Methodology and measured precision

Every number on this page is measured on labeled synthetic ground truth. None of it is field performance. Field performance requires a calibration pilot on the customer's own data, and until that pilot has run, we will not quote one.

The full rule table

27 of 28 rules have measured precision. The table is sorted by precision, and the low rows are published on purpose. A rule that is not ready is labeled, not hidden. Rules version: 2026-07-22-w4-sparse-referral.

All 27 measured rules All precision figures are measured on synthetic ground truth, not on field data.
Rule Precision measured on synthetic ground truth Labeled cases True positives False positives
New provider anomaly 1.00 synthetic ground truth 40 30 0
Impossible day hours 1.00 synthetic ground truth 388 6 0
Deceased patient 1.00 synthetic ground truth 315 285 0
Procedure frequency outlier 1.00 synthetic ground truth 340 1 0
Cci unbundling 1.00 synthetic ground truth 24 8 0
Referral loop 1.00 synthetic ground truth 1,046 4 0
Consent defect 1.00 synthetic ground truth 48,854 3,820 0
Note provenance 1.00 synthetic ground truth 800 60 0
Rpm coherence 1.00 synthetic ground truth 400 80 0
Authority scope 1.00 synthetic ground truth 100 66 0
Udt mill 1.00 synthetic ground truth 210 180 0
Hcc inflation 1.00 synthetic ground truth 150 30 0
Dme mismatch 1.00 synthetic ground truth 20 10 0
Wound care hospice 1.00 synthetic ground truth 16 8 0
Document fingerprint 1.00 synthetic ground truth 40 10 0
Coercion escalate 1.00 synthetic ground truth 40 15 0
Telemetry liveness 1.00 synthetic ground truth 325 116 0
Device signature index 1.00 synthetic ground truth 325 109 0
Missing pathology 0.98 synthetic ground truth 312 103 2
Pathway expectation 0.80 synthetic ground truth 282 33 8
Upcoding distribution 0.75 synthetic ground truth 2,974 3 1
Cleaned scheme 0.73 synthetic ground truth 290 80 30
Panel overutilization 0.50 synthetic ground truth 116 1 1
Missing anesthesia 0.48 synthetic ground truth 3,646 1,219 1,303
Missing facility 0.33 synthetic ground truth 1,352 670 1,360
Pos mismatch 0.15 synthetic ground truth 154 16 89
Missing post op pt 0.10 synthetic ground truth 1,970 267 2,477

Field performance is different from synthetic performance. Before any production use we run a calibration pilot on the customer's own data and report those numbers instead.

How to read these numbers

Precision answers one question: when this rule flags a claim, how often is the claim actually bad? A precision of 1.00 means every flag in the labeled set was a true positive. A precision of 0.48 means roughly half were not, and that rule is not ready to act on alone. We publish both because a buyer who only ever sees perfect numbers is being managed, not informed.

The corpus

The test corpus contains 521,997 synthetic claims across 2,000 synthetic providers. It is generated from a seeded, deterministic generator, which makes it byte-reproducible: the same code and the same seed produce the same corpus, byte for byte. Fraud archetypes are injected with known labels, which is what makes ground truth possible.

It contains zero PHI. No real patients, no real providers, no production claims data. Every name, identifier, and record in it is synthetic.

What we do not verify

  • Medical necessity. The rules check whether records contradict each other, not whether care was clinically warranted.
  • Clinical accuracy of readings. We verify that readings exist, cohere, and match the device on file. We do not verify that a blood pressure value was medically correct.

Every bundle repeats these limits in its own gap list. An artifact that overstates what it checked would not survive expert scrutiny, and it should not.