The interpolation radius factor is regional — 1.70 is a Texas artifact
2026-06-14. The conformal interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable.-radius factor (Q̂ = 1.70, #569) was calibrated on one regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. (Texas / Travis County E-911). This re-gates it on four more statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. spanning the density spectrum, using each stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.'s situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. address pointssitus dataA dataset of exact address-point coordinates (rooftop-level). Mailwoman's geocoder uses a national situs layer (124.9M US points built from state address-point sources) as the highest-precision tier of the geocode cascade. (OA/NADNAD (National Address Database). A US Department of Transportation dataset of structured address points, added to the training corpus as a major source of real US addresses.) as independent ground truthground truthThe correct answer for an example, used as the standard a prediction is graded against. Mailwoman's ground truth is the hand-labeled golden set; its quality caps achievable accuracy. for the TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable. tiertierInternal versioning of which label classes the model emits. Tier 1 is the coarse components (country, region, locality, postcode); Tier 2 adds venue, street, house_number; Tier 3 (future) would add attention, po_box, and POI venue subtyping. Historically called 'Stage 1/2/3' before the runtime-pipeline naming made that ambiguous. — a non-circular holdout available for all 50 statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.. The factor is not regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.-invariant: it runs 1.53× (dense NY) to 2.85× (rural MT), a ~2× spread that tracks rurality. The shipped single 1.70× is overconfident in rural America and over-conservative in dense cities. The parked "per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. vs single-factor" decision resolves toward per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.; this records the evidence, ships a seed calibration table + reusable tooling, and flags the wiring decision.
Why a second look
#569 turned the heuristic interp radius (half the matched TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. length) into a conformally calibrated 90%-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. interval: multiply the claimed radius by Q̂ = 1.70 and you cover 90% of held-out error. But that Q̂ came from a single regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.. A confidence interval calibrated on Austin and shipped nationwide is only honest if the factor generalizes — and there's a physical reason to doubt it does: TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. geometry and address-point spacing differ between Manhattan and rural Montana, so the ratio of true error to segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. length (what Q̂ captures) may differ too.
Method — non-circular, by construction
For each stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.: synthesize a {input, lat, lon} holdout from the stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.'s situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. shardshardA partial output file of the corpus build, written in Parquet format. The training pipeline streams shards row by row. (OA/NADNAD (National Address Database). A US Department of Transportation dataset of structured address points, added to the training corpus as a major source of real US addresses.
address pointssitus dataA dataset of exact address-point coordinates (rooftop-level). Mailwoman's geocoder uses a national situs layer (124.9M US points built from state address-point sources) as the highest-precision tier of the geocode cascade. — the ground-truth coordinates), then run conformal-calibrate.ts interp-only (the
situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. tiertierInternal versioning of which label classes the model emits. Tier 1 is the coarse components (country, region, locality, postcode); Tier 2 adds venue, street, house_number; Tier 3 (future) would add attention, po_box, and POI venue subtyping. Historically called 'Stage 1/2/3' before the runtime-pipeline naming made that ambiguous. no-op'd via the #568 tableless guard) so every resolved row is a TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable.,
scored against the true situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. coordinate. TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. (Census streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. ranges) and OA/NADNAD (National Address Database). A US Department of Transportation dataset of structured address points, added to the training corpus as a major source of real US addresses. (address pointssitus dataA dataset of exact address-point coordinates (rooftop-level). Mailwoman's geocoder uses a national situs layer (124.9M US points built from state address-point sources) as the highest-precision tier of the geocode cascade.) are
independent sources, so this is non-circular — the same provenance separation the Texas/Travis
calibration relied on, now available 50× over from the national situssitusThe physical site address of a property, as opposed to the owner's mailing address. Parcel records often carry both; the divergence is a real-world data-quality challenge. build. n ≈ 1500 interp hits per
stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality., 50/50 cal/test split, α = 0.90.
Result
| stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. | character | interp Q̂ (90%) | 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. @ Q̂ | uncalibrated (Q̂=1) | median err | median claimed r |
|---|---|---|---|---|---|---|
| NY | dense urban | 1.53 | 90.6% | 79.7% | 45.2 m | 104 m |
| TX | urban (Travis) | 1.70 [#569] | — | — | — | — |
| CA | large / varied | 1.87 | 91.0% | 76.1% | 40.6 m | 82 m |
| MI | mid | 1.93 | 89.7% | 77.1% | 51.1 m | 113 m |
| MT | extreme rural | 2.85 | 90.2% | 62.7% | 61.3 m | 83 m |
Every stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.'s own Q̂ lands 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. within 3pp of 90% — the conformal machinery is sound. What varies is the factor itself, monotonically with rurality: dense NY needs 1.53×, extreme-rural MT needs 2.85×. The uncalibrated column shows why MT is the extreme — its raw radius covers only 62.7% as-is (rural addresses are spaced far less uniformly along their long TIGERTIGERThe US Census Topologically Integrated Geographic Encoding and Referencing database. Used as a corpus source for street-segment data. segmentssegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. than the interpolationinterpolationA geocoding technique that estimates a coordinate along a street segment based on the house number range. Used as the middle tier of Mailwoman's geocode cascade when exact address-point data is unavailable.'s uniform-spacing assumption allows), so it needs the biggest correction.
Wider sample confirms it (12 states)
A partial 50-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. sweep (abandoned at the >85 °C heat ceiling on the lab CPU — a sustained run of the neural cascade) extended the five above to twelve, and the trend holds and widens. Ordered by Q̂:
| Q̂ | statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. |
|---|---|
| 1.4–1.6 | DC 1.44, NY 1.53 (densest urban) |
| 1.7–2.0 | TX 1.70, AK 1.72, CA 1.87, CT 1.91, MI 1.93 |
| 2.2–2.9 | AR 2.24, CO 2.29, AL 2.79, MT 2.85 |
| 3.0+ | AZ 3.12 (sprawl / long rural segmentssegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.) |
A 2.2× spread (DC 1.44 → AZ 3.12), monotonic with rurality. The five-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. read wasn't a small-sample
fluke — it's the real shape. The seed table (data/calibration/interp-radius-conformal.json) carries all
twelve; the full 50 is a turn-key follow-up (mind the heat ceiling on a sustained sweep).
What the shipped 1.70 actually does off-Texas
A single nationwide 1.70× is wrong in both directions, and one direction is dangerous:
- Rural statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. are overconfident. MT needs 2.85× for 90%; at 1.70× the radius claims a 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. it doesn't have — a user is told "90% within R" and gets materially less. Overconfidence is the failure mode honest confidence exists to prevent.
- Dense cities are over-conservative. NY needs only 1.53×; at 1.70× the radius is wider than it needs to be — honest, but it throws away 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. the data supports.
Texas (1.70) sits in the middle, which is exactly why a single-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. calibration looked fine and shipped — the artifact is the regional mean masquerading as a constant.
Decision and recommendation
Resolve the parked decision toward per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.. Two shippable shapes:
- Per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. table (seed shipped here).
data/calibration/interp-radius-conformal.jsoncarries the five measured statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. + a conservative default of 1.95 for unmeasured statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. — deliberately on the high side, because under-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. (overconfidence) is the harmful error and most statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. skew rural.geocode-corewould load the factor by parsed regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. instead of the hardcoded 1.70. Cheap; the full 50-stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. table is a turn-key follow-up (the tooling is committed — ~2 min/stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.). - Per-segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.-length bucket (the principled refinement). The real driver is segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context. length / local density, not the stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. line — a stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. like CA holds both dense LA and rural North StateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.. A Q̂ indexed on the claimed radius (segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.-length bucket) would generalize within a stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. and to unmeasured statesregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality.. More work; the better long-term answer. Recommended as the follow-up to the seed table.
Flagged, not auto-wired. Loading a per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. factor changes the shipped uncertainty_m on every
interpolated geocode — a behavior change. Per the merge-wall discipline this is PR-and-flag: the seed
table + this evidence land; the operator decides whether to wire per-regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. now (table) or hold for the
per-segmentsegmentA punctuation-bounded chunk of the normalized input — the comma-separated parts of 'Portland, OR' — used to give downstream stages structural context.-length version. The single 1.70 stays the default until then — with this report on record
that it under-covers rural geocodes.
Reproduce
node scripts/eval/build-situs-holdout.mjs --shard <state-situs.db> --region <ABBR> --n 2000
node scripts/eval/conformal-calibrate.ts \
--holdout /tmp/<abbr>-situs-holdout.jsonl \
--address-points /tmp/empty-situs.db \ # tableless → situs no-op → interp-only (#568)
--interpolation <state-interp.db>
# or scripts/eval/run-conformal-multistate.sh for the whole sweep
See also: 2026-06-14-interp-radius-calibration.md (the
original Texas 1.70 calibration, #569).