Skip to main content

The sub-premise nexus — pre-tagging as information, not decision (direction note)

Operator instinct, 2026-06-11: "if there's any utility in pre-tagging 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. as a pre-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.' 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)., so be it… there's some nexus of secondary address 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). — apartments, suites, unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise., levels." Written down here with the bigger shape it connects to.

The nexus is real and has one grammar​

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. 123 · Apt 4B · Suite 500 · UnitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 9 · Level 12 · 3rd Floor · Private Bag 39990 · Drawer 61547 · CMB B99 — every one is 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. + IDENTIFIER: a closed-vocabularyvocabularyThe fixed set of tokens a tokenizer can produce. Mailwoman's SentencePiece vocabulary is tens of thousands of subword pieces, with byte fallback for anything outside it. leader (now provenance-tracked across codex/us|fr|ca|au|nz) followed by a short id with per-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. shape rules. Addressing standards treat them as one family too (USPS "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. 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."; AU's subpremisesubpremiseA secondary address unit inside a single street address (apartment, suite, floor), typically a designator plus an identifier such as 'Apt 4B'. The same proposer family as PO box. forms in the same AMAS table the #517 slices mined; NZ Post's delivery-service types). 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.' currently learns each tag's instances separately (unit shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row., po_box/cedex shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.) and shares nothing across the family.

Why pre-tagging-as-OVERRIDE stays rejected​

The deterministic pre-tagger was measured and reversed once already (#464: 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. override — perfect on a no-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. evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error., wrong on ethos and on contested surfaces). "Drawer" is a filing cabinet in a 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. name; "Level" is 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. name in places; "Box" is a surname. A pre-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.' 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). that DECIDES steals exactly the contested cases 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.' exists for.

The three slots pre-tagging-as-INFORMATION already has​

The architecture has rightful homes for this instinct, in escalating order of ambition:

  1. Input-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. clue (knowledge ladder rung 3.2 — cheapest, partially built). 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's 5 dims already include a po_box bit. The nexus version: a subpremisesubpremiseA secondary address unit inside a single street address (apartment, suite, floor), typically a designator plus an identifier such as 'Apt 4B'. The same proposer family as PO box. clue dimension that fires on ANY codex 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. (unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise., level, 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.-class, all localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.) — one bit + 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.'s tag-class, the same multi-hot pattern. 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.' conditions; it never obeys. Cost: a lexicon rebuild + a channel-dim bump (retrain-coupled — rides a scheduled run, not its own).
  2. 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). 2.7 phrase proposal (the operator's instinct, structurally). The phrase grouperphrase grouperStage 2.7 of the runtime pipeline: proposes coherent input units (street phrase, locality phrase, postcode, etc.) with structural kind hypotheses. Decouples boundary discovery from type classification so the classifier answers 'what type?' not 'where?' already carries 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. gate"; the generalizationgeneralizationA trained model's performance on data unlike its training set — new regions, new input distributions. The property honest eval is designed to measure. is a sub-premise proposer: one rule 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. over all codex 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. tables emitting typed 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. proposals ({span: "PO Box 123", kind: PO_BOX_PHRASE, confidence}), consumed as phrase priors today — the classifier conditions on the boundary hypothesis and can still disagree. This also subsumes the #518 bracket/paren proposer's machinery (same 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)., same output contract) — one proposer, two cue families (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. + paired delimiters).
  3. 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). 5 second emitter (#478's capstone — where this is ultimately heading). Codex matchers emit full CANDIDATES (po_box="P.O. Box 19" at chars 0–11, conf from matcher 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.) into reconcile's beam alongside the classifier's top-K; 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. picks per evidence. This is "the rules system lives on as a candidate source" — 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.'-first answer to v0's remaining edge-format wins (the postal 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.'s 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. formats are exactly where v0 still beats neural).

Sequencing against the live board​

Slot 2 is the natural first build: it needs no retrain (priors are 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.-side), its machinery is shared with the #518 revival verdict, and the v0.5.0 char-offset format makes its proposals directly storable as supervision later. Slot 1 rides whichever retrain comes after v0.5.0. Slot 3 is #478's existing capstone — this note adds the sub-premise family as its first candidate-emitter vocabularyvocabularyThe fixed set of tokens a tokenizer can produce. Mailwoman's SentencePiece vocabulary is tens of thousands of subword pieces, with byte fallback for anything outside it., with the proposer (slot 2) as its dress rehearsal.

Levels/floors join the family when their codex sourcing pass lands (noted on #517 — the AMAS documentation already mined covers AU Level/UnitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. forms).