v3.0.1-span-head-lr — the trained span scorer beats the token decode

Gate (config-canonical, quoted verbatim from v3.0.1-span-head-lr.yaml):

seg@1 > token@1 on the parity 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., via scripts/eval_seg_at_1.py. v3.0.0 read seg@1 0.004 vs 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.@1 0.348 (CHANNEL-STARVEDstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed). numbers). Secondary read: the 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. NLLNLL (Negative Log Likelihood). The standard form of a loss function in probability-based models: loss = −log P(correct label | input). The CRF uses the NLL of the entire label sequence. must actually converge. If lossloss functionA number measuring how wrong the model's predictions are on a batch of examples. Training minimizes it. Mailwoman's loss combines per-token negative log-likelihood with the CRF sequence loss. again ends ~17 and still falling, the LR is not the binding constraint and the fork reopens. If seg@1 STILL loses with a converged 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. lossloss functionA number measuring how wrong the model's predictions are on a batch of examples. Training minimizes it. Mailwoman's loss combines per-token negative log-likelihood with the CRF sequence loss.: that IS the falsification.

Verdict: PASS

	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.@1	seg@1
v3.0.1 (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. LR 1e-3)	131/267 = 0.4906	152/267 = 0.5693
v3.0.0 (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. LR 1e-5)	93/267 = 0.348	1/267 = 0.004

seg@1 beats 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.@1 by +7.9pp under identical conditions. PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1's question — does a segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. decode over LEARNED 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. scores beat the 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. decode? — is answered yes.

The secondary read confirms the diagnosis, not just the outcome

The v3.0.0 write-up called the failure under-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., not falsification, on the strength of a lossloss functionA number measuring how wrong the model's predictions are on a batch of examples. Training minimizes it. Mailwoman's loss combines per-token negative log-likelihood with the CRF sequence loss. that was still falling at 2k with raw 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. NLLNLL (Negative Log Likelihood). The standard form of a loss function in probability-based models: loss = −log P(correct label | input). The CRF uses the NLL of the entire label sequence. ~35. One variable (the 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.'s LR) tested that claim:

	step 400	step 2000	shape
v3.0.0 (1e-5)	~24	17.77	still falling — never converged
v3.0.1 (1e-3)	1.69	1.37	flat from ~1800 — converged

Raw 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. NLLNLL (Negative Log Likelihood). The standard form of a loss function in probability-based models: loss = −log P(correct label | input). The CRF uses the NLL of the entire label sequence. ~2.7 (1.37 ÷ the 0.5 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.) against v3.0.0's ~35. The LR was the binding constraint. A fresh 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. cannot train at a pretrained encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters).'s fine-tuning LR — the probe was mis-specified, exactly as diagnosed, and repairing it changed the answer completely.

Functional evidence (the aggregate is not the verdict)

'Korunni 810, Praha'  → Korunni:street  810:house_number  Praha:locality
'350 5th Ave, NY…'    → 350:house_number  5th:street  Ave:street_suffix  New York:locality  NY:region  10118:postcode
'Rue Montmartre'      → Rue Montmartre:locality        ← still wrong

• Korunní 810 is the #727 archetype — the case v264 mangles into street="Korunní 8" + house_number="10", the boundary-inside-a-number failure that motivated this whole arc. The 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. decode gets it right. Compare v3.0.0's random-per-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. output (▁Kor:street un:region ni:subregion).
• Rue Montmartre → localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. is still wrong, and that is expected: it is the bare-fragment 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. class (66% of streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. failures, night-3 partition), which the 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. 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. was never going to fix. It is option C's target (kind-posterior soft channel + 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.-weighted lossloss functionA number measuring how wrong the model's predictions are on a batch of examples. Training minimizes it. Mailwoman's loss combines per-token negative log-likelihood with the CRF sequence loss.) and is deliberately out of PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1's scope. Its survival here is the plan's prediction holding, not a surprise.

Bonus finding: the span loss improved the BIO 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.@1 went 0.348 → 0.4906 on the same channel-starvedstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed). harness between v3.0.0 and v3.0.1, and val macro F1macro F1The unweighted average of per-class F1 scores — treats every class equally. Mailwoman's primary label-level eval metric. went 0.7405 → 0.7479. The two runs differ only in the 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.'s LR, so the co-trained 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. objective is shaping the shared encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters). to the BIO 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.'s benefit — the same "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.-consistency pressure helps" effect 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).-1's aux 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. showed, but larger. Worth confirming on production config in PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2; not claimed as a shipped win here.

What this does NOT establish

• These absolutes are channel-starvedstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed).. eval_seg_at_1.py feeds no 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./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. channels, no postcodeRepair, no word-consistency heal (the #718 trap). Its 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.@1 reads 0.4906 where the JS harness reads 0.573 on v264. Do not compare across harnesses. The gate is valid because it is a relative comparison — both headsattention 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. read the same starvedstarvedA tag with too little training representation to learn — near-zero F1 — because the corpus adapter never emits examples of it (intersection tags sat at 0% until an intersection synthesizer existed). encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters). stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality..
• The 0.90 parity floor is untouched by this. PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 1 was never a promote gate; this checkpointcheckpointA saved snapshot of the model weights and optimizer state during training. Mailwoman saves a checkpoint periodically so training can resume after a GPU hang. ships nowhere. Whether seg@1 clears 0.90 under production config is a PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary).-2 question, and the night-3 oracle@10 ceiling (0.749 over v264's emissions) says the floor needs the 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. class too, not just the boundary class.
• 2k steps, one seed. No claim about the 8k shape.

Run facts

Run	ap-XoqeNxSPUsObFIpsM7nXZL, A100, 2000/2000, ~23 min
Variable	train.span_head_learning_rate: 1e-3 — the ONE change vs v3.0.0
Param groups	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. 101,076 @ 1e-3 | encoderencoderThe part of a transformer that turns input tokens into contextualized vector representations. Mailwoman's classifier is a small encoder-only transformer (~30M parameters). 39,259,055 @ 1e-5 (printed by the run)
Init	init_from v264 step-008000, missing=9 unexpected=0
NaNNaN (not a number). A floating-point result for an undefined operation (log of a negative, 0/0). Appearing in the training loss usually halts the run; recovering from it follows the NaN protocol.	none — fp32fp32 / fp1632-bit and 16-bit floating-point formats. Mailwoman trains in bf16 (a 16-bit variant) and exports the ONNX model in int8 for size. partition math held again
CheckpointcheckpointA saved snapshot of the model weights and optimizer state during training. Mailwoman saves a checkpoint periodically so training can resume after a GPU hang.	output-v301-span-head-lr-s42/checkpoints/step-002000

Next

PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2 (ONNXONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. export of the 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. scores + the #378 SLO check) is now justified — it was gated on exactly this crossing. Then Phase 3 (JS k-best) and PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 4 (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. rerank + option C).

The one thing worth deciding first: an 8k run at this LR would tell us whether seg@1 keeps climbing before PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2's export work commits to a checkpointcheckpointA saved snapshot of the model weights and optimizer state during training. Mailwoman saves a checkpoint periodically so training can resume after a GPU hang. — cheap (~1.5h A100) and it is the same one-variable step the campaign already knows how to grade.