Explainable does not automatically mean accurate
IPBot exposes the evidence behind a decision so it can be audited. We do not currently publish population-level precision, recall, false-positive rate, or probability calibration because an independently labelled benchmark has not completed review.
No public outcome-accuracy claim
The 0–100 risk score is an evidence-derived model value, not a fraud probability. A smoother score distribution, stable production behavior, or agreement with a model input is not proof of real-world accuracy.
What is measured today
- Contract regression: fixed cases keep score, verdict, action, and explanation semantics stable.
- Safety invariants: protected services, direct threats, special-use addresses, and decision guardrails are checked before release.
- Shadow behavior: candidate models run without changing public responses while distribution, divergence, latency, and policy transitions are observed.
- Source readiness: data freshness, coverage, conflicts, and runtime errors are monitored separately from accuracy.
What remains unmeasured
- Population-level proxy, VPN, Tor, residential-proxy, and abuse precision or recall.
- False-positive rates for real login, signup, payment, or API-abuse traffic.
- Country, city, or coordinate accuracy against independently verified physical locations.
- Probability calibration, transaction loss, retention impact, or third-party validation.
Benchmark publication gate
IPBot now has a reproducible benchmark runner, but it deliberately reports
claim_status=not_approved. A public result requires all of the following:
- Independent labelsModel inputs and provider ranges can test conformance, not independent accuracy.
- Time-bounded provenanceEvery label needs an observation time, expiry, source reference, review state, and license decision.
- Positive and negative coverageHard negatives and abstentions remain visible; unknown is never silently counted as clean.
- Reproducible snapshotsResults pin the API build, data edition, label window, task definition, and IPv4/IPv6 cohort.
- License reviewCommercial databases and third-party comparisons are not published without explicit permission.
- Human approvalPassing automated sample gates never authorizes a marketing claim by itself.
How to read an IPBot result
Treat location as an estimate, network fields as identity context, anonymity fields as detection signals, threat evidence as scoped observations, and the Decision section as versioned policy advice. Do not use one score as the sole basis for a consequential decision.