No-street counter-example corpus shard — 2026-05-28

The corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus.-side fix for v0.6.1's dependent_locality regression. Per the 2026-05-28 night-2 postmortem and the Layer 1 morphology FST eval: the morphology FSTmorphology FSTA finite-state transducer encoding street-typing affixes (Avenue, rue, Calle) from libpostal dictionaries, for morphological street recognition. as a decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs.-only fix is insufficient because the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is too overconfident on its synth-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-induced dep_loc predictions. The recipe per DeepSeek's turn 2 consult is to add explicit counter-distribution 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. data: addresses where 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.' should NOT emit streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. labelscomponent tagOne of the 33 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. at all.

This evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. doc captures the new synthesizer + pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. glue that produces that counter-distribution shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row..

What lands

• corpus/src/synthesize-no-street.ts — six-template synthesizer:
  • venue-adversarial (35% weightparameterA 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.) — venuesvenueA 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. whose names contain streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-typing tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. (Wall Street Industries, 5th Avenue Theatre, Park Avenue Dental, Highway 61 Diner, etc.). The critical slice: these are exactly the rows that v0.6.1's "decompose mode" mis-tags. Explicit B-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. / I-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. labelscomponent tagOne of the 33 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. on tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. like Street/Avenue/Highway teach 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.' to suppress streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-side predictions when the surrounding context is 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.-shaped.
  • venue-plain (25%) — venuesvenueA 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. without streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-typing tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words..
  • locality-region-postcode (20%) — "Boston, MA 02101".
  • locality-region (12%) — "Boston, MA".
  • postcode-only (6%) — "02101".
  • country-only (2%) — "United States".
• corpus/src/synthesize-no-street.test.ts — 8 unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. tests covering each template, including a 500-iteration "never emits any streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.-side tag" contract check and a 1000-iteration template-distribution sanity check.
• scripts/build-no-street-shard.mjs — mirrors build-po-box-shard.mjs: reads {locality, region, postcode, country} tuples from JSONL stdin and emits aligned LabeledRow JSONL ready for the parquetParquetThe open columnar file format the corpus is written and streamed in. The training pipeline reads shards row-by-row from Parquet. sharding step.
• corpus/src/index.ts — exports the new synthesizer.

Verified end-to-end

Smoke testverdict-smokeA short diagnostic training run (≈1,500–3,000 steps) that must match the full run's gradient-noise profile, used to catch recipe instability before committing to a full 50k-step launch. with 5 base tuples (US, FR, DE) and --variants 4:

read 5 tuples, emitted 20 rows, skipped 0
template distribution:
  venue-plain:           8 (40.0%)
  locality-region:       5 (25.0%)
  venue-adversarial:     4 (20.0%)
  postcode-only:         2 (10.0%)
  country-only:          1 (5.0%)

Critical check — 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.-side BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components. across all 20 synthesized rows. The contract holds.

Sample of the critical adversarial output:

raw: 7th Street Bistro, Boston, MA 02101
  tokens: ['7th', 'Street', 'Bistro', 'Boston', 'MA', '02101']
  labels: ['B-venue', 'I-venue', 'I-venue', 'B-locality', 'B-region', 'B-postcode']

raw: Memorial Drive Medical Center, München, Bayern 80331
  tokens: ['Memorial', 'Drive', 'Medical', 'Center', 'München', 'Bayern', '80331']
  labels: ['B-venue', 'I-venue', 'I-venue', 'I-venue', 'B-locality', 'B-region', 'B-postcode']

raw: South Park Children's Center, München, Bayern 80331
  tokens: ['South', 'Park', "Children's", 'Center', 'München', 'Bayern', '80331']
  labels: ['B-venue', 'I-venue', 'I-venue', 'I-venue', 'B-locality', 'B-region', 'B-postcode']

The tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. Street, Drive, Park carry explicit B-venue / I-venue labelscomponent tagOne of the 33 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. — the 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.' sees direct evidence that these tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. are NOT streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. components in this surrounding context.

v0.6.2 recipe (proposed)

With this shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. landed, the full v0.6.2 retrain recipe per DeepSeek turn 2 is runnable:

1. Reduce synth-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. weightparameterA 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. 2.0 → 0.5
2. Add synth-no-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. at weightparameterA 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. 1.0, 50K–100K rows, US-primary with FR + DE + GB localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. sprinkled in proportion to the base tuple availability.
3. Enable 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. 1 morphology FSTmorphology FSTA finite-state transducer encoding street-typing affixes (Avenue, rue, Calle) from libpostal dictionaries, for morphological street recognition. at 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. (infrastructure already landed). Provides additive anchoring on top of the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus.-side fix.
4. 2D pre-publish evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. gate before promoting:
   • (recall drop >2pp AND baseline >10%) OR
   • (hallucination spike >100 AND rate >20% of golden occurrences)
5. Re-run v0-vs-neural harness post-train and target the failure clusters surfaced by the harness eval.

What this doesn't fix

• Non-US localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. gap. The harness showed neural at 0% on most non-US/non-FR localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.. This shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. helps with US/FR/DE/GB 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.-vs-streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. confusion but does not fix the localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.-distribution gap. Per-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. expansion is a v0.7+ topic.
• UnitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise. 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. schema gap. The Australian unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.-notation tests (Unit 12/345, Apt 12) fail because the 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). 3 schema has no unit_designator tag. This is a schema change requiring a separate retrain, not a corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. tweak.
• Tokenization on non-ASCII. Czech, Portuguese diacriticsdiacriticAn accent mark that modifies a letter (é, ñ, ç). Address normalization must fold diacritics for matching without discarding the information a user typed. are destroying 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. boundaries in the harness output. Independent of the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. path.

See also

• v0-vs-neural harness eval — establishes the failure clusters this shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. addresses
• Layer 1 morphology FST eval — proves the decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs.-only fix is insufficient and the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. side is required
• 2026-05-28 night-2 postmortem — original postmortem driving the v0.6.2 recipe
• Street-supplement architecture — the layered design context