CAVAL™
Clinical AI Validation, Attribution & Lineage — Interactive Walkthrough
Step 1 of 6 — the clinical encounter
Procedure
Colonoscopy with multiple techniques
A patient undergoes a colonoscopy. During the procedure, the gastroenterologist performs three techniques: cold snare excision, cold forceps excision, and tattooing of the polypectomy site. One procedure. Multiple techniques. The operative report documents all three correctly.
What the operative report says
One colonoscopy. Three techniques applied during a single procedure. The documentation is accurate and complete. A coding AI will now generate CPT codes from this documentation.
Step 2 of 6 — the AI generates erroneous output
AI Output — Coding System
Three separate colonoscopy codes generated
The coding AI reads the operative report and misinterprets technique variation as distinct procedures. It generates three separate colonoscopy CPT codes for the same visit. The codes are plausible. The documentation appears complete. The output is wrong.
"ai_output": {
"tool_id": "coding-ai-v4.1",
"output_type": "cpt_code_generation",
"codes_generated": ["45380","45381","45385"],
"surfaced_at": "2026-05-14T09:14:22Z",
"content_hash": "sha256:c8a1..."
}
"tool_id": "coding-ai-v4.1",
"output_type": "cpt_code_generation",
"codes_generated": ["45380","45381","45385"],
"surfaced_at": "2026-05-14T09:14:22Z",
"content_hash": "sha256:c8a1..."
}
What CAVAL captures at this moment
The AI output is recorded — what coding tool generated it, what codes were produced, the exact timestamp, and a tamper-evident hash. The record is immutable. Three codes were generated. That fact is now permanently captured regardless of what happens next.
Step 3 of 6 — the human reviews and corrects
Human Review — the moment CAVAL was built for
Coder reviews operative report and overrides AI output
A certified coder reviews the operative report. She recognises immediately that multiple excision techniques during a single colonoscopy do not constitute separate procedures. The coding logic misinterpreted technique modifiers as distinct encounters. She corrects the codes.
"human_review": {
"reviewer_id_hash": "sha256:7b3e...",
"review_started_at": "2026-05-14T09:23:10Z",
"ai_output_reviewed": "sha256:c8a1...",
"admissibility": "override",
"codes_corrected_to": ["45385"],
"rationale": "single_procedure_multi_technique",
"review_duration_sec": 497,
"decision_timestamp": "2026-05-14T09:31:07Z"
}
"reviewer_id_hash": "sha256:7b3e...",
"review_started_at": "2026-05-14T09:23:10Z",
"ai_output_reviewed": "sha256:c8a1...",
"admissibility": "override",
"codes_corrected_to": ["45385"],
"rationale": "single_procedure_multi_technique",
"review_duration_sec": 497,
"decision_timestamp": "2026-05-14T09:31:07Z"
}
Step 4 of 6 — what happens without CAVAL
Without the attribution record
The correction happened. The reasoning disappeared.
The claim goes out with the corrected code. But there is no record of what the AI generated, why it was wrong, that a human reviewed it, what the reasoning was, or that the correction was made. The EHR shows one colonoscopy. The billing system was corrected. Everything looks fine.
What propagates downstream — invisibly
Claims data: the original three-code error may have already transmitted
Quality metrics: procedure counts are potentially inflated
AI training data: the erroneous output is not flagged as incorrect
Utilization review: three colonoscopies show in the record
Fraud detection: pattern looks like upcoding
The coding AI makes the same error on the next similar case
The system believes three colonoscopies occurred. Nobody can prove otherwise.
Step 5 of 6 — what CAVAL changes
With the attribution record
The correction is documented. The reasoning survives.
CAVAL captured what the AI generated, that a human reviewed it, how long the review took, what the correction was, and why. That record is tamper-evident, append-only, and exists independently of the EHR, the billing system, and the coding AI that produced the error.
Without CAVAL
AI generated 3 codesunknown
Human reviewed itunknown
Reasoningunknown
Correction madeunknown
System can learnno
With CAVAL
AI generated 3 codes✓ recorded
Human reviewed it✓ 497 seconds
Reasoning✓ captured
Correction made✓ tamper-evident
System can learnyes
Step 6 of 6 — the complete attribution record
CAVAL Attribution Record — Complete
What exists when a payer, auditor, or attorney asks
encounter_idenc_20260514_0092
ai_toolcoding-ai-v4.1
output_typecpt_code_generation
codes_generated45380, 45381, 45385
ai_surfaced_at09:14:22 UTC
admissibilityoverride
codes_corrected_to45385 only
rationalesingle_procedure_multi_technique
review_duration497 seconds
decision_at09:31:07 UTC
record_typetamper-evident
append_onlytrue
vendoragnostic
What this answers
When a payer audits the claim three years from now. When a patient disputes being billed for three colonoscopies. When a compliance officer asks how the coding error occurred and whether it was caught. When the coding AI vendor asks why their model keeps generating this error. The answer exists. The system can learn. Without CAVAL — none of this is possible.
1 of 6