FR / non-US eval-coverage scorecard (#229 Phase A) — 2026-06-22

What this is: the consult's PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). A, grounded. Before the next FR/CJK capability push can be aimed, we need honest held-out floors on the fine + non-US components — and an honest read of which strata our current golden can't yet measure. This grades the production 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.' (v4.11.0 = v1.8.0, defaultVersion live) on the existing per-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. golden (data/eval/golden/v0.1.2/{us,fr}.jsonl, anchor + 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. fed, the per-locale-f1 ship-config harness), flags each floor's reliability by support size, and maps exactly what data each thin stratum needs. It is measurement + a data plan, not 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.' change.

The scorecard — production model (v1.8.0), per-locale per-tag

Reliability: ✓ reliable (n ≥ 100) · ~ thin (10 ≤ n < 100) · ✗ unmeasured (n < 10). Coordinate-relevance = does the tag reach the 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. we ship (the only metric that promotes).

FR (fr.jsonl)

tag	n	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.	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.	F1	reliability	coord-relevant
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.	1262	99.5	99.8	99.7	✓	✓
house_number	665	99.7	99.5	99.6	✓	✓ (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.)
streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.	665	90.1	90.1	90.1	✓	✓
localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.	1537	86.3	86.5	86.4	✓	✓
regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.	219	57.6	34.7	43.3	⚠ non-representative (95 multi-script + order-perms)	✓ (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.	93	43.0	92.5	58.7	~ thin	✗ invisible
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.	6	83.3	83.3	83.3	✗	~
dependent_locality	12	0	0	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.	1	—	—	0	✗ unmeasured	✓ (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.)
unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.	0	—	—	—	✗ absent	~
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. / street_prefix(_particle)	1 / 7 / 6	—	—	0	✗	~

US (us.jsonl, for contrast)

tag	n	F1	note
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. / house_number	1695 / 1031	98.6 / 98.4	✓
regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.	2956	89.6	✓ reliable (FR's 219 is non-representative, so US is the only honest regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. floor)
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.	1075	90.8	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.' can do 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. — FR's 0% is 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, not 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.' limit
localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.	1792	75.0	✓
streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.	2216	81.0	✓
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.	150	68.2	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.-bound here too (p=52.3) — 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. over-emission is global
unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.	2	9.1	unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. is under-measured everywhere, not just FR

Findings

1. The reliable FR floors are good — 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, house_number 99.6, streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 90.1, localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. 86.4. The shipped FR address resolves; this is the v1.8.0 win holding.
2. The region floor (43.3) is an ADVERSARIAL-STRESS number, not a representative real-FR measurement — corrected on inspection. A dump of the 219 région rows shows they are entirely synthetic stress permutations: 95/219 are multi-script (Cyrillic Шом, Han 富尔内勒, romanized Runik yozuv localities — the #555 non-Latin class), and the rest are order-permutations of a handful of rural communes (Creuse / Lozère / Thauron / La Ronze). 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.' emits région 99.6% on the in-distribution Locality, Département admin-split format. So the 34.7% 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. is a multi-script / OOD-robustness signal, not a representative real-FR région floor — and there is no representative real-FR région evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. on disk (the admin-split golden is in-distribution; using it would game the floor to ~96%). The honest statement: real-FR région performance is unmeasured; the only signals we have are the gamed in-distribution 99.6% and the adversarial-stress 34.7%. (country, by contrast, is genuinely coordinate-invisible — see below.)
3. country is 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.-bound and coordinate-invisible. 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. 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. 43.0 (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. 92.5) — 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.' over-emits 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., hallucinating it on rows with 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. 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.. This is global (US 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. 52.3 too). It's the exact tag tonight's shelved v1.8.1 tried to fix by adding France examples (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.↑/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.↓ — wrong direction, falsified). Fixing it is a 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.-only win.
4. venue and unit are unmeasured for FR (n = 1 and 0). The "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. 0%" that's haunted the FR narrative since #330 is measured on a single row — it is not a reliable signal, it's an absence of test data. US proves 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.' can emit 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. (90.8); the FR gap is that 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.' was never trained on FR 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. and we have no FR 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./unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. truth to grade it.

Spot-check — the shipped model on REAL Spanish addresses (novel, #148)

Since OA-ES is on disk (oa-cache/es__countrywide.zip), a 120-row held-out set was built from real Spanish cadastral addresses (52 provincesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality., named-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-types only, rendered in three natural orders) and graded on the production 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.' — a localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. 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.' has never been trained on (it's en-us + fr). The honest result:

tag	n	F1	read
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.	120	86.3	numeric — the one thing that transfers
localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.	119	49.6	half-right
house_number	120	38.4	trailing-number position (ES Calle X 1) trips the US/FR lead-number prior
streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.	120	24.9	genuinely weak
regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality./unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise./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.	—	0	not in OA-ES

Macro-F1 28.5%. Verify-before-verdict applied: the low streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. score is NOT a Calle-prefix labeling artifact — re-grading with a bare-name gold (street_prefix split out) drops streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. to 2.0%, i.e. 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.' does not emit Calle as a prefix; the weakness is real OOD. This quantifies the cost of the held #148 multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. retrain on a real, third localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. (not FR): a Latin-script EU localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. 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.' wasn't trained on resolves its 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. but mangles streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels./house_number/localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.. The OA-ES builder is a spot-check here; a committed build-oa-golden (ES + IT, both on disk) is the follow-up that would make this a standing non-US floor.

The honest non-US measurement — an 8-locale ASSEMBLED-COORDINATE panel (the headline)

OA carries truth coordinates, so held-out sets for eight localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. (150 rows each, all real, natural orders, build-oa-coord-golden.py; provenance + filters in the builder docstring) can be graded on the metric we ship — 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. → resolve → great-circle error — separating the resolve rate (did it produce a resolvable 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.?) from the resolved-only coordinate (how accurate when it does). This is the honest dial; 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 is confounded (the ES streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. 24.9% is a Calle-boundary artifact, confirmed by a bare-name A/B that drops it to 2.0%).

localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.	resolve rate	n resolved	p50 (resolved)	p90 (resolved)	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.
FR	80%	120 / 150	1.3 km	191 km	top (trained)
IT	79%	119 / 150	2.1 km	272 km	top (in #149 EU shardsshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.)
LU	57%	86 / 150	0.3 km	2 km	mid (small dense 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. — 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.-tight)
PL	53%	79 / 150	5.8 km	405 km	mid
PT	52%	78 / 150	1.2 km	216 km	mid
AT	50%	75 / 150	5.2 km	171 km	mid
CZ	43%	64 / 150	44 km	278 km	low-mid (loose resolve too)
AU	28%	20 / 72	234 km	2366 km	low (collisions)

Read the right column as a ceiling, not an average. The resolved-only coordinate is over only the addresses 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.' chose to resolve, and the ones it drops are disproportionately harder — so the resolved coord flatters 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.', most of all where the resolve rate is low (AU's 234 km is on 20 points, CZ's 44 km on 64 — treat both as noisy). The unbiased signal is the resolve rate, which is over the full sample. 150/localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. also carries a ±~8% band — rank the 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., don't over-read small mid-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. deltas.

8 resolvable localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. — the panel is ~complete for 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.-bearing OA. 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.: top FR/IT ~80%, mid LU/PL/PT/AT ~50–57%, low-mid CZ 43%, low AU 28%. Most resolved coords are citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.-to-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. tight (0.3–6 km) — the "where it resolves, it's accurate" rule — except CZ (44 km) and AU (234 km), which resolve loosely too (wrong same-name place; the dual-axis-worst localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.). DE/BE/DK/FI OA lack 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. column → not cleanly coordinate-gradeable (the resolve path needs the 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.); ES is cadastral → 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.-only. So 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.-bearing set is essentially mapped; broadening further needs 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.-complete sources.

This is the night's load-bearing finding — two axes, and it reframes #148.

• 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. is good where it resolves (EU). A resolved EU address lands citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.-accurate (p50 1–6 km) — so the 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. + resolution quality, conditional on resolving, is fine. This is why 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 misleads: 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.' gets locality + postcode right enough to geocode the right citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. even when it mis-tags 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, and 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 charges those boundary errors while the coordinate doesn't. (The #566 / v1.7.0 "grade the coordinate" lesson, confirmed on non-US with hard coordinate truth.)
• The real gap is resolve RATE (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 it tracks 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. REPRESENTATIONhidden stateThe model's internal vector for a token after the encoder has mixed in surrounding context. The contextualized representation the classifier head reads to assign a label.. The split is clean: FR 80% / IT 79% (the trained / well-represented localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. — FR is v1.8.0's home turf, IT rode the #149 EU shardsshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.) resolve at ~80% with citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy.-tight coords, while PT/PL ~52% (present but under-represented) and AU 28% (barely represented, + cross-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. name collisions like Windsor) fall off. 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.' fails to produce a resolvable 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. for ~half of mid-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. EU addresses and ~72% of AU. So the gap 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. of 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. distribution, exactly the #148 lever.

So the #148 multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. retrain's value is lifting 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. 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. on the non-IT localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. (+ AU collision handling), not fixing coordinate 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. — a more precise, cheaper-to-justify target than "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.' can't do non-US," now quantified across four real localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for..

Root-cause check — 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. gap, not 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 (so the lever is 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.', not 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.). PT resolves 52% but its 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. locality-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. is 39% (equally low), while where it resolves the coordinate is tight (1.2 km). If the unresolved half were 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. miss (localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. extracted, 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. can't place it), 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.-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. would be high and resolve-rate low — the opposite. Instead both are low: 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.' fails to extract a resolvable localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. for ~half of non-IT EU addresses. So #148 (retrain to lift 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. 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.) is the justified lever; 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. 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. is not the bottleneck (the EU candidate is comprehensive — see the #734 retirement). ⚠ Earlier-in-the-night caveat (verify-before-verdict): the IT-only read (p50 3 km → "median non-US geocodes well") was the BEST case; the panel corrected it — IT is the exception, not the rule.

Artifacts: scripts/eval/build-oa-coord-golden.py, data/eval/external/oa-{it,pt,au,pl}-coord-150.jsonl, scripts/eval/fr-admin-split-gate.ts --default-country <CC> (+ resolved-only metric).

Failure taxonomy (#375) — what kind of gap each is

stratum	gap class	fix lever
FR région (43.3 = adversarial-stress; real-FR unmeasured)	evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.-non-representativeness — the 219 rows are synthetic multi-script + order permutations; real-FR région is unmeasured (in-dist = 99.6% gamed)	a representative real-FR région held-out set (natural orders, Latin) — the prerequisite to even stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. the gap
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. (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. 43%)	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.' — over-emission; but coordinate-invisible	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. lever (suppress 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.-without-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.) — low priority, 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.-only
FR 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. (n=1)	evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.-thinness + 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. — no FR 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. truth, no FR 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. 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.	fetch FR POIspoint 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. → held-out 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. set + 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. 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. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row.
FR unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. (n=0)	evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.-thinness + 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. (global)	fetch 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 addresses (FR + thicken US)

Data-acquisition plan (the real Phase-A unblock)

⚠ Correction (verify-before-verdict, my own miss): FR address data is NOT blocked. fr/countrywide.csv (BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure.) is in openaddresses/europe.zip all along — I checked only the empty /tmp/oa-cache and wrongly called it gone. The FR coordinate evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. is now built from it (oa-fr-coord-150.jsonl, FR 80% / p50 1.3 km — top 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.). The genuinely-blocked FR strata are narrower: 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./unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. (no 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./unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. source — OA carries neither) and a real-FR region stratum (the OA-FR REGION column is empty, unlike OA-IT). For those, the next shift fetches, via the mailwoman CLI (never ad-hoc duckdb — the Overture OOM lesson):

• FR 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. → Overture places theme for FR (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. name + address) or 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. FR POIspoint 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. → render Venue, NN Street, PPPPP City, 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. = 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. name. Unblocks both the held-out 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. set and the T2 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. 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. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row..
• FR OOD région → re-fetch OA FR (BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure.) and render département in the varied real-world orders 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.' misses (NOT the in-distribution admin-split format, which would game the floor to ~96%).
• FR/US unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. → a real 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 source (unit-real-designators.jsonl exists for US as an external evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.; fold it into the golden + find an FR analogue).

GPU decision — no training tonight ($20 unspent)

Both planned GPU-stretch levers fail their bar on inspection:

• T1 (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. 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.) — HELD: coordinate-invisible. The gap is real but 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. never reads 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.''s 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. tag (placer-sourced in evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. + prod). Per "grade the 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., never 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," a 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.-only fix doesn't justify GPU — the same logic that correctly shipped v1.8.0 despite 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.
• T2 (FR 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.) — HELD: data-blocked. Coordinate-relevant, but 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. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. needs FR 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. strings we don't have on disk. Blocked on the same fetch as the 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. evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. set.

No lever clears the coordinate bar with on-disk data, so the disciplined call is no GPU. The substantive build pivots to the on-disk, coordinate-relevant 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. win (#734 AT/SK bilingual depth) — keeping the week's zero-GPU, real-data, coordinate-graded throughline.