Methodology metrics

Published quality numbers

These are the public operating metrics behind the methodology. The point is not to make StreetProof look perfect. The point is to show what has been measured, what is only directional, and what is not ready to quote.

139

public observations

public entities

redacted rows

reviewed candidates

Reporting window: Q3 2026. The machine-readable version is published at /api/v1/methodology/metrics.

Headline metrics

Metric	Value	Sample	Status	Notes
Entity-match review support rate	not yet reportable	0	not yet reportable	No reviewed candidate sample is available in the current live database.
Review overturn rate	not yet reportable	0	not yet reportable	Reviewed machine candidates rejected by a human reviewer. Merged candidates are tracked separately from outright rejections.
False-positive correction rate	0%	139	reportable	Redacted observations divided by public plus redacted observations. This captures material corrections, not every low-confidence candidate stopped before publication.
Average time to dispute resolution	not yet reportable	0	not yet reportable	No resolved removal or dispute requests are present in the current live database.
Redaction count	0	139	reportable	Observation rows hidden from public surfaces through the audited redaction path. Rows are not deleted from the append-only ledger.
Average candidate review age	not yet reportable	0	not yet reportable	Average elapsed time from entity-match candidate creation to human review. This is an operations metric, not a model accuracy metric.

OCR precision and recall

Precision sample

A labeled OCR holdout set has not closed yet. Current public numbers show OCR confidence distribution only; precision will publish once labels are available.

Recall sample

Recall needs a labeled set of visible text spans. StreetProof will not infer recall from model confidence.

Current live observations with OCR confidence: 139. Average OCR confidence: 0.942. Confidence is useful triage data; it is not the same thing as precision or recall, so those fields stay unreported until a labeled sample exists.

Entity-match sample

Entity matching is the trust-critical step. The public number below is an operations proxy from reviewed candidates, not a blinded model benchmark.

Status	Rows	Reviewed	Avg confidence
No candidate sample yet.

Coverage by city

Coverage confidence decides how strongly a “not observed” result can be interpreted. Low coverage means the honest answer is: StreetProof has not seen enough yet.

City	Confidence	Cells	Public obs	Public entities	Latest
Calgary calgary	low	97	139	12	2026-06-30
Edmonton edmonton	none	0	0	0	-
Red Deer red-deer	none	0	0	0	-
Reykjavik reykjavik	none	0	0	0	-
Saskatoon saskatoon	none	0	0	0	-

Limits

Pre-launch data may include synthetic fixture records where explicitly labeled elsewhere on the site.
OCR precision and recall require a labeled holdout set; model confidence is not treated as ground truth.
Entity-match support rate is an operational proxy until blind review samples reach reportable size.
Redaction count captures published-row corrections; candidates stopped before publication are not false positives.
Resolved dispute redaction rate is not yet reportable for the current live database.

Related pages: methodology, ledger transparency, and corrections.