license: agpl-3.0 language:

• en library_name: onnx pipeline_tag: token-classification tags:
• token-classification
• named-entity-recognition
• address-parsing
• postal-address
• geocoding
• onnx
• sequence-labeling

mailwoman — neural address-parser weights (en-gb)

The trained-model bundle (@mailwoman/neural-weights-en-gb) for Mailwoman, a postal-address parser. This package is data only — a model-card.json, the GB postcode-anchor binary, the placetype-pair retrieval index, and the shared gazetteer lexicons. It has no JavaScript logic of its own; it is loaded at inference time by @mailwoman/neural.

This locale is served by the shared multi-locale model. The en-gb bundle ships no model.onnx/tokenizer.model of its own — it declares @mailwoman/neural-weights-en-us as its mailwoman.baseWeights and resolves the base package's model + tokenizer at runtime (byte-identical artifact; one encoder serves both locales). What this package ships is the GB-specific soft-feed data: the outward-code postcode-anchor binary (postcode-gb.bin) built from the WOF GB postcode shard, the placetype-pair retrieval index (pair-index-gb.bin, built from the HM Land Registry PPD tuples — see Evaluation below), plus the shared gazetteer/country lexicons.

What this is

Mailwoman is a calibrated, retrieval-augmented sequence labeler over a microlanguage — coupled to a gazetteer that resolves its output to coordinates. This bundle is the GB-facing half of the retrieval side: the en-us encoder's soft anchor + gazetteer channels, fed with UK-specific data so the model never has to memorize GB postcodes or place names — that knowledge arrives at inference as a retrieval, not a weight.

• Base model: none of its own — see @mailwoman/neural-weights-en-us.
• Postcode anchor: GB is aggregated to the outward code (SW1A, not the full unit SW1A 1AA) — 2.7M unit postcodes is both too large for the browser budget and finer-grained than an anchor needs. A full unit code that misses the exact lookup falls back to its outward code automatically (see @mailwoman/neural's extractPostcodeAnchors).
• Placetype-pair index: a retrieval-augmented dependent_locality prior built from real (child, parent) place-name pairs (e.g. "Fishburn" is a real child of "Stockton-on-Tees") — a soft decode-time bias, hard-gated to GB input only, that never fires for any other locale. See Evaluation.

Intended use

Parsing free-text UK postal addresses into structured components (country, region, locality, dependent_locality, postcode, street, house_number, …) for geocoding — resolving a parsed address to coordinates via a gazetteer/resolver. The model is the parsing front-end of that pipeline, not a standalone geocoder.

Ship-config requirement (read before using)

The Mailwoman model expects the soft anchor + gazetteer channels fed at inference. Running it with those channels off is out-of-distribution and silently collapses the admin tags (country/region/locality/postcode) — an anchor-off metric on an anchor-trained model is systematically misleading. Construct the scorer through @mailwoman/neural's createScorer (the canonical ProductionScorer), which reads the bundle's requires/channel contract and fails closed if a declared channel isn't fed. Do not hand-wire the raw ONNX session with the anchor input zero-filled.

Evaluation

First graded GB story, 2026-07-23. Previously this package was packaging scaffolding only — the runtime resolution path was exercised and tested, but no GB-specific numbers existed. That has changed: the base encoder was fine-tuned on a dependent_locality-feed corpus that includes a real GB shard, and this package's own pair-index-gb.bin supplies a calibrated retrieval prior on top. Full-pipeline dependent_locality recall, GB golden board (69 rows carrying the tag), prior ON at the calibrated δ=5.0: 69/69 emission, 67/69 tag-correct (97.1%). The two misses are pre-existing, independently characterized parser-level cases, not prior artifacts. A three-way ablation shows this recall is carried almost entirely by the prior, not the fine-tuned checkpoint's own classifier row — see model-card.json's notes and eval blocks for the full breakdown, including the δ-calibration sweep, the venue-confound false-positive rate (0.738% at δ=5.0), and the comma-stripped-input trade (fully inert by design in the current probe mode).

The base encoder itself is a STAGED candidate, not yet promoted — see @mailwoman/neural-weights-en-us's model-card.json phase field for the open ship blocker (a Gauntlet metamorphic-layer regression, unrelated to GB). Until that resolves, treat the numbers above as graded-but-unshipped: real measurements against the actual candidate artifact, not yet the production default. docs/articles/evals/ carries the full scorecard once promoted.

Calibration

Mailwoman confidences are isotonic-calibrated (PAVA) against held-out data and applied opt-in via @mailwoman/core's createCalibrator; default parse output is byte-stable when calibration is omitted.

Limitations

• Expects its channels. See Ship-config requirement — anchor-off is OOD.
• Base encoder not yet promoted — see Evaluation; the numbers here are real but not yet the production default.
• GB postcode anchor is outward-only — expect district-level (not unit-level) centroid precision from the anchor channel itself; the neural parser + gazetteer resolver still handle full unit-postcode text.
• Placetype-pair prior is comma-segment-scoped — it matches whole comma-delimited input segments only (the "segment" probe mode, default since d2a1242f), so comma-stripped GB input gets no boost from this channel (fully inert, not degraded — a documented v1 trade, not a bug). It is also hard-gated to GB: it structurally cannot fire on non-GB input.
• All-caps / shouting input degrades the admin tags (mixed-case training); @mailwoman/neural's normalizeCase opt recovers detected all-caps ASCII.
• Non-Latin scripts (CJK, Cyrillic) fall through to byte-fallback tokens; quality there is unmeasured.

License & links

• License: AGPL-3.0-only.
• Docs & concepts: https://mailwoman.sister.software
• Loader / scorer API: @mailwoman/neural (createScorer).
• The functional contract for this bundle is model-card.json (added 2026-07-23, now that a GB-graded checkpoint exists); this README.md is the HuggingFace-facing card.