The digit-ownership defect is cross-lingual, coverage-independent, and length-conditioned

2026-07-16. The measurement that unifies H3H3Uber's hexagonal hierarchical geospatial indexing system. Mailwoman uses H3 cells at resolution 9 (~0.03 km²) for geo-first blocking in the matcher and for stable address primary keys in @mailwoman/address-id., B0, and the one PL failure B1's 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. story could not explain — and that explains why B4's probe barely moved.

Finding: the intra-word digit incoherence B0 measured on Norwegian is the same defect on correctly-parsed Polish streetsstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. that are in-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. and admitted by the filter. It is 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., it is not the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.→localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. leak, and it is length-conditioned — 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.' reproduces a real 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. prior on long digit-run continuations. A targeted fragment shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. fights that prior uphill, which is why B4 (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. 12, 2k steps) moved the target only +1.7pp.

---

The one row coverage couldn't explain

B1 showed most Track B failures were the Norway YAML bug (25k rows dropped) — 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., 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.' defect. But one parity row survived: aleja Wojska Polskiego 178 → 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.. PL is in 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. and admitted by the filter, and the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. parsesaddress 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. correctly, so neither 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. nor the Track A streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.→localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. leak explains it. It is the single clean digit-ownership failure.

The piece-level trace (shipped v310, package-shaped)

Every streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. piece reads B-street/I-street — the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. is fine. The failure is entirely in the digit run:

aleja Wojska Polskiego 178   ▁1 START B-hn=0.397 B-pc=0.487   7 I-pc=0.417   8 I-pc=0.735   -> postcode
ulica Marszałkowska 140      ▁1 START B-hn=0.655 B-pc=0.217   4 I-pc=0.552   0 I-pc=0.800   -> postcode
aleja Jerozolimskie 91       ▁9 START B-hn=0.665 B-pc=0.196   1 I-pc=0.540                  -> postcode
aleja Wojska Polskiego 12    ▁1 START B-hn=0.823 B-pc=0.088   2 I-hn=0.646                  -> house_number ✓

This is exactly B0's Norwegian signature (Tindvegen nedre 44B: B-hn 0.604 on the first piece, I-pc 0.587/0.765 on the continuations). On Marszałkowska 140 and Jerozolimskie 91 the first digit piece is confidently B-house_number (0.655, 0.665) — 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.' starts the run correctly — and the continuations flip to I-postcode, so ViterbiViterbi decodingA dynamic programming algorithm that finds the most likely sequence of hidden states (labels) given a sequence of observations (token emissions). Mailwoman uses Viterbi over a linear-chain CRF to produce globally coherent BIO label sequences from per-token model scores.'s legality mask resolves the incoherent B-house_number → I-postcode to 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.-consistent path. Same mechanism, three languages (French, Norwegian, Polish).

It is length-conditioned, and that closes the loop with H3

aleja Wojska Polskiego 12 (2 digits) reads house_number correctly. aleja Wojska Polskiego 178 (3 digits, same streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.) fails. The only variable is run length, and the continuation 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. mass climbs with it (I-pc 0.417 → 0.735 across the two continuations of 178).

That is precisely what the H3 piece-prior measured in 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.: P(I-postcode | digit-run continuation) rises with length — 0.043 at 2 digits, 0.879 at 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.' is not malfunctioning; it is reproducing its 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. prior faithfully, and that prior says a long digit run's continuations are 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.. PostcodespostcodeThe 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. are long, house numbers are short, digits tokenize one piece per character, so length is the signal 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, and it uses it.

Three measurements, one mechanism:

	what it showed	unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.
H3H3Uber's hexagonal hierarchical geospatial indexing system. Mailwoman uses H3 cells at resolution 9 (~0.03 km²) for geo-first blocking in the matcher and for stable address primary keys in @mailwoman/address-id.	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. teaches P(I-pc | continuation) rising with length	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., per piece
B0	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.' reproduces it: B-hn first, I-pc continuations	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.', per piece (Norwegian)
this	same on correctly-parsed, in-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. Polish — 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.	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.', per piece (Polish)

Why B4 barely moved — this is the reason, not just the ratio

B4's verdict named --bare-street-prob 0.30 as the likely reason its target moved only +1.7pp. This probe adds the deeper reason: a fragment shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. is fighting a real, strong, length-conditioned 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. prior. At 3+ digits 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. itself pushes continuations toward 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. with mass up to 0.879. A targeted signal 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. 12 for 2k steps dents that; it does not overturn it. That is why the guards held (no harm) but the target barely moved — the shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. and the prior are pulling against each other on the same piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated)..

Two implications for B4b, both to be pre-registered not spun tonight:

1. Oversample long house numbershouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales.. The defect is worst at 3+ digits, where 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. prior is strongest. A shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. drawn from real NO/PL numbers is mostly 2–3 digits; deliberately weighting 3–4-digit house numbershouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales. aims the signal where the prior is hardest to beat.
2. The 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. direction is the real lever. The #727 research says a lower-fertilityfertilityHow many tokenizer pieces it takes to spell a word: fertility 1 means whole words survive tokenization, fertility 3 means they arrive in fragments. High fertility hurts sequence labeling directly — the model has to reassemble a word before it can label it, and fragment boundaries are where labels break. Mailwoman measured this on French street names: accented words splitting at the accent ('René' → '▁Ren' + 'é') caused real geocoding misses, and lowering French fertility was worth a tokenizer revision plus a retrain. vocabvocabularyThe 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. is upstream of any headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads., and this is the digit-specific evidence: a 3-digit number is 3 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). with 2 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.-leaning continuations because digits tokenize one-per-character. A vocabvocabularyThe 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. where 178 is one piece removes the continuation-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. mass entirely. That is a bigger change than a shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row., and it is the operator's call whether Track B justifies re-opening the vocabvocabularyThe 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. work.

The vocab lever, quantified

The tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. (v0.9.0-multisplice, 73,143 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated).) carries exactly two multi-digit numeric piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). — ▁10 and ▁16 — and none at 3+ digits. So:

   16 -> ['▁16']            0 continuations   (one of the only two)
   24 -> ['▁2','4']         1 continuation
  178 -> ['▁1','7','8']     2 continuations
90210 -> ['▁9','0','2','1','0']  4 continuations

Digit fertilityfertilityHow many tokenizer pieces it takes to spell a word: fertility 1 means whole words survive tokenization, fertility 3 means they arrive in fragments. High fertility hurts sequence labeling directly — the model has to reassemble a word before it can label it, and fragment boundaries are where labels break. Mailwoman measured this on French street names: accented words splitting at the accent ('René' → '▁Ren' + 'é') caused real geocoding misses, and lowering French fertility was worth a tokenizer revision plus a retrain. ≡ digit count. 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 only length signal is the continuation count, and the continuation is where 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. mass lives. That is not a tuning accident; it is the mechanism's root. 178 fails and 24 succeeds because 178 has two 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.-leaning continuations and 24 has one.

This makes the #727 "lower-fertilityfertilityHow many tokenizer pieces it takes to spell a word: fertility 1 means whole words survive tokenization, fertility 3 means they arrive in fragments. High fertility hurts sequence labeling directly — the model has to reassemble a word before it can label it, and fragment boundaries are where labels break. Mailwoman measured this on French street names: accented words splitting at the accent ('René' → '▁Ren' + 'é') caused real geocoding misses, and lowering French fertility was worth a tokenizer revision plus a retrain. vocabvocabularyThe 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. is upstream" claim concrete and scoped: a number-piece splice — add single piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). for common house-number and 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. ranges, exactly the shape of the FR multisplice that added 2,406 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). — would make 178 one piece with zero continuations, removing the continuation-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. mass for that class entirely. 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.' would then decide house_number-vs-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. on one piece in full context (is there a streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.? 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. slot?), not on a length-driven continuation prior. PostcodespostcodeThe 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. becoming single piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). too is a featurefeatureAn 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., not a bug: a 5-digit piece and a 3-digit piece are different piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated)., so 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 no longer forced to use continuation-count as its only length discriminator.

Cost, stated honestly: a splice grows the vocabvocabularyThe 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. and the embeddingembeddingA vector of numbers representing a token (or other item) so that similar items sit near each other in vector space. The first thing the model does is turn each token into an embedding. table; the new rows need gradientgradientThe direction and rate at which the loss would change if each parameter were nudged. Training follows the gradient downhill to reduce error. Huge gradients are tamed by gradient clipping. (the init_from fine-tune path the multisplice used). It is a bigger, coordinated change than a shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. — a tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. + 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.' bump, 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. edit — so whether Track B justifies it is squarely the operator's call. But it attacks the root the shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. only dents.

What this does NOT change

• The fix is still a shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. OR the vocabvocabularyThe 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. — not a validator. The house rule holds: this is positive evidence and 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., never a hard 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. veto. 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 first-piece B-house_number is usually right; the fix is to stop the continuations from overriding it, not to forbid postcodespostcodeThe 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..
• B4b is still the operator's call. This sharpens the hypothesis (oversample long numbers; or go to vocabvocabularyThe 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.) but does not license a solo 3am run.

Reproduce

node scratchpad/pl-piece-probe.run.ts