EVIDENCE & VALIDATION

Confidence needs a denominator.

Source-reported retrospective findings, with the cohort, method and limitations kept visible. Evidence reviewed November 10, 2026.

DATASET ACCOUNTING

The corpus is not the blind test.

4,404

Historical cardiovascular claims

3,568

February–August 2026 development claims; 3,545 scored

836

September–October 2026 reserved claims

Within the reserved set: 802 main-cohort claims, of which 794 were scored and eight excluded. Another 34 remote/hospital cases were evaluated separately.

Knowledge freeze: August 31, 2026. First blind scoring: November 8, 2026. Development data informed the rules; it is not independent validation.

TASK-SPECIFIC RESULTS

Keep the unit of measurement visible.

Task	Agreement	Numerator / denominator	Interpretation
Office primary-diagnosis agreement	97.6%	449 / 460	Adjudicated historical agreement
Office non-E/M service-line agreement	99.77%	433 / 434	Judged service lines
Modifier 25 agreement	98.84%	427 / 432	Judged claims; strict 100% gate failed
Stress-test template agreement	100%	69 / 69	Answered templates
Echocardiogram template agreement	98.6%	73 / 74	Answered templates; one revenue-losing case
30-day monitor template agreement	98.5%	66 / 67	Answered templates
Patch / walk-monitor template agreement	100%	47 / 47	47 answered of 48 cases
Correctness of READY parts	99.78%	905 / 907	Parts only; zero complete claims READY

The original autonomous-release verdict: NO-GO.

The study did not pass every safety and operational gate. No complete claim was marked READY. These results do not establish complete-claim correctness, validated ICD-to-CPT medical necessity, authorization success or collections uplift.

METHODOLOGY

Frozen logic. Sealed answers. Disclosed limits.

How the held-out test worked

Rules used knowledge dated at or before the freeze. Creation-time inputs were separated from sealed final claims. Proposals were saved before scoring. A separate critic recomputed result groups.

“Independent” means held-out evaluation with frozen logic. It does not mean a third-party audit or certification.

What limits interpretation

Some holdout aggregate counts were known during design. Payment posting timing and captured note dates weaken aspects of blinding. Remote-monitoring rules were tuned on the rolling-test months. Historical agreement is not authoritative current coding policy.

The original frozen source was retained as hashes and later modified, limiting exact replay. Source reports disclose these weaknesses.

RELEASE CONDITIONS

Good component results are not enough.

SAFETY

Stop conditions must hold

The original test reported four harmful follow-up actions and incomplete interception of mutation patterns.

QUALITY

Review the consequential misses

Revenue-losing omitted service lines, Modifier 25 requirements, scrubber false blocks and excessive false follow-up work remained issues.

VALIDATION

Fresh cases are required

Post-test fixes improved development replays. Already-seen data and spent holdouts are not a new independent evaluation. No successful fresh retest is established by these sources.

SOURCE REGISTER

Trace the evidence.

Review-draft source references require repository access. Approved public evidence editions are needed before public release.

A measured beginning

Choose one workflow.
Establish the evidence. Measure the result.

Start with your practice’s real bottleneck and a clearly defined pilot.

Request a workflow assessment →