resolver_score cannot reject garbage — a characterization

Date: 2026-08-04 · Issue: #40, mailfail finding 5 · Scope: what metadata.resolver_score is, what it can and cannot decide, and what an abstention surface would have to be built on instead.

This is a characterization only. No scoring code was changed, and none of the designs at the end were implemented. Everything numeric here was re-measured today against the committed fixture (mailwoman/eval-harness/fixtures/mailfail.jsonl, 105 rows) plus a real-address control (data/eval/external/oa-us-coord-150.jsonl, 150 US addresses with government coordinates), on both 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. backends.

---

Verdict in one paragraph

resolver_score is not one number. On the FTS backend it is a negated bm25 plus about six points of additive tiers; on the candidate backend it is exactly log10(population + 1). The two live on scales that do not overlap (FTS localities 9–44, candidate localities 0–7), and within the FTS backend a second code path returns a bounded [0, 1] blend instead. Its own type says so: "Scale is implementation-defined; callers should treat as ordinal." So the interesting question is not "what threshold abstains" but "is this number ordinal within one query, or is it a magnitude you can compare across queries?" — and the answer is the former. Measured on the fixture, a single score threshold separates a correct localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. from a garbage localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. with Youden J = 0.357 (FTS) / 0.573 (candidate). The classifier's own 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. confidence, already on the same node and already in [0, 1] for both backends, gets J = 0.929 / 0.917 on the same populations. The signal that abstains is upstream of 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..

---

1. Where the number comes from

resolver/resolve.ts:1119 is the only place it is written, and it copies the backend's value verbatim:

node.metadata = { ...node.metadata, resolver_score: resolved.score, resolver_name: resolved.name }

There is no normalization, calibration, or clamping between the backend and that line.

FTS backend — resolver-wof-sqlite/lookup.ts

Two distinct regimes, chosen per query.

Regime A (the normal path), lookup.ts:756-843. bm25(place_search) AS rank from the FTS5FTS5SQLite's built-in full-text-search module (with BM25 ranking). Mailwoman uses it for prefix and name matching against the gazetteer. query, negated, plus additive tiers:

// SQLite's bm25() returns a lower-is-better score (negative for matches). Negate so we
// start from a higher-is-better baseline.
let score = -row.rank

then +0.5 placetypeplacetypeThe Who's On First hierarchical classification of places: planet → continent → country → region → county → locality → neighborhood. The resolver uses placetype to rank candidates — an exact locality match outranks a county-level match. match, +0.2 implicit localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy., +0.3 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. match, +0.5/+0.2 direct-child/descendant, −0.1 × extraLen/10 length penalty, up to +0.8 proximity, up to +4.0 population (resolver-wof-sqlite/ranking-weights.ts:119-148). The bm25 base is unbounded and the whole additive stack tops out around +6.3, which is why exactMatchTiering exists — the code's own note records "BM25 gaps of 1.5-3.0 between famous places and tiny same-name peers", i.e. the base term is noisy at exactly the scale the boosts operate on. exactMatchTiering re-sorts by match tier first and uses this score only to break ties within a tier.

Regime B (coordinate-first postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. path), lookup.ts:1019-1146. A weighted blend of three [0, 1] terms:

score: w.pc * sPc + w.name * sName + w.pop * sPop // WORLD_DEFAULT weights { pc: 0.6, name: 0.3, pop: 0.1 }

WeightsparameterA 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. sum to 1, so this score is bounded in [0, 1] — a different scale from Regime A, out of the same backend, on the same field.

Candidate backend — resolver-wof-sqlite/candidate-lookup.ts

neg_rank is precomputed at build time as -log10(pop + 1) (build-candidate.ts:176), and:

// `score` stays the RAW population rank (`-neg_rank`) — it feeds the resolver walk's absolute
// `minWinningScore` gate (`resolve.ts`), which must see real prominence, never a penalized value.
score: -Number(row.neg_rank),          // candidate-lookup.ts:482

So score = log10(population + 1), exactly. Two exceptions: a postal-citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy.-exact hit short-circuits with a flat score: 1 (candidate-lookup.ts:328), and a place with no population data scores 0. The bounded cross-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. primary preference (PRIMARY_PREFERENCE_LOG10 = 1) and the optional proximity re-rank both write prominence, never score.

This is confirmed by measurement: every postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. node the candidate backend resolved in this run scored exactly 0.00 (n=155 across garbage + control), and Seoul scored 7.01 — log10(10.2M) ≈ 7.01.

Who consumes it today

Almost nobody, and nothing blocks on it by default.

• resolve.ts:1076 — if (top.score < state.minWinningScore) return null. This is the one site that can reject a result. minWinningScore defaults to 0 (resolve.ts:757) and no production caller sets it (grep: only resolver/resolve.test.ts:311). Its own docstring says "Score scale is implementation-defined; tune per backend."
• resolver-wof-wasm/browser-cascade.ts:302 — a tiebreak within a placetypeplacetypeThe Who's On First hierarchical classification of places: planet → continent → country → region → county → locality → neighborhood. The resolver uses placetype to rank candidates — an exact locality match outranks a county-level match.-rank tier (b.rank - a.rank || b.hit.score - a.hit.score). Same backend, same query, so scale-consistent and safe.
• core/pipeline/reconcile.ts:565 — normalizeResolverScore clamps to (0, 1] for a multiplicative beam combiner. Because both real backends routinely return scores above 1, this saturates almost every candidate to exactly 1, contributing "was there a match at all" and nothing else. It is on the joint-reconcile path, which is default-OFF (runtime-pipeline.ts:479, jointReconcile ?? false, retired as default 2026-06-14), so this is latent rather than live — but it is a trap waiting for whoever re-promotes that path.
• mailwoman/geocode-core.ts reads primaryNode.alternatives to build the geocode candidates array, and never reads the score. The API-facing confidence field comes from the coarse-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. placer, not 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..

Everything else is observational.

---

2. What was measured, and how

Two runs per backend (105 garbage probes + 150 control addresses each), through the shipped createRuntimePipeline({ classifier, resolver }) with locale: en-US, capturing every node that came back carrying coordinates: its tag, value, resolved place nametoponymA proper name for a geographic place., resolver_score, classifier confidence, alternatives count, the distinct tags present in the whole 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., and (for the control) the great-circle error to the government point.

• FTS backend: admin-global-priority.db (5.3 GB, sealed, md5-stamped 2026-08-02). No postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. database, so postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. nodes never resolve on this leg.
• Candidate backend: candidate.db → candidate-global-1026.db (1.65 GB, sealed).

Scripts live in the gitignored scripts/diagnostic/ (score-probe.ts, score-report.ts, score-bars.ts, score-signals.ts, source-split.ts, conf-dist.ts), following the same convention as the mailfail probe generator.

Definitions, stated because the answer depends entirely on them:

• A garbage resolution is any coordinate produced for a fixture row. A violation is narrower: a coordinate produced for a row whose committed expect bar is no-resolve (35 of the 105 rows carry that bar).
• A correct control resolution is a control localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. node within 25 km of the government point. 25 km is deliberately loose because admin-centroid resolution legitimately misses an edge address by tens of km.

---

3. The reproduction, including what did not reproduce

Resolve rates

backend	probes resolving ≥1 coordinate	no-resolve bar violated	resolved nodes
FTS	33 / 105	14 / 35	56
candidate	41 / 105	18 / 35	62

The mailfail report's headline was "35 of 110 probes produced at least one coordinate". Re-measured on the 105 committed rows that is 33 (FTS) — consistent, since the fixture drops the oversized size rows.

The 33 is not 33 defects. Nineteen of the resolving FTS probes carry a bar weaker than no-resolve, and the fixture is right to grade them that way: 350 5th Ave, New York, NY (leading BOM), Café de Flore, Paris (NFD accents), רחוב דיזנגוף 100, תל אביב, the U+2028-separated address, the repeated-address size row, the CSV and YAML rows that contain New York, and Springfield are real place namestoponymA proper name for a geographic place. inside awkward wrappers. Resolving them is correct behavior. Only the 14/35 no-resolve violations are the defect: 1 → Zona 1, 0 → Purwa 0, a → A, home → Home, Aug → Aug, boom → Boom, amet → Amet, all → All, null → Null, DROP → Drop, Quote → Quote, NEAR → Near Acres Estates. One row (g76, a SQL statement containing 'New York') carries no-resolve and contains the citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. name New York — that bar is arguably mis-set.

Score distributions (locality nodes)

population	FTS	candidate
no-resolve violations	n=14 mean 21.05 p50 23.92, 13.45–28.74	n=12 mean 2.26 p50 2.65, 0.00–6.95
correct control localities (≤25 km)	n=149 mean 25.63 p50 24.58, 18.73–43.98	n=150 mean 4.47 p50 4.58, 0.00–6.44

Best single threshold, sweeping every observed score:

backend	threshold	keeps correct	drops violations	Youden J
FTS	≥ 18.73	100.0%	35.7%	0.357
candidate	≥ 3.67	74.0%	83.3%	0.573

Nine of the 14 FTS violations score above the lowest-scoring correct control localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy.. That is the finding, stated without the sign confusion: there is no threshold on this field that separates the two populations, and the thresholds that do best are different numbers on different backends because the scales are different.

What did NOT reproduce

The framing this task arrived with was that wrong answers score higher than right ones — FTS 25.67 vs 24.33, candidate 6.52 vs 6.07. On the definitions above, that inequality does not hold: correct localities score higher than garbage localities on both backends (25.63 vs 21.05; 4.47 vs 2.26).

The sign flips when you pool tags. Grading every resolved node by distance-to-truth, on the candidate backend:

control right <=25km   n=323 mean=2.53
control wrong  >25km   n=127 mean=6.43     <- "wrong scores higher"

That "wrong" bucket is 125 regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. nodes and 2 postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. nodes. A regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. node that resolves California correctly still sits 200 km from a Mill Valley address, so distance grading marks it wrong; and regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. rows carry the highest scores on that backend (5.31–7.59, vs 0.00–6.44 for localities and a flat 0.00 for postcodespostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.). The pooled statistic is therefore dominated by tag, not by correctness — 6.43 is close to the 6.52 in the original claim, which makes this the likely mechanism behind it.

That is not a correction that rescues the score. It is a second, independent reason not to threshold on it: resolver_score is not comparable across tags within one backend, let alone across backends. Any pooled mean over this field is measuring composition.

The scale table, which is the whole argument

backend	localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy.	regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.	postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.
FTS	9.11 – 43.98	4.80 – 28.98	(not wired)
candidate	0.00 – 7.01	5.31 – 7.59	0.00 (always)

A threshold of 5 means "reject nearly everything" on FTS and "accept most cities" on candidate. A threshold of 20 is unreachable on candidate. There is no shared unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise..

---

4. What separates the two populations, since the score does not

Same populations, same runs, other fields already present on the node:

signal	FTS J	candidate J	note
resolver_score	0.357	0.573	backend-specific scale; no shared threshold
classifier 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. confidence	0.929	0.917	already [0, 1], already on the node, backend-independent
corroboration (another component present)	drops 7/14	drops 6/12	costs 0 on this control — but see the caveat
alternatives count	1.50 vs 1.50	2.33 vs 0.81	no signal on FTS; some on candidate

Confidence detail (FTS leg):

violations  n=14  min=0.062  p50=0.500  max=0.964
correct     n=149 min=0.918  p50=0.928  max=0.945

At a 0.918 threshold, 13 of 14 violations drop and no correct control localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. is lost. The single survivor is +1 (555) 867-5309 → 1 → Zona 1 at 0.964, the highest-confidence violation in the set — a phone number 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.' is more sure about than any real address it read.

Three caveats, because the numbers above are seductive:

1. The correct-control confidence band is absurdly tight (0.918–0.945 across 149 addresses). A signal that near-constant on clean structured US input will spread on a harder control, and the 0.918 threshold is nothing more than the minimum of that cluster. Any threshold work must re-derive this on a multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for., fragment-shaped control before it means anything.
2. Corroboration costs "0 of 149" only because every control row is a full streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. address. Springfield — a bare citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. name a user types, and a row in this very fixture — is uncorroborated by construction. A corroboration-only rule would delete the bare-localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. search path, which is the map-search registerinput modeThe Decision-A register switch: 'fragmented' (human-typed fragments — feeds the evidence channels) vs 'formatted' (complete records — runs the trained absence identity). Explicit on CLI/API; per-endpoint defaults (batch→formatted, autocomplete→fragmented); kind-derived otherwise. the product is aimed at.
3. n=14 and n=12. These are directional results on a small violation set, not calibrated numbers.

Fields already on ResolvedPlace that nothing currently reads at resolve.ts:1076, and that a design could use: exactMatch (match-quality tier — already the primary sort key elsewhere), prominence (bounded ~[0, 8], defined the same way on both backends, and therefore a far better candidate for a shared threshold than score), mismatch (an explicit postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one./localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. conflict flag), resolutionQuality (an explicit fallback-tier flag), and the top-vs-runner-up score margin, which is computed nowhere today even though alternatives is retained.

---

5. Three candidate designs for an abstention surface

Presented in ascending order of cost. None of them is a recommendation; each carries what would have to be measured to promote it.

Design A — per-backend calibration of the existing minWinningScore

Keep resolver_score as-is. Give minWinningScore a per-backend default, sourced from the backend rather than the caller: a scoreProfile on the PlaceLookup/ResolverBackend interface declaring { scale: "bm25" | "log10pop" | "unit", abstainBelow: number }, which resolveTree reads instead of the current opts.minWinningScore ?? 0.

• Cost. Small and contained: one optional interface field, one line in resolve.ts:1076, two backend implementations. No new 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., no 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.' work, no artifact rebuild.
• What it buys. At the measured optimum it drops 36% (FTS) / 83% (candidate) of no-resolve violations. The candidate number is decent; the FTS number is not.
• What it does not buy. It cannot fix the tag-incomparability — 5.31 is a plausible regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. and an implausible localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. on the same backend — so the profile would really need to be per-(backend, placetype), and by then you are hand-tuning six constants against a 14-row violation set.
• Evidence required to promote. A per-backend, per-placetypeplacetypeThe Who's On First hierarchical classification of places: planet → continent → country → region → county → locality → neighborhood. The resolver uses placetype to rank candidates — an exact locality match outranks a county-level match. threshold sweep on a violation set 10× this one, plus a no-regression run on the fragment boardsfragment boardThe standing targeted-failure-class eval: ~400 rows per pre-registered class (homonym, bare-street, bare-locality, housenumber variants, particles, date-names) sampled from a national source, graded with Wilson intervals, with the NEGATIVE class scored (a hallucinated street on a bare locality is visible by construction). The FR board (ban-fragments-fr) is the P0 instrument; the DE board (overture-fragments-de) is the v7-fold gate's leg. and the gauntletgauntletThe three-layer input-robustness gate: regression fixtures (named past failures, must-pass), metamorphic perturbations (label-preserving edits must not move the coordinate — INV/DIR/BAND relations), and a held-out fresh draw. Known failures are tracked xfails; a NEW failure blocks. (the bare-localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. and homonym rows are exactly what a score floor kills). The D-ruleD-ruleIron rule 6: no default-on mechanism (model or decode-time) ships with a known regression vs the shipped model on any tier-1 locale — fix, per-locale-gate, or opt-in. The gate every cut passes through. applies: default-on with a known tier-1 regression is not shippable.
• Assessment. This is the cheapest thing that could work and the one most likely to be re-tuned forever. The measurement above says the ceiling is low on the FTS backend.

Design B — abstain on a separate, backend-independent feature set (do not touch scoring)

Leave resolver_score alone permanently — accept that it is ordinal-within-query — and add an explicit abstention decision computed from 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. that already exist and are already on a shared scale: classifier 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. confidence, exactMatch, corroboration (does 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. carry a house numberhouse 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. / streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. / postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. / unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.?), prominence, mismatch, and the top-vs-runner-up margin. Surface the outcome as structured metadata (resolution_abstained + reason) rather than by suppressing the coordinate, so consumers choose the policy — matching the postcode_city_mismatch / resolution_quality idiom already in decorateNode, and matching what the #40 PipelineResult.faults work landed today (report the degrade; do not silently change the answer).

• Cost. Medium. A new decision function in @mailwoman/resolver plus plumbing for two 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. 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. cannot currently see (the classifier's 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. confidence is on the node, so that one is free; corroboration needs the whole tree, which resolveTree already walks). No 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.' retrain, no artifact rebuild, no backend change.
• What it buys. On the measured signals, confidence alone separates at J ≈ 0.92 on both backends with a single shared threshold — an order of magnitude better than any score threshold, and it works identically on both backends because the 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. is upstream of them.
• Risk. The confidence band on clean input is suspiciously tight, so the first shippable version of this is advisory metadata with no default suppression while the threshold is calibrated on real traffic-shaped input. Shipping a suppression rule off a 149-row control would be exactly the "reasoned to instead of measured" failure the house rules warn about.
• Evidence required to promote. (1) The confidence distribution re-derived on a multi-localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. control including fragments, bare localities, and non-Latin scripts — does the 0.918 floor survive? (2) A held-out violation set ≥150 rows, ideally mined from real queries rather than authored. (3) Fragment-board + gauntletgauntletThe three-layer input-robustness gate: regression fixtures (named past failures, must-pass), metamorphic perturbations (label-preserving edits must not move the coordinate — INV/DIR/BAND relations), and a held-out fresh draw. Known failures are tracked xfails; a NEW failure blocks. no-regression with the rule advisory-only, then again with it enforcing. (4) An explicit answer for the one case a confidence rule misses: high-confidence single-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. digit localities (1 → Zona 1 at 0.964), which probably wants the corroboration 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. rather than the confidence one.

Design C — normalize the score into a shared unit

Replace the raw pass-through with a calibrated, backend-declared [0, 1] score: each backend maps its native scale to a probability-like quantity (FTS via a fitted logistic on bm25 + tiers; candidate via min(1, log10(pop+1) / 9.15) or a fitted equivalent), and resolver_score ships the calibrated value with the raw one preserved beside it as resolver_score_raw.

• Cost. High, and it is the only option with a compatibility surface: resolver_score is a published metadata field, browser-cascade.ts tiebreaks on it, and reconcile.ts clamps it. Changing its meaning is a breaking change to a shipped contract, so it wants a major.
• What it buys. One threshold that means the same thing everywhere, and it repairs normalizeResolverScore's saturation as a side effect (a calibrated [0, 1] score is exactly what that combiner was written expecting).
• What it does not buy. Calibration does not create separation. On the measured data the ordering is barely separating; a monotone remap of a barely-separating signal is a barely-separating signal in nicer unitsunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.. Design C is a prerequisite for cross-backend comparison, not a solution to abstention.
• Evidence required to promote. A held-out calibration set per backend (score → observed correctness rate), a reliability curve showing the calibrated value is a probability, and a demonstration that the ordering survives the remap on 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. evalsevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error.. Plus the usual major-version consumer sweep.

If you only take one thing: Designs B and C are complementary, not alternatives, and B does not depend on C. Design C makes the number comparable; Design B makes the decision possible. Doing C alone would produce a well-calibrated number that still cannot abstain.

---

6. What was not tested

• One localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.. en-US weightsparameterA 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., US control set. The confidence distribution is 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.' property and will differ per localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.; the score scales are backend properties and will not.
• One control shape. All 150 control rows are full US streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. addresses. No bare localities, no fragments, no non-Latin script in the control (there is some in the garbage set). The corroboration signal's zero measured cost is an artifact of that shape and should not be quoted without this caveat.
• No FTS postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. database. The FTS leg ran against the admin database alone, so it produced no postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one. nodes and never exercised the coordinate-first Regime B scorer. The [0, 1] regime is read from source, not measured.
• No calibration fitting. Design C's mappings are sketches. No logistic was fitted, no reliability curve drawn.
• No prominence measurement. It is argued for on the strength of its definition (bounded, same meaning on both backends) rather than a measured separation — it was not captured in these runs. That is a one-line probe change and should be the first thing anyone does before adopting Design A or B.
• Small violation set. 14 (FTS) / 18 (candidate) rows clear the no-resolve bar. Every J above should be read as directional.