Why Transcript Feedback Needs Longitudinal Patient State

A transcript and a session rubric are valuable, and they are incomplete. The transcript says what was asked, explained, and left unsaid in one encounter. The rubric says how that encounter was judged. Neither one is the patient. Feedback that is supposed to matter at a later encounter needs longitudinal state: what changed, under which rules, against which evidence version, in a history that can be replayed. Without that, the feedback is a comment on a document that the next session is free to ignore.

What a transcript is good for

A transcript is a record of the interaction. In MediTutor's current product, healthcare learners meet synthetic patients by voice or text, and the session produces a transcript alongside formative scoring. That is a better review object than a memory of the conversation. A reviewer can point at a turn. The learner can see the words they used. Formative feedback can be tied to those words instead of to a general impression.

Transcripts are not training fuel for a foundation model in this account. They are material for formative feedback and for product design. Any use of them for training or evaluation is subject to applicable consent, data rights, de-identification, and governance controls. A transcript also does not become longitudinal just because it is long. Length is still one sitting. When the sitting ends, the current product's case context ends with it.

What a session rubric can and cannot see

A session rubric can mark coverage, omissions, and the quality of an explanation against criteria for that case. It can be reviewed. Session-level scoring is what the shipped product does. Expert comparison across encounters, and a rubric that reads a carried state, are the direction being built. They are not already delivered.

The rubric's blind spot is time. It can say that a fact was not asked for in this transcript. It cannot say what that omission did to the next encounter unless the next encounter exists as state rather than as a new vignette. A missed item on a score sheet is a label. The same miss, stored as still undisclosed, is something a later encounter can meet. Feedback that only has the label is teaching the transcript. Feedback that has the state is teaching the consequence inside the simulation. Neither version is a claim about real patients.

State is the delta, not a summary

Longitudinal state is not a paragraph that begins "previously on this patient." A summary is another text, and text can drift. A state delta is the set of changes the world accepted: which facts were disclosed, which commitments were written, what the clock did, and which allowed changes the scenario actually recorded. If the scenario did not record them, they did not happen, no matter how moving the dialogue was.

Governed transitions are what keep that delta honest. The world should refuse an update that the rules do not allow, including a model's impulse to improve the story. The current session product can constrain a reply inside one case. The runtime being built has to constrain the write between cases. Otherwise the second encounter is just a chat that was shown a synopsis.

Evidence versions belong in the same record. Feedback that cites a guideline or a case source is citing something that can change. If the evidence pack is updated and the old feedback is not pinned to a version, a later reviewer cannot tell whether the learner missed a criterion that existed at the time. Why Every AI Patient Run Needs a Simulation Receipt is the same discipline applied to a whole run: name the versions, or do not pretend the judgment is repeatable. A rubric without a version is a rubric that will silently move.

The later encounter is the test

The test of longitudinal feedback is rude and simple. Start a second encounter from the stored state, not from the original vignette. Then ask whether the feedback from the first encounter is visible in the world. If the learner omitted a question, is the fact still hidden? If they explained a plan, did any allowed state change follow, or was the explanation only words? If the case records something like trust, did it move under a rule, or did the next encounter start from a default the first encounter never wrote down?

If the second encounter cannot answer those questions from state, the first encounter's feedback was a session artifact. That is acceptable for one-sitting practice. It is not longitudinal. MediTutor's early sessions were one-sitting practice at scale: more than a thousand completed simulations by 31 August 2026, useful as an operating snapshot, and explicitly not clinical validation. They helped show that session feedback stops at the transcript. The design response is described in From MediTutor to Patientic. It is a persistent patient, not a longer score report.

encounter transcript
  questions, explanations, omissions
        |
        v
state delta
  disclosed facts, commitments, clock
  only the changes the rules allowed
        |
        v
later encounter
  reads the stored state
  does not restart from the original vignette
        |
        v
feedback and replay
  rubric cites the transcript and the delta
  replay walks the same events and versions

The diagram is the dependency order. Feedback that skips the middle and jumps from transcript to advice has nowhere to attach a later encounter. Replay that skips the versions cannot show that the delta was produced by the same rules.

Replay is how the feedback stays checkable

Replay is the property that makes longitudinal feedback auditable. Someone who was not in the session should be able to walk the events, see the state before and after, and see which evidence version the rubric used. If the only artifact is the transcript plus a score, a disagreement becomes a debate about tone. If the artifact includes the delta, the disagreement can be about whether a transition was allowed and whether the rubric cited the right event.

That standard is stricter than "the feedback sounded fair." Fair-sounding feedback is cheap. Checkable feedback is a record. The record is still a simulation record. It can show that an omission was carried, or that a transition was refused. It cannot show that the same omission would harm a real person, and it must not be written as if it could.

What remains intentionally incomplete

A transcript should stay. It is the evidence of the interaction, and throwing it away to "simplify" the state would hide the words a reviewer needs. A session rubric should stay. It is how an encounter has criteria at all. Longitudinal state does not replace either one. It gives them a patient to attach to, a versioned evidence pack to cite, and a history that a later encounter and a replay can both read.

Until that attachment exists, the honest feedback is session feedback. Calling it longitudinal because the learner came back tomorrow, and sat down with a reset case, would be a labeling error. The state has to be the thing that returned.

Related reading

• What a Thousand Patient Simulations Helped Us Notice About Medical AI Interaction
• What Is a Persistent Virtual Patient?