Parser-improvement backlog (2026-05-30)

Scoped from the three-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. capability evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. (libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. + 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.-perturbation

• postal-standards). Those arenasarenaA 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. — unlike our own PeliasPeliasAn open-source geocoder, Mailwoman's spiritual predecessor.-derived 376-assertion suite — surface where the neural parser actually fails. This doc turns those failures into a prioritized backlog, categorized by fix mechanism, because the mechanism (not the symptom) determines cost and risk.

The capability picture

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.	what it tests	v0	neural	takeaway
libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it.	clean/canonical	29%	rules-favoured	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. + rules win on their turf
perturbation	noisy/degraded	39%	64%	neural is the robustness 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.
postal-standards	edge formats	26%	TBD@v0.7.2	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. gaps on military/PO-box/rural-route

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 the robustness 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., not a worse v0. The backlog below sharpens that 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. where the arenasarenaA 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. show it bleeding.

The real neural failures (from arena spot-checks)

1. Drops secondary unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. — … Apt 456 → no unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 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.; postal 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. edge class 0% neural.
2. Over-spansspanA 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. streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. boundaries — Stone Way North Seattle → street: "Stone Way North Seattle" (the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. bleeds 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.).
3. Confuses 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. — UK labeled regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.; 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. names dropped or mis-tagged.
4. County → localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. — UK/IE counties (Co. Mayo, Derbys) land as localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. instead of regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality..
5. LocalelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. sparsity — AU/NZ/GB/IE canonical underperform US (the data-imbalance constraint, already on the 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. list).

Categorize by fix mechanism

The cheapest, lowest-risk lever is a deterministic post-decode repair pass (the postcodepostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.-repair #35 family): detect a rigid surface shape with a regex, snap/add the BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components. after decode, before tree-build. The modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is untouched, the pass is opt-in, and 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. guards keep it from regressing a confident 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.. 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.-repair earned default-on at a measured +135/0. Not every failure fits that mould — some need 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. (retrain) or a boundary 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.'.

#	failure	mechanism	cost	risk	status
B1	drops unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.	post-decode repair	low	low (opt-in)	✅ built (unit-repair.ts)
B2	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.	post-decode repair (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)	low	low (closed set)	scoped
B3	county→localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.	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. (lexicon or synth)	med	med	scoped
B4	streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. over-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.	decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs./boundary (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.-clip or morphology-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. anchor)	high	med-high	scoped
B5	localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. sparsity	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. (real EU/Oceania data, #40)	high	low	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.-listed

Why unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. fit the repair mould and streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-overspan doesn't: a unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. has a self-announcing shape (Apt/Ste/Unit/# + identifier). A 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 is defined by what comes after it — there's no local shape that says "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. ends here," so a regex can't draw the line. B4 needs either a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 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. to clip the tail or 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.-morphology FSTmorphology FSTA finite-state transducer encoding street-typing affixes (Avenue, rue, Calle) from libpostal dictionaries, for morphological street recognition. to anchor the 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.; both are real work and belong after the cheap repairs land.

B1 — unit-repair (BUILT)

neural/unit-repair.ts + unit-repair.test.ts, wired as opt-in ParseOpts.unitRepair and harness --unit-repair. Mirrors 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.-repair:

• Detect explicit designatorsdesignatorThe closed-vocabulary leading word of a secondary-address phrase — 'Apt', 'Suite', 'Floor', 'PO Box', 'Level' — paired with an identifier to form a complete subpremise. (Apt, Apartment, Ste, Suite, Unit, Rm, Room, Fl/Floor, Bldg, Dept, Lot, Flat, PH, …) + an identifier (4B, 12, single letter STE D), plus bare #104.
• ADD a unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 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. over O or a geographic-container tag (locality/dependent_locality — see the v0.7.2 finding below); never over house_number / streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. / 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. / 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. / 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. / 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.. SNAP an existing unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 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. to the full shape. Local smear-clip on the flanks.
• 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. guards: word-boundary after 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. (so Unit ≠ inside United, Fl ≠ Florida); excludes Box (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.), bare F/No, Space/Stop (common words); single-letter ident only fires on a standalone 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..

v0.7.2 measured result. The 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. re-run showed the unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-drop has TWO failure modes, not one:

1. Mislabel-as-localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. — bare 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.-led unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. (Flat 2 14 Smith St, APT 2 …) → 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.' 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 whole Designator N run as locality. The original ADD-over-O-only guard correctly refused to fire (the tokenstokenOne 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. weren't O). Allowing ADD over locality/dependent_locality (an explicit 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. is a high-confidence unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. shape) fixes these: postal unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-tag 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. 0/8 → 2/8, 0 localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. regressions (Flat 2, APT 2 reclaimed; F 2 skipped — bare F is deliberately excluded as too greedy).
2. Absorbed-into-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. — STE 12 / STE D mid- or post-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. (MAIN ST NW STE 12) gets swallowed into the street 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.. A regex repair cannot safely carve a unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. out of a confident streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 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. — that's a coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present. gap (unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. underrepresented in 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 same lesson as 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.. The fix is a unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-bearing synth shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. (a B-list item), not a post-decode pass.

Netneural networkA model made of layers of simple numeric units whose connection strengths (weights) are learned from data. The transformer encoder at Mailwoman's core is a neural network.: the repair patches the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.-mislabel half cleanly and safely; 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.-absorption half needs 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.. Address-level pass on those cases doesn't flip (they also fail on house_number/streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.), so unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-repair's value is measured on unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-tag 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., not address-level pass. Stays opt-in; promote to default-on once a golden-set unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-tag delta confirms the +N/0 holds there too.

B2 — country-repair (scoped, not built)

Same mould as B1. A closed lexicon of 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. names/codes (UK, United Kingdom, USA, U.S.A., Canada, Australia, New Zealand, Ireland, GB, …) → force B-country when 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.' tagged it regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality./localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy./O at the tail of the address (position guard: only the trailing run, so Washington the citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. isn't clobbered). Closed set + position guard = low risk. v0.7.2 motivates it strongly: trailing AUSTRALIA → locality, and country 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. −0.8pp in the gate. Build next.

Adjacent finding (new): the v0.7.2 postal v0-only cluster shows a related failure — last-line-only addresses (NEW YORK NY 10025, CANBERRA ACT 2614, SYDNEY NSW) mislabel the leading localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. as streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. when 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. is present. That's neither unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. nor countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head.; it's a coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present. gap (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.' rarely sees streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-less addresses). Folds into the same 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. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. as B5.

Sequencing

1. ✅ B1 unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-repair — built + measured (+2/8 postal unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-tag, 0 regressions, localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.-mislabel half). StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-absorption half is 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. (see B3'). Opt-in.
2. B2 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.-repair — now well-motivated by v0.7.2; build next (post-decode).
3. B3' unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. + streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-less 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. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. — synth unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. mid/post-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. (STE 12) and streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-less last-lines; folds with B5 real-data 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.. This is the half post-decode repair can't reach.
4. B4 streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-overspan — the one true 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.'/decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs. change; its own investigation.

All deltas measured through scripts/eval/external-arenas.sh against the v0.7.2 modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' — the same three-bucket harness that surfaced the failures.