The resolution ladder — coordinate resolution plan (2026-06-11)

This document replaces two working drafts (the operator's interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. musings and DeepSeek's plan synthesis, both reviewed 2026-06-11 and deleted in favor of this). It keeps what survived review, drops what conflicts with settled architecture, and re-sequences the correctives. It is the planning home for #483 (interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable.), the unbuilt spatial tierstierInternal 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., and the confidence 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.' that sits over all of them. #484 (reverse) ships and continues on its own issue.

The principle, and the ladder

Prefer exact locations; interpolate only as a fallback; expose confidence and provenance on every answer. That has been this project's posture all along (resolution_tier, interpolated: true, uncertaintyM, approximate containment, source/release per row) — the musings' contribution is a shared 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. for the whole ladder, which we adopt:

Rung	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.	Status
1	Entrance point	unbuilt (future, behind footprints)
2	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. / building centroid	unbuilt — rides Overture buildings (#470)
3	Address point	✅ #476, wired (applyAddressPoint)
4	ParcelparcelA property polygon or record carrying a situs (site) address and often a separate owner mailing address. County GIS parcel aggregations are a training source for address-point variety and situs-vs-owner divergence. centroid	unbuilt, deliberately deferred (US parcelparcelA property polygon or record carrying a situs (site) address and often a separate owner mailing address. County GIS parcel aggregations are a training source for address-point variety and situs-vs-owner divergence. data is patchwork-licensed)
5	Address-point interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. ("Method 2")	the next corrective — see PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1
6	StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-range interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. (TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data.)	✅ #483 pilot, standalone, gate MISS on record
7	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. interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. ways	blocked on ODbL treatment (#26)
8	StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. centerline	implicit today (segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. match without number)
9	Admin centroid (localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. → regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.)	✅ the 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. 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.

Every 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. answers the same shape — (street, number, postcode/locality scope) → flagged coordinate — so 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. walks an ordered list and the first hit wins. TierstierInternal 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. are data problemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed., not architecture problemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed., from here on.

Stack positions (reaffirmed, not open)

The musings recommended OpenSearch/Elasticsearch for search and PostGIS for the spatial 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.. We decline both, on existing grounds: the stated goal explicitly excludes Elasticsearch (hierarchy lives in-DB via the ancestors table), and the deployment story — slim DBs over httpvfs, the browser 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., static-asset hosting — depends on sqlite everywhere. Our search 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. is FTS5FTS5SQLite's built-in full-text-search module (with BM25 ranking). Mailwoman uses it for prefix and name matching against the gazetteer. + the 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. + the typo-tolerant retrieval 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. now scoped in #531; our spatial 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. is the R*TreeR*TreeSQLite's spatial index of bounding boxes, enabling fast geographic range and nearest-neighbour queries in the resolver. + ray-cast PIP that #484 shipped. The musings' fuzzy-matching section is #531 in different words; its autocomplete section is #190. Both were already queued.

Likewise "avoid full reindexing": our artifacts are immutable provenance-tracked shardsshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. rebuilt per source vintage (TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. yearly, Overture monthly). That cadence is the update story; we do not take on incremental index mutation.

Phase 1 — the gate corrective, re-sequenced

The #483 pilot missed its pre-registered gate (p50 66 m vs ≤ 50, p90 249 m vs ≤ 150, VT, n=5000). The dominant error term is TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data.'s uniform-spacing assumption over long rural segmentssegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.. Two correctives were on the table; review collapsed them into one decision:

Method 2 — address-point interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. — is the primary corrective, promoted from "PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 5" to now. SegmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.-subdivision-at-anchors and address-point interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. are nearly the same computation (bracket the query number with known points, interpolate between them), but Method 2 is the simpler of the two: it needs no TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. at all — the #476 Overture shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. alone supplies the known points — and it replaces theoretical capacity ranges with real occupancy, which is precisely the failed gate's error term. TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. range interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. demotes to what it should always have been: the fallback for streetsstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. too sparse to bracket.

Order of work:

1. Density characterization first (nearly free): re-run the existing evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. on a dense county (Cook IL or Kings NY class). If dense passes and VT fails, the miss is geometry-capped and the gate becomes county-stratified — by stated re-baseline with operator sign-off, never a quiet edit. If dense also fails, the method itself needs the corrective regardless.
2. Method 2 implementation: given (street, number, scope), find bracketing address pointssitus dataA dataset of exact address-point coordinates (rooftop-level). Mailwoman's geocoder uses a national situs layer (124.9M US points built from state address-point sources) as the highest-precision tier of the geocode cascade. on the normalized streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.; interpolate linearly between them; single-sided bracket = extrapolation with an explicit uncertainty penalty; no bracket = fall through to TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data.. Same honest-evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. harness, same gold, held-out points interpolated only from non-held-out neighbors (non-circular by construction).
3. Side-of-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. offset (~10 m perpendicular on parity match) only after the above — it is not the dominant term.
4. ZIP+4 snapping stays deferred behind #525.

The gate does not move except by stated decision. A second miss is a second postmortem.

Phase 2 — resolver wiring + the workspace split

• 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. interface: the ordered list (DeepSeek's Option B, endorsed). ResolveOpts grows a spatialTiers ordered array of lookups sharing the find() shape; addressPoints becomes the first entry rather than a special case. We have two tierstierInternal 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. today and a ladder of nine above — per-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. opts members do not scale.
• Workspace: split. @mailwoman/resolver-interpolation as its own workspace — TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data./Overture vintage lifecycle vs 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. continuous, national shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. size (hundreds of MB) vs the 9 MB 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., independent versioning. The shared streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. normalizer stays the single build-time/query-time function it already is. Do the split before national builds so they land in the right home.

Phase 3 — nationalize

Multi-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. + Overture shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. orchestration (the builders are per-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. already); ZIP→localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. scoping so "123 Main St Springfield" works without a 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. (build-time join of segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. ZIPs to localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. names via 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. ancestry); 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. interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. ways for EU 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. only after #26 resolves ODbL treatment. The interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. module stays source-agnostic — TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. edges and 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. ways are both (street, from, to, parity, polyline) rows; only builders know sources.

Phase 4 — building centroids (the highest unbuilt rung)

The best marginal-accuracy 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. not yet built (~10–20 m vs interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable.'s 50–200 m), and closer than it looks: Overture's buildings theme (Microsoft footprints folded in) rides the #470 ingestion epic already staged. Spatial-join footprints to address pointssitus dataA dataset of exact address-point coordinates (rooftop-level). Mailwoman's geocoder uses a national situs layer (124.9M US points built from state address-point sources) as the highest-precision tier of the geocode cascade.; 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. centroids where a point is missing. ParcelsparcelA property polygon or record carrying a situs (site) address and often a separate owner mailing address. County GIS parcel aggregations are a training source for address-point variety and situs-vs-owner divergence. stay deferred behind this.

Phase 5 — confidence, calibrated not hand-assigned

The musings proposed static 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. (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. 0.95, interpolated 0.50, …). We keep the ordering as a prior and reject the constants: this project has a calibration discipline (the isotonic work), and 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. confidence should be measured — P(error < X m) per 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. per density stratum, fitted from the evalsevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. we already run, recalibrated when shardsshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. rebuild. Hand-assigned constants are make-or-break trivia in number form. Provenance extends to a structured chain (dataset, release, 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 flag) — the fields already exist per row; the chain is assembly.

Phase 6 — learned placement (research track, strictly gated)

The novel idea in the musings: replace linear interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. with a 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.' that predicts where addresses sit on a block. Verdict from review: keep it, as research, with two corrections. First, drop the location-encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters). framing — GeoCLIP/SatCLIP encode global position for geo-localization; this problem is where along a 200 m segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context., and the informative featuresfeatureAn input signal a model conditions on. Beyond the raw tokens, Mailwoman feeds soft features — gazetteer-membership channels and the postcode anchor — that inform predictions without overriding them. are segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.-local (bracketing-number positions, density, footprint geometry). SatCLIP's own card rules out fine-grained many-close-location use. Second, the baseline to beat is Method 2 itself — which is already nearest-neighbor regression on real data. The bar: beat it per density stratum on the same held-out gold, with the deterministic 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. as the abstain fallback. No ship commitment until that bar clears.

Sequencing

Now    Phase 1  density characterization → Method 2 → re-gate
Next   Phase 2  ordered spatialTiers + workspace split          (then) Phase 3 nationalize
Then   Phase 4  building centroids (rides #470)
Later  Phase 5  calibrated confidence  ·  Phase 6 learned placement (parallel research)

Decisions (ruled 2026-06-11 unless marked open)

Decision	Ruling
Search/spatial stack	sqlite + FTS5FTS5SQLite's built-in full-text-search module (with BM25 ranking). Mailwoman uses it for prefix and name matching against the gazetteer. + 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. + R*TreeR*TreeSQLite's spatial index of bounding boxes, enabling fast geographic range and nearest-neighbour queries in the resolver./PIP — no ES, no PostGIS (standing)
Primary gate corrective	Method 2 promoted to PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1; TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. demotes to fallback
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. interface	Option B ordered spatialTiers list
Workspace	split to @mailwoman/resolver-interpolation before national builds
Gate re-baseline	only county-stratified, only by stated sign-off
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. ways	blocked on #26 (ODbL)
Learned placement	research-only; gate = beat Method 2 stratified
ParcelparcelA property polygon or record carrying a situs (site) address and often a separate owner mailing address. County GIS parcel aggregations are a training source for address-point variety and situs-vs-owner divergence. 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.	deferred behind building centroids