Clinical Proof Program
Krasyn has registered a prospective, multi-organization quality-improvement program on clinician review of transcript-derived AI notes. The protocol was written and frozen before any data was collected. This page exists so the protocol can be read before there is anything to argue about.
The proof lock has not started. No organization is enrolled, no baseline period is running, no encounter has been accrued, and no result of any kind exists. Nothing on this page is a finding. When there is a finding, it will arrive with its dates, its denominators, its exclusions, and its limitations attached.
The whole method was fixed in advance and committed to the repository: who is eligible, which encounters count, how they are selected and deduplicated, what is measured, which denominators are used, how missing data is handled, and what may and may not be said about the outcome. Freezing all of that before the first observation is what stops a study from being reshaped after the fact to flatter the product.
Current version registered on 2026-08-29, before any data was collected.
Program identifier clinical-documentation-accountability-qi-v1, protocol version 1.2.0.
Version 1.1.0 (hash fbb5dcf239ef092ebb232fee72bc349f351251b9423e6eef77c1941ad0d6de8f) was superseded on 29 August 2026 after KRA-1960 found that accepted long transcripts were only partially supplied to the judge. Version 1.2.0 closes that hidden-coverage gap by matching both limits at 60,000 characters and prevents a planned excision margin from hiding an omitted pathology-margin status.
Version 1.0.0 (hash 80b749b4d039d65074633f0e95f39815ebd124a053b4647ba66fa6414901d456) was superseded on 28 August 2026, when the first public-benchmark run found two defects in the deterministic layer and fixing them changed its output. No cohort was affected: the proof lock had not started and no organization or clinician was enrolled. Neither superseded version produced a program observation. Nothing was regraded and no engine versions were mixed.
A change that could affect output stops accrual to the affected group. Stored reports are not regraded, engine versions are not mixed, the deviation is disclosed, and a new protocol version with a new hash starts a new period. A safety or security fix ships immediately even when it costs the cohort. Safety outranks finishing the target.
A missed gate, a null effect, an unfavorable confidence interval, a high dismissal rate, or a commercial failure is recorded exactly as it came out. It cannot be rescued by a new denominator, a later batch of encounters, or a subgroup nobody named in advance. The primary dataset is the first block of chronologically eligible encounters after the lock starts, so a later encounter can never replace an earlier one that went the wrong way.
Exploratory analysis is labelled exploratory and cannot raise the claim ceiling. A pooled average that hides a failing practice is not an acceptable report, so organization-level distributions are reported alongside it. Missing outcomes are reported as missing and counted in attrition. Unknown is never converted to zero.
Each rung names the evidence that has to exist before a sentence at that level may be published. A claim never skips a rung, and no amount of evidence at one rung licenses a sentence from a higher one.
Internal protocol, implementation, test, and content-free telemetry assertions.
Participant-observable production service plus reconciled, auditable operational counts under this protocol.
Independent third-party attestation matched to the exact claim, including a written quality-improvement or research determination and licensed-clinician adjudication.
Settled live-mode payments, paid seats, and retained paid organizations through the existing revenue evidence spine.
Until an external determination and a licensed-clinician adjudication are on file, Krasyn publishes no Note Check accuracy percentage, no sensitivity or specificity figure, no error-prevention rate, no patient-outcome improvement, no malpractice-risk reduction, no clinical validation, and no causal time-savings figure. This holds for the program, for benchmark publicity, and for anything a salesperson would like to say.
Clinician dispositions collected during the program are treating-clinician workflow judgements. They are not blinded adjudication, and they cannot be turned into an accuracy estimate. There is no blinded clinical adjudication in this protocol, and the protocol says so in its own text.
Six readiness gates govern the start of the proof lock. Two pass today. Four do not, and until they do, no recruitment can be credited as enrollment and no accrual can begin. Observed 2026-08-27.
The protocol carries a recorded SHA-256 and repository commit, and every proof-lock field is machine-readable.
Note Check sits in the AI surface safety matrix and the tenant, PHI-logging, budget, honesty, failure-state, and clinician-review guards pass. This is an operational fact and not a finding about clinical accuracy.
A production-verified, multi-clinician paid seat entitlement has to bill and meter allowances per seat before any organization can be counted.
The content-free per-encounter ledger is readable and writable in production, but no enrolled producer and no eligible production observation exist. Availability of the store is not an observation.
Review-start, clinician-verified completion, visibility and idle intervals, and structured finding dispositions have to be production-verified without recording clinical text.
Live checkout, seat quantity, paid activation, cancellation, and 30-day retention have to be machine-readable, with internal, demo, free, trial, and complimentary organizations excluded.
The eligibility rules were fixed with the rest of the protocol, so they are the same for the first practice that asks and the last. A practice qualifies when all of the following hold.
Each participating clinician is attested by the clinic as currently licensed and authorized to finalize the notes involved, works from their own organization-scoped account, and completes the program acknowledgement personally. Nobody may accept that acknowledgement on a clinician's behalf, and the clinic attestation creates no public claim about anyone's license.
Transcript text, note text, statement text, omission excerpts, patient identifiers, report contents, and exact corrections stay inside the organization-scoped covered service. The central program ledger carries counts, durations, statuses, and pseudonymous keys. No program or customer clinical content is used to train a model.
The form below adds a practice to an interest list and does nothing else. It does not enroll anyone, it does not open the program, it does not start a baseline period, and it does not cause a single encounter to be accrued. Enrollment happens later, in writing, after the readiness gates pass, and it requires a current Business Associate Agreement, a paid multi-clinician entitlement, and a clinician acknowledgement completed by each clinician.
The separate 50-note documentation QA pilot is a different thing entirely. It is a fixed-scope paid evaluation that stays local to one organization, and it neither enrolls a practice in this program nor contributes evidence to it. You can also read the Note Check product page and the Trust Center first.
Tell us about your practice and we will get in touch when the readiness gates pass and enrollment opens. Sending this form places your practice on an interest list. It does not enroll you, it does not start the program, and it does not cause any encounter to be recorded.
Status
registered, not open
The proof lock has not started and no results exist.
This form
interest only
Recruitment is not enrollment. Nothing starts here.
Before enrollment
gates, BAA, acknowledgement
Four readiness gates are still open today.