#655 — cross-source threshold sweep: can a re-thresholded GBT beat FS?
TX-scoped, ≤2000 rows/source (NPPES org + TX HHSC nursing = eligibility-ish; FCC-RHC = funding), geocoded once then resolved per arm. Phone-corrob = of the cross-source entities whose records carry a phone in ≥2 different sources, the fraction where those phones MATCH — 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.-free 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. proxy (phone is not the join key). Higher cross-source + higher phone-corrob = better.
| arm | threshold | total entities | cross-source links | triple-source | phone-corrob (of checkable) |
|---|---|---|---|---|---|
| FS baseline | 0 | 3355 | 27 | 0 | 15/27 (56%) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -8.00 | -8 | 106 | 60 | 7 | 23/60 (38%) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -6.00 | -6 | 106 | 60 | 7 | 23/60 (38%) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -5.00 | -5 | 106 | 60 | 7 | 23/60 (38%) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -4.00 | -4 | 106 | 60 | 7 | 23/60 (38%) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -3.00 | -3 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -2.00 | -2 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ -1.00 | -1 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 0.00 | 0 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 1.00 | 1 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 2.00 | 2 | 4927 | 0 | 0 | 0/0 (—) |
| GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 2.83 | 2.8324 | 4927 | 0 | 0 | 0/0 (—) |
| cross-GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 1.47 | 1.4748999999999999 | 3239 | 33 | 1 | 17/32 (53%) |
| cross-GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 2.47 | 2.4749 | 3239 | 33 | 1 | 17/32 (53%) |
| cross-GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. @ 3.47 | 3.4749 | 4927 | 0 | 0 | 0/0 (—) |
Verdict
FS baseline: 27 cross-source links (0 triple), phone-corrob 56% (15/27).
No GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. threshold dominates FS — none matches FS's 27 cross-source links at ≥ its 56% phone-corrob without over-merging (entity count collapsing below 3019). At its dedup threshold the GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. finds FEWER cross-source links than FS; lowering the threshold to admit more only over-merges (the over-merge featuresfeatureAn input signal a model conditions on. Beyond the raw tokens, Mailwoman feeds soft features — gazetteer-membership channels and the postcode anchor — that inform predictions without overriding them. REPLACE the FS 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., so true cross-source pairs share a logitlogitA raw, unnormalized per-label score the model outputs before softmax. Priors and biases are added in logit space, then softmax turns logits into probabilities. band with genuine over-merges). Threshold alone (option 1) is INSUFFICIENT — FS stays pinned (correct + best-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. for this objective); a cross-objective retrain (option 2), gated on cross-source 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., is the only lever. See #655.
Candidate verdict (org-cross-GBT, added at review — the run predates the harness's candidate-verdict wiring)
The --candidate org 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.' (org-crosssource-gbt-en-us.ts @ its recommended 2.47): 33 cross-source
links vs FS's 27 (+22%) at 3,239 entities (above the 3,019 no-collapse floor), phone-corrob 17/32
(53%) vs FS 15/27 (56%) — more corroborated links in ABSOLUTE terms (+2), a −3pp rate that is exactly
one-link noise at n=32 (the harness's own caveat: the phone proxy "corroborates but isn't decisive").
No dedup-GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. collapse signature (its arms sit at 106 entities or 0 links).
Reading: the org-cross-GBTGBT (Gradient Boosted Trees). A non-linear machine learning model that combines many weak decision trees into a strong predictor. Mailwoman uses a GBT as an optional learned scorer for single-dataset dedup, improving F1 by 5–7 percentage points over the Fellegi-Sunter baseline. is the first scorer to EXCEED FS's link discovery on the org flows without over-merging. It misses STRICT dominance only on the noisy proxy's rate. Recommendation: un-pin FS for the org-level cross-dataset flows as an operator decision (the flip is a config default, not this evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.'s to make); alternatively hold FS pinned and re-judge after widening the 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. set with the other Care Compare families (dialysis/hospice/SNF/HHA), which should tighten the proxy's n.