STATMAP RESOLVE — BENCHMARK AND SERVICE NOTES
Prepared 8 September 2026

ACCURACY COMPARISON
StatMap's recorded August 2026 comparison uses 114,522 prepared Hackney
addresses and the same PublishedGold UPRN labels for MatchGB and local UKAM.
At a precision floor of 99.8%:
                         MatchGB      UKAM 1.2.4
Recall                   99.066%      91.959%
Correct accepted         113,452      105,313
Incorrect accepted       227          210
Difference in correct accepted: 8,139.

Recall is the proportion of reference records correctly accepted.
Precision is the proportion of accepted matches that are correct.
These are operating points meeting a shared minimum precision, rather than
an assertion that the two measured precision values are exactly equal.

UKAM version 1.2.4, commit 6899748f36168d2b5e202289708a8841f7890b99,
uses its native accuracy_analysis precision-recall curve. This is a
StatMap-run benchmark, not an independent evaluation or a comparison
against subsequent UKAM releases. UKAM's own published headline figures
use a different population and are not substituted for our local run.
Version: https://github.com/moj-analytical-services/uk_address_matcher/tree/6899748f36168d2b5e202289708a8841f7890b99

THROUGHPUT
MatchGB completed the 114,522-row Hackney corpus in 104.440 seconds:
1,096.5 addresses per second with one matcher worker, after ingestion.
This lean run calculates all decisions but omits retained top-candidate
audit snapshots. The full-audit counterpart reached 996.2 addresses/sec.
Upload, download and human review are outside this matching timer.
These results are dataset- and configuration-specific, not a guarantee
of accuracy or speed for every customer file. No universal claim of
being faster than UKAM is made.

DEMONSTRATION
The films show 231 original reviewed Mid Sussex address inputs under the
illustrative working filename Sussex Commercial Properties.csv.
The recorded UI initially classified 228 as matched and 3 for review.
Those UI totals are not independently adjudicated accuracy figures and
are separate from the Hackney benchmark above.
Basemap: Ordnance Survey, NativeGB / TinyMM.

CSV AND STREAMING
Resolve preserves input columns and row order. The service appends match
status, accepted UPRN, easting, northing, score, matched address, decision
explanation and constructed query. Only accepted matches receive UPRNs
and coordinates. Consumers should inspect every status, including errors.
Raw text/csv supports simultaneous upload and streamed results. It needs
a compatible duplex client and proxy configuration. Multipart upload is
also supported and validates the file before beginning the response.
Upload and record limits are configurable; CSV size is not unlimited.
The repository defaults are 25 MiB / 1,000 CSV records, and 1,000,000
records for the separate uploaded-job workflow. Deployment settings vary.

NGD MAINTENANCE
The managed offering uses automated daily checks for available licensed
Ordnance Survey NGD Address FULL and Change Only Updates. New generations
are validated before activation; matching continues on a valid generation
during an update. A daily check does not mean OS publishes new data daily.

EVIDENCE
Figures transcribed from the StatMap Resolve repository at d22ab0d:
README.md, UK Address Matcher comparison;
docs/MATCHGB_V4_SINGLE_THREAD_THROUGHPUT_V2.md;
data/benchmarks/results/ukam-1.2.4-hackney-full/precision-recall.csv;
data/benchmarks/results/ukam-1.2.4-hackney-full/metadata.json;
docs/RESOLVE_WEB_API_V1.md;
docs/OS_NGD_AUTOMATIC_UPDATES.md.
Contact StatMap to discuss the evidence and evaluation on your own data.