subpremise., not the digit. FR fragment board bare-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 0.215→0.715, streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-particle 0.273→0.855, admin-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-homonym 0.087→0.517; the bare-localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. counter-guard held exactly (0.980→0.980) — 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.' learned the distinction rather than flipping its default. WeightsparameterA 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.-only.	v310 fr-fragment verdict
6.3.0	2026-07-15	v2.6.4-country-softguard (v264, #1104) — init_from FINE-TUNE of v263 with ONE variable: country_ambiguous_scale 1.0→0.5, which softens 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. channel's homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. guard; baked into the exported ONNXONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. via a forward-pass buffer (no lexicon/inferenceinferenceRunning the trained model on new input to get predictions, as opposed to training, which produces the model. In Mailwoman that means a small transformer encoder reads an address string and classifies every token — house number, street, locality, region, postcode, and the rest. A Who's On First gazetteer can feed soft location hints into the pass, but the model makes the final call on every label. Where a generative model writes text token by token, Mailwoman's output is a retrieval-augmented token classification: one label per input piece./browser change), tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. unchanged v0.9.0-multisplice (73,143, md5 5c01cdcd…)	The #1104 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. softguard — v263's HARD homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. guard (the country_ambiguous bit fully suppressing 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. surfaces that are also US statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. / cities / common words: Georgia, Jordan, Turkey) over-suppressed 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. broadly. Softening it to 0.5 strictly dominates v263 with no trade: 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.-homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. F1 82.6→85.1, golden 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 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. 89.3→91.1% (the softening helped BOTH distributions — the guard was over-firing), real-postal 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. 3/4 + hallucination 1% HELD, held-out coordinate z-test PASS both localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for., aggregate golden labelcomponent 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.-fails 1101→1098 (−3). The v265 (scale 0.25) probe recovered more of the narrow homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. slice (87.5) but slightly less 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; 0.5 chosen for balance. 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. lexicon + @mailwoman/neural-web unchanged from 6.2.x — the softening rides the same ONNXONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. the 6.2.1 browser 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. wiring already feeds, so the live demo flips to v6.3.0 with no browser change.	v264 country-softguard grade
6.2.1	2026-07-14	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.' unchanged (v263, card 6.2.0)	Browser 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. channel (#1104): wires the dedicated 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. soft-feed channel into @mailwoman/neural-web. web-onnx-runner.infer() now feeds country_features (real when supplied + a zero-fill fallback for the required input) and the loader fetches country-surface-lexicon-v1.json beside model.onnx and passes it to the shared isomorphic classifier — un-breaking in-browser loading of the v263 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.' (whose ONNXONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. declares country_features as required) and enabling the live-demo flip to v6.2.0. Code-only patch.	n/a — code-only; 11/11 neural-web unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. tests + the real-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.' web-onnx-runner.test.ts (WebONNXRunner × v263 via onnxruntimeONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime.-web WASM — the browser code path) pass
6.2.0	2026-07-14	v2.6.3-country-channel (v263, #1104) — init_from FINE-TUNE of v261 that turns ON a dedicated 2-dim [country_surface, country_ambiguous] 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. soft-feed channel (own projection + confidence, immune to near-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. gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture. suppression); tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. unchanged, v0.9.0-multisplice (73,143, md5 5c01cdcd…)	The #1104 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. fix — the atlas channel that closes v261's cosmetic exception. Golden 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. recovers 84.8→89.3% (clears the 88.6% v241 bar; fixes 19 / breaks 9); the real-postal 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. guard holds (recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. 3/4 + hallucination 1% identical to v261, DeepSeek-signed falsifier); held-out coordinate z-test PASSES both localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. (US z −0.16 on 300 FDIC, FR z −0.17 on 300 BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure.); gauntlet regression+metamorphic PASS. It carries one coordinate-INVISIBLE golden regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. per-tag trade (556→580 fails, concentrated on redundant trailing-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. tails like VT 05161, US where the channel out-competes the adjacent regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. abbrev) — operator-approved. 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.-surface lexicon now ships as a soft-feed sibling (country-surface-lexicon-v1.json, requires.country) in the weightsparameterA 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. packages, HF, and R2.	v263 country-channel grade
6.1.0	2026-07-14	v2.6.1-span-boundary-full (v261, #727 stagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone).-1) — init_from FINE-TUNE of the stable v257 fragment-campaign line + a 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.-only spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree.-boundary auxiliary headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. (BCE weightparameterA 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. 0.5, 8k steps, gentle schedule), discarded at export; tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. unchanged, v0.9.0-multisplice (73,143 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated)., md5 5c01cdcd…)	The #727 stagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone).-1 spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree.-boundary promotion: shrinks the regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.→streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. boundary-absorption class — US regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.→streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. flips 5→2 vs v257, parity streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 0.5281 (+13pp vs v241), parity house_number +7pp, parity 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. 0.9861 PASS, gauntlet regression+metamorphic PASS (Dublin bare-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. coordinate pin held). Golden 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. 82.0% vs shipped 88.6% (−6.6pp) trips the 2pp pre-publish gate and ships anyway as a documented cosmetic exception (#1104): a DeepSeek-signed falsifier (session 019f5f2c) shows the drop concentrates entirely on 220 of 224 golden 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.-gold rows that are non-postal 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-hierarchy strings ("United States of America, Wyoming, Лорейн"), not addresses anyone types — on the 4 real-postal rows and 300 real no-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. rows, 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. and hallucination rate are identical across the whole fragment lineage (v241/v257/v261). The permanent fix — a 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.-lexicon soft-feed channel mirroring the gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture. channel — SHIPPED one day later in 6.2.0 (v263, above), recovering golden 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. to 89.3%. Also landed this window: the docs-architecture cleanup (nav restructure + the flagship concept-page series) and the #1104 corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. counterweight/leading-position 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.-row work that fed the v261 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. data.	v261 promotion + gate exception, golden failure report
6.0.0	2026-07-12	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.' unchanged (card 5.9.0)	The Hono API surface cutover: all four HTTP surfaces — libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. (#1075), photon (#1082), nominatim (#1083), and the native mailwoman serve surface via the new @mailwoman/api package (#1084/#1085) — now run on one shared Hono stack, with typed OpenAPI 3.1 schemas and gated client generation (#1086). The express server is deleted in full — mailwoman/server/ (routers + the static PeliasPeliasAn open-source geocoder, Mailwoman's spiritual predecessor.-debug pages), express dropped from package.json, a 333-line yarn.lock collapse — with no compatibility shim; the retired endpoints (/api/parse, /api/resolve, /api/resolve-tree) are not aliased. core/formatter's pre-#599 fragaria stub is also removed, superseded by @mailwoman/formatter.	n/a — code-only major, 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.' unchanged; full workspace test suites green (mailwoman 565 passed / 12 skipped, api 37/37, libpostal 25/25, photon 31/31, nominatim 24/24), smoke-clean-install across 28 packed workspaces, live-smoked graceful-shutdown drain on the compiled CLI
5.10.1	2026-07-10	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.' unchanged (card 5.9.0)	Stranger cold-start fixed (#1009): the drop-in servers install-and-run from npm without manual weightparameterA 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. wrangling, with actionable errors for missing data; streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-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. results decorate streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-grade (#1050); the demo's FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. pinned to the dated 2026-07-10 data release.	gauntlet PASS
5.10.0	2026-07-10	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.' unchanged (card 5.9.0)	The scripts→Pastel arc: every build/evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error./data tool becomes a mailwoman command (corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus., gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture., evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error., dev, registry, placer namespaces) — the scripts/ drawer closes. The public demo serves FR national rooftopsrooftopGeocoding precision at the building or parcel level — coordinates within a few metres — the highest tier of the geocode cascade. Sourced from address-point and situs data. (hosted BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. + streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-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. fallback). PLM arrondissement communes fold into the address-point key on both sides (#1045); 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.-less FR citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. queries reach the national 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. via bbox fall-through; chromium-i18n address metadata refreshed (252 countries).	gauntlet PASS
5.9.0	2026-07-09	v2.4.1-fr-nsplice-ft — init_from fine-tune on the v0.8.0-fr-nsplice spliced tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. (12k steps, ~33 min A100; int8 md5 121162e6…)	The FR diacriticdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. fix, the principled way (#444): isolated é SentencePieceSentencePieceA language-independent subword tokenizer that splits text into pieces using a unigram language model. Mailwoman uses a SentencePiece tokenizer with a 48,000-token vocabulary and byte-fallback, trained on address data rather than general text. piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). were tagged as noise, mangling 23.5% of accented FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. keys. The 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.-free vocabvocabularyThe 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. splice was tried first and falsified with mechanism (mean-init inherits the é-piece's trained O-signal; mangle got worse, 26.4% — #1047); the falsification specified this fine-tune: mangle 23.5→17.7%, Honoré + René-class fixed, held-out FR z=+0.90 / US z=+0.29 (holds), non-FR drift ~0. Also: the FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-centroid 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. (#1042, 2.2M streetsstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. derived from the sealed BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. artifact — streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-only voies 0/8→7/8) and house-grade rooftoprooftopGeocoding precision at the building or parcel level — coordinates within a few metres — the highest tier of the geocode cascade. Sourced from address-point and situs data. decoration (#1041, komoot-verbatim place/house schema).	gauntlet --candidate --candidate-tokenizer + held-out z-tests (full table on #444)
5.8.0	2026-07-09	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.' unchanged (card 5.4.0)	The BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. FR address-point 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. (#1012): 26M rooftopsrooftopGeocoding precision at the building or parcel level — coordinates within a few metres — the highest tier of the geocode cascade. Sourced from address-point and situs data. from France's open register as a sealed 6.9 GB artifact — FR @1km 37.4→87.9% in-BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. (independent Nominatim spot-check 13/14 within 200 m), first publish of @mailwoman/ban. Plus the #1023 matchCountry fall-through (the Tbilisi-class foreign-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. namesake fix — caught by the gauntlet pre-dispatch, the gate doing its job) and the #1024 md5-vs-card guard.	gauntlet PASS (+3 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.-namesake cases)
5.7.0	2026-07-08	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.' unchanged (card 5.4.0)	Gazetteer CLI + sealed-artifact discipline (#1027–#1031): the data build path becomes mailwoman gazetteer, artifacts sealed readonly with recipes in-repo; the #1026 corrected admin rebuild restored 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./regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. nodes for ~95 countries flattened by the prior rebuild.	gauntlet PASS
5.6.0	2026-07-07	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.' unchanged (card 5.4.0)	The wire-truth release: forward /api decorates from the resolved place — proper names, 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., osmOpenStreetMap (OSM). A community-curated global map database (ODbL-licensed) with addr:* tagged features and place hierarchies. A secondary corpus source and a source of street names. schema (#1014, real Photon clients stopped crashing); limit returns ranked top-N + lat/lon proximity bias honored (#1016); BE reverse no longer crosses borders (#1015, Overture real bboxes + 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. nodes). Every fix pre-verified through real third-party clients (@openrunner/photon-geocoder, the QGIS French Locator Filter) before release.	gauntlet PASS + real-client integration transcripts
5.5.0	2026-07-06	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.' unchanged (card 5.4.0)	Native CORS on the photon/nominatim/libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. drop-in servers (#1017) — browser map widgets can call self-hosted endpoints without a proxy shim.	gauntlet PASS
5.4.0	2026-07-05	v2.3.0-nl-postcode — init_from FINE-TUNE of v2.2.0-full-family @ 100k + synth-nl-postcode@6.0, the four leading-name-boundary family shardsshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. pinned @12 (int8 md5 ea785a70…; tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. v0.7.1-nsplice unchanged)	The NL digits-first 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. fix (#924): 5.3.0 read the Dutch full-form 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. 1012 LG as house_number(1012)+street(LG), pulling the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. into the US situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. 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. (1012 LG Amsterdam → Amsterdam, NY); this 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. postcode "1012 LG" natively and resolves to the Netherlands (52.37, 4.9). A DeepSeek-signed pinned init_from fine-tune (~0.9 A100-h vs a 3.6 h from-scratch): the "any fine-tune tips the #901 family" lore is surgery-lineage-specific, and v5.3.0 is from-scratch, so it took the pin without tipping the sentinels (a falsification pilot confirmed it, then the full run pinned the family @12). Gains beyond NL: US inverse-risk 9→10/10 (4-digit house numbershouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales. not misread as postcodespostcodeThe 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. — the fine-tune even fixed a pre-existing 6789 El Camino leak); FR bare-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. chevaleret 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. fixed (locality=Paris, street="Rue du Chevaleret" vs 5.3.0's Chevaleret Paris/Rue du fragmentation) — its metamorphic gauntlet rows were reclassified #829 (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. case-norm exposed BY the better 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., not a regression). Gate: 17/17 floors (us.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. 96.2, homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. 85.1, streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 82.1, fr.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. 99.3), fr.bare_street_intact 85% (#949 leg), mask-regression + arenaarenaA standardized test set probing one capability: libpostal (clean canonical), perturb (noisy and degraded), postal (edge formats). Each arena answers a different question about where rule vs neural wins. + gauntlet PASS. Carried int8-homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. exception (delta 2.4 vs the 1.5 cap; fp32fp32 / fp1632-bit and 16-bit floating-point formats. Mailwoman trains in bf16 (a 16-bit variant) and exports the ONNX model in int8 for size. passes all floors, int8 homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. 85.1 clears its 83.3 floor). macro_f1 0.7613→0.7558 (noise). Calibration carried from 5.3.0.	gate v2.3.0-nl-postcode + gauntlet PASS (5 #901 sentinels held)
5.3.0	2026-07-04	v2.2.0-full-family @ 100k — FROM SCRATCH on the nsplice vocabvocabularyThe 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. (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.' md5 a64ad2e6…; tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. v0.7.1-nsplice unchanged from 5.2.0)	The FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-fragmentation fix (#949): 5.2.0 shipped an ungated bare-AND-anchored FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. break (16/40 bare on the standing harness, Allée→e); this reads 34/40 bare / 37/40 anchored, plus CZ wrong-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 13.7→6.6%, SK 13.3→6.6%, SE +4.6pp resolve, Åbo 0.12 km vs 859. US/FR coordinates identical (ni, deconfounded). Three operator-approved trades on the record: SI no-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. +2.7pp wrong-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. (floored 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.'-independently by postalCompoundRecovery, #942), fr.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. floor 99.3→99.2 (gate v5.3.0-family.json), int8 homographhomonymyOne surface, many referents: 'Georgia' is a US state in 'Atlanta, Georgia' and a country in 'Tbilisi, Georgia.' Handled by disambiguation, split across two stages — the parser resolves the tag from in-string context; the resolver late-binds the referent with geographic context. delta 2.4 vs the 1.5 cap (per-artifact exception; fp32fp32 / fp1632-bit and 16-bit floating-point formats. Mailwoman trains in bf16 (a 16-bit variant) and exports the ONNX model in int8 for size. passes all floors). Promoted from the 2026-07-03 archive after the re-gate under the #942 floor — the salvage that replaced a ~4 A100-h rerun (#901).	gate v5.3.0-family + re-gate record (gate-v193/RESULTS.md)
5.2.0	2026-07-03	nsplice-v2-expanded — the v1.9.3a3 encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters). + the SECOND vocabvocabularyThe 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. splice (v0.7.1-nsplice, mean-init surgery; 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.' md5 2c18d854…)	The nordic splice: extends the diacriticdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. piece 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. beyond CZ/PL (the ø/æ/å class the bsplice piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). missed). Gate v5.2.0-nordic; us.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. floor 95→94.8 (measurement-identity revision — shipped 5.1.0 reads the identical 94.8 on the same harness, int8 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. 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. byte-identical on all 2,660 gate rows). Shipped with #949 latent (the FR streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. fragmentation, found 2026-07-03 and fixed by 5.3.0 — no fr.streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. floor existed to catch it).	gate v5.2.0-nordic.json
5.1.0	2026-07-02	v1.9.3a3 + spliced tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. v0.6.0-bsplice (58,582 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated).) + embeddingembeddingA vector of numbers representing a token (or other item) so that similar items sit near each other in vector space. The first thing the model does is turn each token into an embedding. mean-init (bsplice-meaninit; 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.' md5 12fa2d64…, tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. b0397ce3…)	The diacriticdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. splice (#884) — the first coordinated 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.' + tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. bump. 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.-free vocabvocabularyThe 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. splice + mean-init fixes Slavic diacriticdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. fragmentation: wrong-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. CZ 22.4 → 14.8%, PL 27.9 → 7.9%, SK 22.2 → 13.2%, SI 18.9 → 10.6% (n=1k, pre-registered CIs); US byte-identical on ASCII rows, FR n=3000 coordinate non-inferiority PASS on the codepoint-leak rows (the ASCII-only-guarantee finding → #900). int8 33.8 MB (+3.8 MB over the ~30 MB browser SLO, operator-accepted). Demo repointed to npm parity (#896) and runs the shared resolveTree (#861); the demo-only v4.16.0 fr-bare-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. case rides the next retrain (#901).	post-ship battery 17/17 PASS on the published pair (ledger row a640df0d)
5.0.0	2026-07-01	v1.9.3a3 weightsparameterA 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., unchanged from 4.15.0 (bundled model.onnx md5-verified identical)	The acronym-casing major: breaking rename of public identifiers to whole-component acronym casing (createWOFResolver, parseJSON, WOFSqlitePlaceLookup, …). No 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.' or 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. behavior change. Known-incomplete edges (Us/Json families) batched to the next major (#875).	release 3adb655f
4.16.0 – 4.16.2	2026-06-27	weightsparameterA 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. unchanged (dev-weightsparameterA 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. pinned to shipped, #259)	The drop-in API release: @mailwoman/{nominatim,photon,libpostal} compatibility surfaces, the @mailwoman/annotations contract (OpenCage-style blocks), timezone-lookup/un-locode-lookup/nuts-lookup, and the turnkey mailwoman gazetteer build/promote/publish CLI. .1/.2 were same-day packaging patches.	night postmortem
4.15.0	2026-06-25	v1.9.3a3-anchor-absorption @ 80k (v4.14.0 recipe + shaped anchor + counter-shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.)	The #220/#723 anchor-pollution fix: leading-5-digit house numbershouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales. no longer swallowed as postcodespostcodeThe 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. (SLICE-H anchor-ON 20 → 100). Documented gate revision (operator-approved): us.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. floor 97 → 95, fr.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. 99.5 → 99.3 — coordinate-flat on 2000 real OA-US points (VT, the stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. the trade lives in, flat).	v4.15.0 promotion
4.14.0	2026-06-24	v1.9.2-multilocale-au (v192) @ 80k (adds the G-NAFG-NAF (Geocoded National Address File). Australia's authoritative open address register (CC-BY-licensed), used as a training source for Australian addresses. AU shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.)	Australia: AU resolve @25km 65 → 87 on real G-NAFG-NAF (Geocoded National Address File). Australia's authoritative open address register (CC-BY-licensed), used as a training source for Australian addresses. holdout; US/FR/EU held. Fixed 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. publish gap (publishConfig in .release-it.json; new packages need trusted-publishing setup before their first OIDC publish).	night postmortem
4.13.0	2026-06-22	v1.9.1-multilocale-3-order (the #148 order-overfit fix)	The EU multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. release: renders each localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. in 3 natural orders so the parser stops overfittingoverfittingWhen a model memorizes quirks of its training set instead of learning general patterns, so it scores well in training but poorly on new data. Guarded against with held-out evals and regularization. one template → EU resolve +9..+31pp (IT 79 → 92.7, PT 52 → 82, AT 50 → 81.3), US/FR held. Three coordinate-invisible labelcomponent 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.-F1 deltasdelta F1The F1 difference between two model versions for a tag, used to flag a regression or a gain step-over-step. shipped with written justification (the pattern #885 now backstops).	v4.13.0 postmortem
4.12.0	2026-06-21	parser weightsparameterA 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. unchanged; 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. v743b bundled	The EU 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.-constraint release (#743/#194): 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. widened 11 → 27 in-map classes, the 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.-guarded HARD 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. filter default-ON (in-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-or-unresolved), 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. bilingual-alias fold (FI hard-resolve +16.3pp).	night postmortem
4.11.0	2026-06-19	v1.8.0-fr-admin-split @ 40k (from scratch, clones v1.5.0-fr-order)	First to beat v1.5.0 on the shipped assembled coordinatecoord metricThe primary evaluation metric: distance from the resolved coordinate to the true address point. Measured at percentiles (p50, p90) and as 'within X meters.' Prevents the label-F1 trap where a model scores higher on token labels but geocodes worse. (not labelcomponent 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.-F1): the fr-admin-split shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. (real BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. commune+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., localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. split from its département via codex) teaches the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.↔adjacent-admin-tokentokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. split on non-US formats → held-out FR coord p50 42.5 → 2.2 km (−95%), regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-emit 39.6 → 99.6%, regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-correct 36.4 → 96.4%; US coordinate flat (localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 97.5 → 97.6, p50 3.3 = 3.3). Two coordinate-invisible labelcomponent 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. deltas (us.localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. −2.2 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.-only; fr.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. −3.5, 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. given 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.) operator-approved. Bridge retired.	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.'-card v4.11.0 / #728
4.5.0 – 4.10.0	2026-06-12 → 18	(intermediate — rows pending backfill)	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.'-recovery + release-machinery cuts between the parity consolidation and the admin-split: fr.house_number recovery (4.6.0), 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.'↔code version decoupling (4.7.0), the clean-install/hoisting fix + ci:smoke guard (4.8.0). See the per-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.' ledger (evals/scores-by-version.json) for the authoritative lineage.	ledger scores-by-version.json
4.4.0	2026-06-11	v1.3.0-boundary-consolidation @ 40k (from scratch)	The boundary consolidation: po_boxPO boxA numbered mailbox at a post office used as a delivery address instead of a physical street location. Mailwoman tags it as the po_box component; structurally the same family as a subpremise. 0→89.1, cedexCEDEX (Courrier d'Entreprise à Distribution Exceptionnelle). A French postal routing for high-volume business mail: a CEDEX code delivers directly from a sorting centre, bypassing the local post office. A common negative-space format Mailwoman must parse. 0→96.1, intersectionsintersectionAn address that names a location by two crossing streets ('5th & Main') rather than a number and street. Mailwoman tags the two streets as intersection_a and intersection_b — a negative-space format that starved the early model. 0→100 (the last starvedstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed). tags), perturb arenaarenaA standardized test set probing one capability: libpostal (clean canonical), perturb (noisy and degraded), postal (edge formats). Each arena answers a different question about where rule vs neural wins. 64→72 (the #513 debt, gated), FR regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. 16.2→25.6. Ships the spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. bridge (required decode config; dotted surfaces are corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus.-unlabelable). PASS 17/17 on the densest pre-registered spec to date; three-battery verdict record.	v4.4.0 ship gate
4.3.0	2026-06-11	v1.1.0-relabel-consolidation @ 40k (from scratch, clean lineage)	The affix lever closed by data consistency (#511: the base corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus.'s 1,039:1 contradictory streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 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. relabeled — prefix 64.9→93.6, suffix 48.8→96.6) + the conventions layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6. ships its first slice (#478: locale_logits exported, fr conventions maskconventions maskA decode-time constraint layer keyed by the model's own address-system detection (the exported locale head): tags that are ungrammatical in the detected system are removed from the Viterbi vocabulary, and the system's postcode shape arms a snap-only repair pass. The first slice forbids USPS street-affix decomposition for French. Same knowledge-outside-the-weights property as the gazetteer anchor — add a codex conventions row, no retrain. fixes the leading-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. digit-split its own first gate caught). FR 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. 99.7, FR house_number 97.7, unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 92.1, VT honest-evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. flat. Stated notes: FR regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. −11.4 (unfloored tail), 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. leg re-measured un-starvedstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed).. 12/12 + 3 ship legs.	v4.3.0 ship gate
4.2.0	2026-06-10	v1.0.2-consolidation-runB @ 20k (init_from consolidation v1.0.0 @ 40k)	The v1.0 parity consolidation: localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. +12.8, regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. +10.7, 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. 0→89.8 (gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture. soft anchoranchor inferenceA technique where structured knowledge (postcode locations, gazetteer place names) is injected into the model as soft input features — not as deterministic overrides. The model still decides the final labels, but the anchor signal biases it toward correct admin tags.), affixes exist (0→64.9/48.8), FR house_number 94.6 best-ever, DE 90.9. Stated re-baselines: US streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. −2.3, unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. −1.7 (measured 29M ceiling, #492). 4/4 ship gate.	Parity scorecard 2026-06-10
4.1.0	2026-06-09	v0.9.7-unit-v3 @ 20k (off v0.9.3, off v0.7.2)	Secondary-unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 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. (unit 0 → 92.3 real-OOD) — the first negative-space parity win. Int8 quant toolchain pinned + fixed (Safari WebGPU graph stability). Two-backend artifact release (npm + demo pointer).	Parity scorecard 2026-06-09
4.0.0	2026-06-06	v0.9.3-de-regiontail	First fully-synced release across all workspaces (mailwoman installable again, end to end). Multi-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.' (both-order German synth, regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-tail), live 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.-anchor inferenceanchor inferenceA technique where structured knowledge (postcode locations, gazetteer place names) is injected into the model as soft input features — not as deterministic overrides. The model still decides the final labels, but the anchor signal biases it toward correct admin tags., self-conditioningself-conditioningAn auxiliary model head that predicts the input's locale (language/country) and then FiLM-modulates the encoder so parsing adapts to the detected locale. Structural facts like word order still need explicit training on both orders.. New scoped packages published (codex, localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.-gate, resolversresolverThe 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., neural-web).	ledger v0.9.3 rows
2.x / 3.0.0 (weightsparameterA 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.)	2026-05	v0.4.0 (8.87M params, 21 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.)	The Phase-3 integration era: 6-stagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). runtime 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., CRFCRF (Conditional Random Field). A statistical modeling method that predicts structured outputs by modeling dependencies between adjacent labels. Mailwoman uses a linear-chain CRF as the Viterbi decoder at inference time to enforce BIO label consistency — a B-street must be followed by I-street or O, never I-locality. decode, hybrid policy registrypolicy registryThe per-component table that decides which classifier (rule or neural) has authority for each address component. The Ship-of-Theseus dial.. WeightsparameterA 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. versioned independently of the runtime packages (3.0.0 weightsparameterA 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. + 2.1.0 runtime) — the misalignment this page exists to prevent.	ledger 0.1.0–3.0.0 rows

Earlier 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.-series numbering (v0.1.0, v0.2.0, the realigned 3.0.0) predates the lockstep policy; treat the evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. ledger as the record there.

In flight

Series	Target	Status
CZ/PL tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. splice	operator-gated	#884: 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.-free vocabvocabularyThe 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.-splice + mean-init (CZ wrong-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 22.4 → 14.8%, PL 27.9 → 7.9% at n=1k; US byte-identical). Coordinated 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.' + tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. bump; int8 is ~3.8 MB over the browser SLO.
arbitrationarbitrationA pipeline stage that compares rule-based (v0) and neural classifier output, resolving disagreements via a policy registry. Built and merged but not promoted — the coordinate gate showed label-F1 gains came at the cost of worse geocoding. layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6.	backlog	#478: built and merged default-OFF (router #682, coherencecoherenceThe property of a parse whose resolved places form a consistent geographic hierarchy — the resolved locality really does sit inside the resolved region. #683, machinery #684); promotion killed by the coordinate gate (#685).
Commonwealth po_boxPO boxA numbered mailbox at a post office used as a delivery address instead of a physical street location. Mailwoman tags it as the po_box component; structurally the same family as a subpremise. formats	backlog	po_boxPO boxA numbered mailbox at a post office used as a delivery address instead of a physical street location. Mailwoman tags it as the po_box component; structurally the same family as a subpremise./cedexCEDEX (Courrier d'Entreprise à Distribution Exceptionnelle). A French postal routing for high-volume business mail: a CEDEX code delivers directly from a sorting centre, bypassing the local post office. A common negative-space format Mailwoman must parse. shipped in 4.4.0; the residual is AU/NZ-style formats (#517).

Capability quick-reference

Capability	Since
6-stagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). runtime 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. (normalizenormalizeStage 1 of the runtime pipeline: deterministic input preprocessing (Unicode NFC, punctuation normalization, whitespace collapse). Returns a NormalizedInput with an offsetMap that maps normalized positions back to the raw input. → … → resolve)	2.x era
Hybrid rule/neural policy registrypolicy registryThe per-component table that decides which classifier (rule or neural) has authority for each address component. The Ship-of-Theseus dial.	2.x era
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. SQLite 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. (exact-match tieringexact-match tieringA resolver ranking strategy that groups candidates by match quality (exact name/alias match vs partial) and ranks within each tier by secondary signals, so a strong population prior can't promote a poor name match., coordinate-first 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. disambiguation)	4.0.0
Postcode anchoranchor inferenceA technique where structured knowledge (postcode locations, gazetteer place names) is injected into the model as soft input features — not as deterministic overrides. The model still decides the final labels, but the anchor signal biases it toward correct admin tags. at inferenceinferenceRunning the trained model on new input to get predictions, as opposed to training, which produces the model. In Mailwoman that means a small transformer encoder reads an address string and classifies every token — house number, street, locality, region, postcode, and the rest. A Who's On First gazetteer can feed soft location hints into the pass, but the model makes the final call on every label. Where a generative model writes text token by token, Mailwoman's output is a retrieval-augmented token classification: one label per input piece. (PCB1)	4.0.0
Japanese 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 resolution (Geographic Rule Engine)	4.0.0
Isotonic confidence calibrationconfidence calibrationThe process of adjusting model confidence scores so that '0.6' actually means the model is right about 60% of the time. Mailwoman uses isotonic regression (PAVA) to calibrate per-span confidences against held-out data. Applied opt-in via createCalibrator. (per-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. calibration.json in the weights bundlemodel weightsThe learned parameters of the neural classifier, shipped as ONNX files in the @mailwoman/neural-weights-* packages. Weights are locale-specific bundles that include the model, tokenizer, and a model-card.json metadata file.)	4.0.0
Secondary-unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. (unit) recognition	4.1.0
Street affixesstreet affixA modifier on a street name indicating type or direction — Street, Avenue, rue, Calle, N, East. Mailwoman tags these as street_prefix / street_suffix, recognized via a morphology FST., 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. soft anchoranchor inferenceA technique where structured knowledge (postcode locations, gazetteer place names) is injected into the model as soft input features — not as deterministic overrides. The model still decides the final labels, but the anchor signal biases it toward correct admin tags., multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. balance	4.2.0
Affix splits at parity (93.6/96.6), address-system conventions maskconventions maskA decode-time constraint layer keyed by the model's own address-system detection (the exported locale head): tags that are ungrammatical in the detected system are removed from the Viterbi vocabulary, and the system's postcode shape arms a snap-only repair pass. The first slice forbids USPS street-affix decomposition for French. Same knowledge-outside-the-weights property as the gazetteer anchor — add a codex conventions row, no retrain., localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. readable	4.3.0
po_boxPO boxA numbered mailbox at a post office used as a delivery address instead of a physical street location. Mailwoman tags it as the po_box component; structurally the same family as a subpremise./cedexCEDEX (Courrier d'Entreprise à Distribution Exceptionnelle). A French postal routing for high-volume business mail: a CEDEX code delivers directly from a sorting centre, bypassing the local post office. A common negative-space format Mailwoman must parse./intersectionsintersectionAn address that names a location by two crossing streets ('5th & Main') rather than a number and street. Mailwoman tags the two streets as intersection_a and intersection_b — a negative-space format that starved the early model. alive, spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. bridge, arenaarenaA standardized test set probing one capability: libpostal (clean canonical), perturb (noisy and degraded), postal (edge formats). Each arena answers a different question about where rule vs neural wins. floor restored	4.4.0
FR admin-split (localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.↔département), first coordinate-gated promote	4.11.0
EU hard 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. filter + 27-class coarse placer	4.12.0
Multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. 3-order EU 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.' (16 localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.)	4.13.0
Australia (G-NAFG-NAF (Geocoded National Address File). Australia's authoritative open address register (CC-BY-licensed), used as a training source for Australian addresses.), AU @25km 65→87	4.14.0
Leading-house-number anchor fix (#723), shaped postcode anchoranchor inferenceA technique where structured knowledge (postcode locations, gazetteer place names) is injected into the model as soft input features — not as deterministic overrides. The model still decides the final labels, but the anchor signal biases it toward correct admin tags.	4.15.0
Nominatim/Photon/libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. drop-in APIs, annotations, tz/LOCODE/NUTS lookups	4.16.0
Whole-component acronym casing (breaking rename)	5.0.0
Slavic + Nordic diacriticdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. splice (CZ/PL/SK/SI wrong-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. halved)	5.1.0
FR bare/anchored streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. de-fragmentation (34/40 bare, 37/40 anchored)	5.3.0
NL digits-first 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. (1012 LG), US inverse-risk hardening (9→10/10)	5.4.0
SpanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree.-boundary headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. shrinks regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.→streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. ambiguity (#727 stagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone).-1)	6.1.0
Dedicated 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. soft-feed channel — golden 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. recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1. 84.8→89.3% (#1104)	6.2.0
Browser 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. channel wired into @mailwoman/neural-web (#1104)	6.2.1

When a release ships, this page gains a row and the status page gets re-verified — they change together or not at all.