Does context earn its keep?

The comparator.

The baseline is a simple visibility rule such as “no visit in 60 days.” That rule identifies a record to check; it is not treated as an instruction to contact a patient.

The KeyRD question.

For a flagged record, what does this person need now? Is follow-up already booked? Is there a reply waiting? Is care paused, completed, or awaiting intervention review? Is a visit limit relevant? Is the information current enough to decide?

A simpler workflow that checks these same facts may reach the same decision. The current work tests whether KeyRD’s extra context is useful enough to justify itself, not whether complexity is automatically better.

What an owner should get from an initial evaluation.

A bounded retrospective should leave the practice with a sample-based picture of continuity work: which sampled active-care patients have no next visit, which gaps need no action, where manual checking happens, what the platform exposes reliably, and what preparation, review, correction, and remaining execution still cost.

It is not a complete caseload audit unless the agreed data actually covers the whole panel.

What saying yes involves.

The discovery conversation comes first: about 50 minutes, no preparation, installation, or patient records, and no commitment to product evaluation.

Any later record-based evaluation receives a written scope before the practice agrees. The scope must specify sample, expected preparation and review hours, duration, any fees, who reviews what, where records and outputs reside, and what deliverables the practice receives.

Dated protocol summary.

Public summary v1 · September 9, 2026. This describes the existing blind-first retrospective design, frozen September 5, 2026. It is not a registry entry or a claim of independent scientific, ethics, or institutional approval. Final practice-specific scientific review and a signed evaluation scope remain pending.

Population and setting.

The initial design concerns one dietitian’s practice using individual Practice Better client exports: people with documented continuing nutrition care and enough dated information to judge the selected historical snapshot. The practice’s complete eligibility, exclusion, and roster-ordering rules must be recorded before selection. This is separate from discovery interviews with owners of multi-RDN, insurance-billing practices using Healthie.

Sampling and historical cutoff.

The plan uses a consecutive, neutral roster block and the first 40 eligible unique clients when available, including quiet and unflagged cases. Every omission and reason is recorded. Fewer than 40 is a feasibility subset, not an equivalent planned sample. The sample must not be extended after seeing results.

If more than 40 records are supplied, the existing design uses deterministic sampling stratified by KeyRD classification, with all safety and re-engagement candidates included. This can exceed 40 and creates a selected comparison sample, not an estimate of practice-wide prevalence. The source frame and this selection consequence require explicit review before use.

The frozen snapshot is September 4, 2026, at 23:59:59 UTC. Undated notes and later information cannot establish what was known then. Booking creation and update history may be unavailable; uncertain historical status requires an insufficient-information judgment. A changed cutoff requires a documented amendment before analysis, not a silent update.

Comparator.

The fixed rule flags a last documented completed visit at least 60 calendar days before the cutoff when no next appointment is documented. A documented next appointment suppresses the flag. A missing or invalid last-completed-visit date produces insufficient information. This is a visibility rule only; it grants no authority to contact a patient. Both methods use the same extracted facts, so shared parsing errors can affect both.

Source-first review, then reveal.

In pass one, the dietitian reviews source records without KeyRD recommendations, derived summaries, or comparator flags. Each case receives a reason and one of four classifications: no action, routine attention, RDN judgment needed, or insufficient information. All selected first-pass judgments are locked before any recommendation is revealed.

In pass two, the same dietitian reviews KeyRD’s proposals and records acceptability, concerns, safety, and whether the proposed response is the minimum sufficient one. First-pass answers cannot be edited after reveal. Prior exposure is logged and interpreted separately. This prevents recommendation priming in the first pass; it does not provide an independent second clinician or eliminate familiarity and anchoring in the second pass.

Outcomes and missing information.

The primary paired comparison is needs-attention versus no-action against the locked source-only reference. Routine attention and RDN judgment needed both count as attention. Report agreement, correctly identified attention and no-action cases, false attention, missed attention, and the coverage of the comparison.

Cases with an unknown reference or comparator result are excluded from both methods’ paired denominator. A KeyRD insufficient-information output on a remaining case stays in that denominator and is not counted as agreement; its handling and separate reporting must be confirmed in the final analysis review. Missing information is never evidence that no care occurred. Suppressed or unavailable results are not zero.

Review burden and disclosure.

Time each review pass and correction explanation, pausing for interruptions. Record roster selection, individual exports, compatibility checks, and setup separately; identify measured logs versus recalled estimates. Expected preparation and review hours must be agreed in the written scope. Subsequent care, scheduling, and other execution are not performed or timed.

Do not subtract second-pass from first-pass time to claim savings: the tasks differ and the reviewer already knows the cases. Do not extrapolate sample times to the whole panel. Only disclosure-controlled aggregate results may leave the practice. At least 11 completed client reviews are needed for aggregate export, and small cells and revealing complements are suppressed. These are disclosure rules, not statistical power or efficacy thresholds.

Interpretation rules.

No numerical efficacy or acceptability pass threshold was frozen. Classify the result as usable directional evidence, comparable but no incremental signal, or inconclusive/invalid. If KeyRD provides no substantiated useful distinction beyond the comparator after accounting for review burden, there is no basis in this sample to prefer it. That does not prove equivalence or establish that every simpler workflow is sufficient.

A favorable result remains one-clinician retrospective feasibility evidence. It cannot establish improved retention, adherence, clinical outcomes, safety, causal time savings, revenue, or adoption. A material integrity problem or unsafe proposal requires investigation before any prospective use. Market interviews cannot rescue an invalid study, and a favorable study cannot establish market demand.

What remains unresolved.

Before real-data execution: confirm the practice’s signed scope and scientific review; eligibility and roster ordering; feasible sample and export burden; historical source fidelity; explicit reporting of insufficient-information outputs; applicable data agreements and ethics determinations; and the completion timetable. Independent second-clinician adjudication is not currently staffed in this design. This page authorizes no record access, patient contact, or clinical-policy change.

What is not established yet.

KeyRD has not established improved continuity, adherence, clinical outcomes, workload reduction, economic benefit, or superiority to a simpler workflow.