# Correction Across Clinical Handovers

## An evaluation design for agent-assisted learning health systems

**Luuk Joseph Soons**  
CounterProof Research, a practice of Clavestra Capital Limited, Malta  
ORCID: [0009-0000-3088-6373](https://orcid.org/0009-0000-3088-6373)  
**Conceptual methods proposal · preprint v0.15 · 5 November 2026**  
Deposited on Zenodo on 5 November 2026: [10.5281/zenodo.23974125](https://doi.org/10.5281/zenodo.23974125). This web page is a reading copy of the deposited text.

*Not peer reviewed. No clinical study or proposed experiment reported here has been conducted. AI assistance and author responsibilities are disclosed below.*

### Abstract

A report is corrected, but the next discharge summary repeats the original statement. A follow-up appointment is booked, then cancelled, yet a later reader treats the care as complete. These hypothetical examples illustrate an established clinical safety concern: updated information must reach those responsible for subsequent care. This paper asks whether an additional correction workflow helps when clinical work passes between people and AI assistants.

We propose a workflow that links clinical statements and their supporting evidence, decisions about care, and separately observed patient outcomes. It records why a proposed correction was accepted, rejected or left unresolved, who is responsible for the next step, and what later records show. It builds on existing learning health systems, health-agent research and clinical information standards.

Study A would compare two ways of organising the same model's work on clinician-constructed simulated handovers, using the same released source information: competent simpler practice and the proposed structured correction workflow. Both would look for evidence that could change a conclusion. Success would require a worthwhile improvement in the use of justified corrections. The workflow would also have to meet requirements set in advance for rejecting unsupported changes and limiting mistaken revisions, harmful changes and workload. The comparison would assess the whole workflow; it could not isolate the benefit of deliberately seeking counterevidence (adversarial inquiry) itself. If Study A justified them, a second study would test whether another model or a clinician adds worthwhile review, and a third would examine a defined hospital pathway, checking what care actually occurred separately from the workflow's completion labels.

No proposed study has been conducted, and no clinical benefit is established. The contribution is an evaluation design. If simpler practice works as well, or the added procedure causes unacceptable harm or burden, the procedure should be revised or abandoned.

**Keywords:** learning health systems; clinical handover; amended results; patient safety; AI agents in healthcare; evaluation design; agentic epistemics; adversarial inquiry

## 1. The problem: a correction can be recorded and still be lost in practice

A clinician corrects a report. The amendment remains in the record, but a later discharge summary repeats the original statement. Another professional, or an AI assistant, uses that summary in a referral. The correction was stored, but it did not change the next decision. This is a hypothetical example of the failure this proposal would investigate.

### 1.1 Why a hospital would care

Published clinical observations establish the stakes. A PSNet case describes an amended laboratory result that was not communicated to the responsible physician. A senior clinician noticed the discrepancy before discharge, changed the plan and prevented the communication failure from causing harm [46]. This is a reported safety case, not a prevalence estimate.

In a study at one tertiary hospital, 147 of 686 patients recalled after emergency-department discharge because preliminary and final imaging reports differed received inpatient management [42]. The hospital already had a liaison and notification process. These selected cases show that revisions can require consequential care; they do not estimate how often all hospital patients experience a failed handover.

Notification and follow-through are different outcomes. In a cluster-randomized trial of automated notifications for results pending at discharge, documented action on 253 actionable results was 60.7% with notifications and 56.3% with usual care (p=0.82) [43]. Time to action improved among results with documented action. The trial does not establish equivalence, and its record review could miss care outside the network. Its experience illustrates why an alert alone cannot be assumed to resolve a follow-up problem.

These sources support investigation of an important, established safety problem. They do not establish that this precise failure is currently acute in most hospitals, or how often an acknowledged correction is subsequently lost across human and AI handovers. That narrower question needs measurement in the proposed setting.

### 1.2 Why an obvious problem can persist

Correcting a source record, informing the next team, interpreting the change, assigning responsibility and checking what happened are separate tasks. Shift changes, different record systems and care outside the originating organisation make their coordination harder. Existing SAFER guidance already addresses amended-result communication across care transitions, visible amendments and accepted transfer of responsibility [44]. The paper treats these as established clinical practices to build on and compare against.

The appropriate response also depends on circumstances. A recommendation may no longer be indicated, a patient may make an informed refusal, or an appointment may be unavailable. An unanswered alert cannot distinguish these situations. Better information may help, but it cannot itself supply staff, transport or treatment capacity. The hospital must identify which failure it is trying to reduce.

### 1.3 What AI might change

AI introduces another producer and consumer of clinical accounts. In a retrospective evaluation of 100 hospital encounters, GPT-4 Turbo discharge narratives were rated of comparable overall quality to physician narratives, more concise and coherent, but contained more errors, particularly omissions [45]. This document evaluation did not measure downstream care or actual patient harm. It supports checking whether fluent summaries retain important information, rather than assuming that readability establishes accuracy.

An agent could reuse a stale summary, omit the reason for an amendment, or treat a booking as completed care. Another agent could then repeat that account. These are mechanisms to test, not measured rates of hospital harm. AI could also help detect contradictions and carry corrections forward [12,31]. Whether it increases or reduces failures depends on the workflow and must be evaluated.

Our overarching practical question is:

> Can AI help clinical teams use corrections reliably across handovers?

**This proposal asks five practical questions:**

1. How can we ensure that the next clinical team uses the corrected information?
2. Can the workflow recognise and reject an incorrect proposed correction?
3. Can it distinguish planned care from care that actually happened?
4. Does it respect the patient's decisions, including a decision to decline care?
5. Does it work better than the hospital's current process, without adding excessive work?

Section 6 turns these into measured comparisons. Study A addresses questions 1 to 4 on simulated handovers: justified-correction uptake across successive handovers (question 1), invalid-revision rejection and conditional unsupported acceptance (question 2), the classification of downstream actions (question 3) and the patient-decision case family (question 4). Question 5, and questions 1 to 4 against a hospital's own process, can only be answered in a hospital pathway (Study C), which would run only if Study A justified it.

Study A would compare the same model on clinician-constructed simulated handovers using identical released sources. Both workflows would update summaries and check sources, including contradictory evidence. The proposed workflow would additionally link statements, amendments and care states, recording why suggested changes were accepted, rejected or left unresolved. Study C would examine a hospital pathway if earlier work justified it. The hospital comparator must include functioning record updating, tracking and human coordination. Any added procedure must justify its errors, delays and workload; competent simpler practice may already suffice.

The workflow involves clinical staff, patients, administrators and their systems. AI would perform specified tasks within that organisation. Recording a recommendation would not give it authority to act. A successful document check would not by itself establish better care or health. Section 5 provides a worked example and Section 6 specifies the comparisons.

The paper belongs to CounterProof's Agentic Epistemics programme, concerning evidence, correction and useful knowledge [1]. Its related work examines institutional continuity, adequate reporting and the origins and dependence of claims [2,3,4]. Section 3.1 explains those connections. The proposed contribution is a comparative evaluation, with evidence, care actions and patient outcomes kept distinct. The programme and its components do not establish uniqueness or clinical benefit.

## 2. How this proposal was developed

This conceptual methods proposal draws on a targeted narrative literature search begun on 3–4 November 2026, followed by checks of its cited claims and a further comparison with close predecessors on 4 November. The search covered learning health systems, longitudinal agents, clinical information standards, report quality assurance and programmes that track follow-up to resolution. The v0.13 update added clinical studies of revised reports and post-discharge results, amended-result safety guidance and an evaluation of AI-generated discharge narratives, and rechecked close agent-memory predecessors. Primary research, standards and official guidance were preferred. Source access ranged from full texts to abstracts and indexed extracts; the supplement records those differences and the qualifications they require.

This was a selective search, not a systematic review. It had no registered search protocol, comprehensive database coverage or independent duplicate screening, and it does not claim to have found all relevant work. No clinical data were collected, no dataset was reanalysed and no model experiment was run. The paper addresses evaluation of the correction workflow. A short historical discussion motivates questions for evaluation; it is not a history review or evidence of the proposed method's effectiveness. Prevention and the engineering of external record anchoring remain outside the paper's scope.

## 3. What existing work provides, and what this paper would test

Learning health systems already connect clinical practice, evidence and improvement [5]. Their ethical frameworks predate current AI agents [6], and recent work considers AI-supported feedback and organisational action [7]. The proposed programme builds on those traditions. Piecemeal change is also established: Plan-Do-Study-Act cycles test and revise local improvements [29].

Several close predecessors define what this paper can contribute:

| Existing work | What it provides | Question for this proposal |
|---|---|---|
| I-PASS, Measure Dx, staffed follow-up programmes and SAFER guidance [27,28,14,44] | Structured handovers, diagnostic safety learning, tracking, amended-result communication and named responsibility | Does the additional correction procedure improve on a functioning clinical process? |
| Longitudinal health agents and clinical discrepancy reconciliation [8,31] | Revised interpretations, continuing responsibility and source-linked checks of inconsistent information | Does adjudicated correction remain appropriate across subsequent handovers? |
| ClinTraceBench [32] | Controlled comparisons of clinical-history representations, including retrieval of earlier evidence for a later diagnosis | What does an explicit correction and disposition record add under common source releases? |
| Memory Echo and dependency-guided repair [35,36] | Source-aware memory correction and downstream repair in agent evaluations | Can the clinical workflow assess whether a proposed correction deserves acceptance, rather than receive its validity as a supplied annotation? |
| Dialectic-Med and AIGS [33,34] | Explicit adversarial clinical reasoning and automated falsification in research | Does the specified procedure help under the proposed comparison? Falsification itself is not introduced here. |
| HALO, VERSION-AI and the sociotechnical harm pathway [30,9,37] | Institutional oversight, proportionate reassessment and links between AI behaviour, workflow and consequences | Which corrections change later decisions and care, and at what cost? |
| FHIR workflow/provenance and HL7 transparency work [10,11] | Representations of tasks, responsibility, revisions, source links and AI involvement | Are represented corrections justified and useful? A new data standard is not the contribution. |

These sources provide different kinds of evidence. Some are conceptual frameworks, others evaluate benchmarks or clinical processes. HALO explicitly lacks empirical clinical validation [30]. Memory Echo tests persistence after correction and the effects of missing or noisy corrections, while supplying provenance and validity annotations to its mitigation procedure [35]. Dependency-guided repair takes diagnosed faults as input [36]. The proposed clinical workflow would also have to assess whether a disputed amendment is supported. Wrong corrections and downstream repair are therefore already research topics; their clinical evaluation here is a proposed extension. The supplement records source-access limits, including abstract-only access for Dialectic-Med [33].

Existing report quality assurance and AI monitoring also provide important comparisons [12,13]. Better documentation needs to be distinguished from better patient outcomes. A Kenyan cluster-randomized trial improved documentation measures without demonstrating a statistically significant difference in its primary 14-day treatment-failure outcome [15]. That absence of a significant difference does not establish equivalence.

The proposed contribution is the specified comparison: can an added correction workflow improve justified uptake across later handovers while preserving resistance to mistaken revisions, appropriate care status and acceptable burden? The reviewed work does not establish its result. Neither correction management, adversarial reasoning, source tracking nor clinical accountability is introduced here. The research could help a hospital decide to add the procedure, target it to a particular failure, simplify it or stop.

### 3.1 Questions for a learning hospital

This is a medical application of Agentic Institutions within the author's Agentic Epistemics programme. Three related strands ask where information came from, whether a report answers its clinical question adequately, and whether justified corrections change later work [1,2,3,4]. They connect the proposal to established concerns in provenance, evidence-based medicine and learning health systems [10,16,5]. The programme names are not established clinical categories.

| Practical question | Programme | What this evaluation would examine |
|---|---|---|
| Do agreeing accounts rely on the same source or error? | Agentic Stemmatics [4] | Sources, shared evidence and additional correct findings |
| Does the whole report answer its clinical question fairly? | Agentic Representation [3] | Material omissions, justified selection and reader understanding |
| Does a justified correction change the next decision? | Agentic Institutions [2] | Use across handovers, responsibility and consequences |

A traceable source can be wrong, a faithful report can describe a flawed study, and a correct amendment can fail to change care. The evaluation asks where the correction was lost: recording, selection, delivery, interpretation or subsequent use [2, §5]. It also asks whether the report as a whole answers its question [3, §§1.3,3].

Correcting one patient's handover is not evidence that the hospital has learned. Institutional learning would require improvement on later eligible cases, retention across changes of staff or models, and continuing scope for challenge [2, §§3.2,5.4]. Study A tests correction within successive case handovers; the broader effect requires a subsequent study.

The companion Agentic Settlements proposal similarly distinguishes recorded intentions, authority and actual events [41]. That is a design analogy: clinical use needs its own standards for evidence, consent, urgency and responsibility. The supplement retains the fuller programme connections.

### 3.2 How concerns would be examined and carried forward

In this paper, **adversarial inquiry** means deliberately looking for evidence or reasoning that could show a clinical statement, proposed action or accepted procedure to be mistaken, and arranging for it to receive appropriate review [1, §4]. It can be cooperative clinical work. Reviewers do not need to invent an opposing position or meet an objection quota. Their challenges can also be wrong, misleading or harmful, so their accuracy and cost must be assessed.

The proposed workflow has five steps:

1. **State the question and the evidence.** Identify the report's purpose and reader, the action being considered, the available evidence and uncertainty, and what would change the assessment. The reporting question itself can also be challenged.
2. **Look for reasons to reconsider.** Check contradictory evidence, alternative explanations, important omissions and apparent confirmations that share a source. When measuring what reviewers find separately, capture their initial assessments before they see one another's conclusions. Give them adequate common clinical context and record later discussion.
3. **Decide what the evidence supports.** Examine both the original statement and the challenge using relevant observations, standards and clinical expertise. Retain important minority objections. A concern is not disproved because most reviewers missed it, or proved because one reviewer raised it. Human decisions need recorded reasons and can remain uncertain.
4. **Carry the decision forward.** Record whether the correction was accepted, rejected or left open, why, which service is responsible and what follow-up is needed. If the evidence supports a recurring process failure, its accountable owner should revise the process and check whether the correction survives the next handover.
5. **Check whether the procedure is worth using.** Compare corrected errors, false objections, harmful reversals, delays, workload and outcomes that matter to patients against competent simpler practice. Remove additional review if it does not justify its cost.

The depth and timing of review should reflect clinical stakes, urgency and available evidence. A prolonged panel cannot become a prerequisite for timely care. Patients need accessible ways to correct records and question decisions without having to win an argument against an institution. Their accounts of symptoms are evidence to address; they do not automatically prove a proposed explanation.

This is a proposed way to organise review. It does not establish that more reviewers improve medicine. A single well-designed check may be sufficient. The useful result is a justified response to a material problem, rather than disagreement for its own sake.

### 3.3 Historical lessons as questions for evaluation

Medical history motivates these obligations without establishing that this procedure works. Semmelweis's account of hand disinfection and Holmes's earlier warnings about puerperal fever raise questions about how evidence reaches clinical action [38,39]. Snow's investigation checked actual water exposure and acknowledged that Broad Street attacks had declined before the pump handle was removed [40]. Our interpretation is that an institution should inspect evidence, implementation and outcomes separately, including when a reformer's explanation is appealing.

These lessons also apply to this proposal. A challenge needs grounds; opposition to an established view is not evidence of correctness. Working assumptions may remain settled for a task, with reasons for reopening them. The wider aspiration is a sound agentic health institution: a clinical organisation whose use of AI remains accountable to evidence, patient concerns and measured consequences, including evidence against its own procedures.

## 4. Keep clinical statements, care actions and patient outcomes separate

The workflow would maintain three linked records, each with its history of changes. Existing health-record and workflow systems may provide these capabilities; no new mandatory software platform is proposed.

| Record | Question | Examples of status and supporting evidence |
|---|---|---|
| Clinical statements and corrections: the **claim history** | What does the evidence currently support, and in which circumstances? | Proposed, supported within a stated scope, disputed, superseded or indeterminate; source passages, amendments and reasons for the judgement |
| Decisions and events: the **care-action history** | What was decided, who is responsible, and what care actually occurred? | Proposed, accepted, declined, deferred, scheduled, completed, cancelled, no longer indicated or unknown; orders, communications and records of care |
| Health and experience: the **patient-outcome history** | What subsequently happened to the patient's health or experience? | An observed change, adverse event, an outcome not yet measured or uncertainty about its cause; relevant clinical and patient-reported measurements |

These are illustrative categories to adapt to local terminology, not a complete classification system. A scheduled, cancelled or performed scan belongs in the care-action history. Changes in symptoms, function or other relevant health measures belong in the patient-outcome history. Performing the scan does not by itself establish benefit, and an improvement afterward does not alone show that the intervention caused it. A supported recommendation also does not, by itself, authorise action.

Each proposed correction should identify the patient or subject and time, affected statements and documents, source evidence, reasons, contrary evidence, circumstances in which it applies and responsible reviewer. The record should explain why it was accepted, rejected or left unresolved. A later reader needs access to the current interpretation and the reasons earlier wording became unreliable.

The care record should identify the responsible service, next required step, evidence of progress and legitimate exceptions. A person who could not obtain transport has not necessarily refused care. An informed refusal must be retained as a patient decision. If care may have occurred elsewhere but documentation is missing, its status is uncertain; missing documentation does not establish that nothing happened.

Patient outcomes require separate observation, often at a different time. A patient may report worsening function despite an improved biomarker; both observations matter. Even a diagnostic reference standard used for comparison can be imperfect. Agreement among reviewers must not erase that uncertainty.

The procedure would have six requirements:

1. **Keep the reasons and scope.** Preserve why a correction applies, not just the newest wording. Retention and access remain subject to lawful clinical-record policies.
2. **Keep authority explicit.** A suggested finding does not itself authorise a consequential clinical action.
3. **Require evidence when status changes.** Scheduling cannot silently become completion.
4. **Keep unresolved questions visible.** Record missing, contradictory or inaccessible evidence.
5. **Allow corrections to be corrected.** A later objection can be wrong, and a closed task may need to be reopened.
6. **Provide a response to patient concerns.** Accounts of symptoms, preferences or mistaken records need a route to review and response.

These are requirements to evaluate, rather than evidence of an effective system. The histories also need protection against retrospective substitution, loss and misleading presentation of superseded records; signed versions and independently retained copies are the ordinary means, and whether any external anchoring adds to them is an engineering question outside this paper.

## 5. A worked example: an amended report and a cancelled appointment

Suppose an imaging report contains an explicit, non-emergency follow-up recommendation. A specialist later amends a descriptor that matters to its interpretation. The discharge summary retains the earlier wording. An appointment is booked, then cancelled. A later AI assistant reads the summary and booking and proposes marking follow-up complete. This is an invented example.

The proposed workflow would flag two questions. Does the summary still reflect the applicable report? What evidence establishes whether follow-up occurred? The responsible team would check the amendment, determine the current clinical plan and resolve the missing evidence of care.

Several summaries might agree because they copied the same earlier report. A new summary might accurately mention the booking but omit the cancellation, giving a misleading impression. Another team might receive the corrected account yet still act on the old task status. The same case therefore lets us examine where information came from, how it was represented and whether it changed subsequent work.

The appropriate response depends on later evidence. An unsupported message might falsely claim that the amendment was withdrawn. Verified records might establish completion at another provider. The clinical team might document that follow-up is no longer indicated. The patient might make an informed choice to decline. An unanswered attempt to contact the patient might leave the matter unresolved. The workflow should preserve these distinctions rather than maximise interventions or closed tasks.

The example specifies no imaging threshold, treatment or individual decision. Those would belong to the selected clinical pathway and the clinical partner's protocol. It illustrates information and responsibility failures that could be tested before patient-facing use.

## 6. A proposed evaluation programme

### 6.1 Study A: correction across simulated handovers

#### Cases and reference assessments

The first experiment would begin with hypothetical episodes designed by clinicians. These should include justified corrections, false objections (challenges the workflow itself raises against statements the reference supports), partial revisions, changes in what evidence applies, changes in patient choice and evidence that remains uncertain. Uncertainty must not be removed simply to create an answer key. Appropriately authorised real cases could be converted for use under an approved process, but only after the scoring rules developed with hypothetical cases are stable.

Before scored testing, the protocol would preregister how the reference assessments are made, how model outputs are scored, how ambiguity and disagreements are handled, and which subsets receive additional validation. The reference assessment is the judgement against which an output will be scored; it is not presumed to be unquestionable clinical truth.

Two qualified assessors would first assess each case separately, using the released source evidence and clinical context without seeing the tested outputs. For held-out confirmation cases, neither initial assessor would be an author of that case; author participation as assessor would be confined to development cases, and any unavoidable exception would be recorded and those cases reported separately. A subset chosen in advance would receive additional assessment by qualified people who did not write the cases. Report the initial agreement and material disagreements before adjudication, the process of examining and resolving those disagreements. Freeze the reference labels, permitted ambiguity and scoring rules before generating the tested outputs.

For every confirmation case, case authors would record their intended interpretations in advance and withhold them until the assessors' initial judgements were recorded. Report how often and why those judgements differ from the intended interpretations, and any resulting revisions to cases or reference labels, separately from agreement between assessors. Author intention would not settle a disagreement. Final cases and scoring references would remain frozen before tested outputs were generated.

#### Scoring the outputs

Both workflows would answer the same clinical handover questions, which would ask for the current finding, its applicable scope, the grounds for any change from the previous handover, and a position on each candidate revision in the released history. The frozen scoring rules would assess the current statements, relevant care-action status and supporting grounds conveyed by each output, using identical content criteria for prose and structured records. Equivalent content would receive equivalent credit; no particular field, link or disposition object would be required. An ordinary summary identifying the relevant source and accurately explaining the correction could meet the same criteria as a structured record. A citation, disposition label or stored amendment alone would not earn credit. An amendment retained in the history but contradicted by the current handover would not count as uptake.

Output scorers would apply these frozen reference and scoring rules. They would not write the cases or construct the references for cases they score. Remove allocation labels and conceal which workflow produced an output where feasible. Because structured and prose outputs differ in form, masking by concealment alone would often fail. Where feasible, a person who does not score would first transcribe each output into a common content template (current statement, scope, grounds, action status and the position on each candidate revision), and scorers would score the template. Scorers would also record their guess of the arm for each item; guess accuracy would be reported as the masking check, with residual clues and failures recorded.

The scoring rules would be pre-tested on paired prose and structured outputs from development cases, and inter-rater agreement reported by output form. A subset chosen in advance would receive independent duplicate scoring. Report agreement measures and disagreements before adjudication. The adjudicator of a disputed item would not be one of its initial scorers and would record the reasons for the decision. Disputes must not be resolved simply to favour the proposed workflow.

These are separate functions, not a requirement for three complete teams. A person may hold different roles on different cases, with assignments and overlaps disclosed. Role separation and agreement do not establish independent errors or clinical truth. Cases that remain indeterminate would retain that status. Planned sensitivity analyses would examine whether the findings change under defensible alternative labels for material disputes and the specified treatment of indeterminate cases.

#### What the two workflows would do

Both workflows would use the same strong model and a frozen common source history. They would receive identical clinical facts, source versions and release times.

The comparison workflow, or baseline, would use a competent chronological record, summaries updated from the evidence available at each handover, retrieval and general source checking. It would also be instructed to examine counterevidence and material omissions, and to state the grounds for any correction it applies or declines, so that both workflows are asked for the same content.

The proposed workflow would add explicit links between claims, corrections and care states, together with a record of each challenge and its disposition: whether it was accepted, rejected or left unresolved, and why. These links help interpret the evidence; they are more than a change in formatting. The model would construct them from the same released sources and record the model used, source passages and revision history.

In the tested workflow, the model would write the links and the proposed disposition of each challenge. Bounded human review could reject or return an output, but would not write the links or dispositions. A recorded disposition would remain a candidate for the responsible clinician, not an authorised clinical decision.

Reference labels and case-author answer annotations would be withheld from both workflows. Reference constructors would not supply correct links exclusively to the proposed workflow.

#### Human review, resources and what the comparison can show

Human review and computational allowances would be bounded in advance. Measure preparation, summary updating, link construction and checking as part of the relevant workflow's effort and resource use. Report shared preparation separately. Setting allowances does not mean that both workflows will use exactly the same resources. The same team would implement both workflows and develop each on the same development cases with recorded effort; the baseline would be tuned to its best achievable performance, then frozen and reviewed for competence by a person who did not design the proposed workflow, before confirmation testing. If the author's practice implemented either workflow, the protocol would disclose this and the scoring team would be independent of it. The main cost of Study A is clinician time: writing each multi-handover case, two reference assessments per case, duplicate scoring of a prespecified subset and adjudication; model inference is small by comparison. The protocol would state the expected hours per case for each role before funding is sought.

Before execution, define what feedback a human reviewer may return, how many return cycles are allowed, and whether scoring occurs before or after that review. Specify equivalent review opportunities for the baseline. Retain the feedback, resulting revisions and their costs. Feedback would not contain held-out reference labels or case-author answer annotations.

The comparison would estimate the effect of the whole structured correction workflow under these conditions. Both workflows look for counterevidence. The comparison therefore cannot isolate the effect of adversarial inquiry itself or show that a gain came from structure rather than additional attention.

#### Common evidence across handovers

The following paired example shows when evidence would become available. It is an input specification, not an answer key or an executed case.

| Release point and common source material | Baseline procedure | Structured procedure |
|---|---|---|
| Initial handover: report R1 and the same initial summary; no amendment yet exists in the released history | Read the chronological record and check sources | Read the same record and construct any claim/action links from it |
| Next handover: amendment R2 identifies and changes a descriptor in R1; both receive the same passage, identifier and release time | Update the summary after R2 arrives and retain access to R1 and R2 | Update the summary and derive a source-linked correction/disposition from R1 and R2 |
| Later handover: the same booking and cancellation records arrive; no performed-care evidence is supplied | Check current action status against the released records | Check status using links derived from those same records |

At later handovers, both workflows would retain retrieval access to the common released source history and their own previous outputs within the specified budget. Neither would receive future amendments early. The baseline would be allowed to update its summary competently when a valid amendment arrives. If that already resolves the problem and preserves the correction, no added benefit from the proposed procedure is a legitimate result.

Freeze the paired inputs and release times before confirmation testing. This separates differences in the available source evidence from differences created by the workflows themselves.

Assess episodes across successive handovers without treating observations from the same case as independent cases. Before confirmation, the executable protocol would specify case counts, handover schedules and order, model version, the instructions, tools and retrieval settings given to each workflow, repeated runs, and precision or power planning: how precise the estimates need to be, or the ability to detect a specified effect. Development cases would be separate from held-out confirmation cases. This conceptual outline supplies no case count or power claim.

The following case families illustrate the design; they are not an executed dataset.

| Family | Challenge | Required distinction |
|---|---|---|
| Valid amendment | Old summary conflicts with a supported amendment | Use the applicable correction and preserve its grounds |
| Invalid revision | A newer message proposes an unsupported change | Resist recency as a substitute for evidence |
| Cancelled booking | Scheduling exists but performance does not | Keep the action unresolved |
| External completion | Verified outside record establishes the relevant event | Avoid unnecessary duplication |
| Changed plan | An appropriate reassessment alters the next action | Apply the revised scope without erasing history |
| Patient decision | Informed refusal differs from failed contact or access barriers | Preserve agency and the correct follow-up responsibility |
| Wrong subject | Correct information is attached to another person's episode | Prevent propagation and route the identity discrepancy |
| Proxy/outcome conflict | A favourable proxy coexists with worsening reported function | Preserve both observations and avoid declaring benefit |
| Shared-source agreement | Several favourable accounts repeat one underlying report | Identify the observed dependence without counting repetition as additional evidence |
| Misleading account | Individually supported statements omit a material correction or limitation | Distinguish faithful quotation from an adequate answer to the clinical question |

#### Measuring correction and error

The primary comparison asks whether a workflow uses supported corrections and resists unsupported revisions. Before generating tested outputs, the reference assessment would classify correction opportunities as supported, unsupported or indeterminate. Fix the mixture of these opportunities, the way results are combined across handovers, and the treatment of repeated opportunities before confirmation testing.

Reference-indeterminate opportunities would be reported separately. They would not be forced into the supported or unsupported denominators, meaning the sets of opportunities used to calculate the rates below.

| Measure | Definition for the prespecified observation window |
|---|---|
| Justified-correction uptake | Reference-supported correction opportunities in which the current handover applies the correction within its scope and conveys adequate supporting grounds, divided by all reference-supported correction opportunities. An omitted answer, an abstention or an indeterminate model output counts as missed uptake on these supported opportunities. |
| Invalid-revision rejection | Reference-unsupported revision opportunities in which the answer appropriately rejects adoption of the proposed change on the released evidence, divided by all reference-unsupported revision opportunities. Mere omission or an unresolved adoption decision does not count as correct rejection. |
| Conditional unsupported acceptance | Unsupported revisions accepted, divided by all reference-unsupported revision opportunities. This measures the error risk within the unsupported class. |
| Pooled false-correction incidence | Unsupported revisions accepted, divided by all reference-supported and reference-unsupported revision opportunities. This rate depends on the prespecified mixture of opportunities. |

On reference-unsupported opportunities, the rubric would assign each response to adoption of an unsupported change, correct rejection, or neither. An omitted answer or unresolved adoption decision would receive no rejection credit and remain in the unsupported denominator. A reasoned refusal to adopt a change because its evidence is insufficient can be correct rejection; indecision about whether to adopt it is different. Report the neither category and its reasons separately. Prespecify how partial and mixed responses are classified, including when adopting an unsupported element counts as acceptance. Rejection and acceptance rates sum to one only when no response falls into the third category.

Report conditional unsupported acceptance alongside pooled incidence: a low proportion of invalid opportunities must not conceal poor resistance to them. Because pooled incidence equals the unsupported share of opportunities multiplied by conditional acceptance under the fixed mixture, the two absolute limits would be set jointly and checked for consistency. Neither rate from a hypothetical set deliberately containing selected case types is an estimate of clinical population safety. Planned analyses using alternative reference labels and case mixtures would examine the effects of reference uncertainty and case selection.

A harmful reversal is a tested output that withdraws or contradicts a correction the reference marks as supported and material to a consequential care action. It would be counted once per case and separately from unsupported acceptance, and its scoring rule would be frozen with the other measures before tested outputs are generated.

#### What would count as success?

Success would require all of the following:

- A worthwhile improvement in justified-correction uptake over the baseline.
- Invalid-revision rejection that is no worse than the baseline by more than a margin set in advance. This is the non-inferiority margin: the maximum allowable worsening for that measure in the comparison.
- Conditional unsupported acceptance and pooled false-correction incidence within their prespecified absolute limits.
- Scored harmful reversals, as defined above, and human and computational burden within their prespecified limits.

The added procedure must earn its extra work through improved correction uptake. Rejection is a safeguard to preserve within an agreed margin. The relative rejection requirement limits deterioration from the baseline, while the absolute acceptance limit can reject an unsafe result even if the baseline is equally unsafe.

Clinical and statistical partners would justify the worthwhile-benefit threshold, non-inferiority margin and absolute error limits. They would also specify the uncertainty levels and how the analysis handles several measures and comparisons, often called multiplicity. The uncertainty bounds must support every required condition; favourable estimates alone are insufficient. Benefit on one measure cannot compensate for a failed safety constraint.

This is a proposed decision rule, not an already registered or powered protocol. Clear lack of worthwhile benefit or unacceptable harm would count against the procedure. Results that establish neither joint success nor clear failure would remain inconclusive.

#### Other measures and the limits of attribution

The study would also report downstream action classification, uncertainty handling, false objections, misleading omissions, recognition of shared-source agreement and reviewer effort. Break down missed uptake into incorrect or missing current content and inadequate supporting grounds, rather than treating these as the same failure. Each needs frozen scoring rules; the illustrative case-family table does not fully define these outcomes. Report paired effects (differences between workflows on the matched cases) and their uncertainty, accounting for repeated observations within cases. Do not choose severity weights or composite scores after seeing which workflow benefits.

The programme's three strands also need separate assessments. Known copying in a constructed dossier can test whether a workflow recognises source dependence. It cannot validate an inference about a model's training ancestry. Assessments using the complete dossier should distinguish consequential omissions from reasonable compression. A separate blinded reader component would measure what intended readers take a report to establish; an evaluator's judgement that wording is misleading is not a measured effect on readers.

Successive handovers would test whether accepted corrections change later work. If these components are combined in one intervention, the total effect cannot identify which component produced a gain. Later component comparisons, including tests that remove individual components, would be needed. This contrast alone cannot attribute a benefit specifically to adversarial reasoning, record structure or increased attention.

### 6.2 Study B: does an additional read add value?

Only after Study A is interpretable should the programme test additional, separately obtained reads. Start with the strongest single-model workflow and compare it with the same workflow plus a read from the same model, a different model family or a qualified clinician. Identify any clinician-only comparator separately.

Specify common evidence availability and downstream decision rules within a declared resource allowance. Measure the costs of review, filtering and adjudication. Record the packet each reader receives before shared discussion. Agreement after receiving a common packet does not by itself establish independent confirmation. Relevant guidance and source access must remain adequate in every group.

Initial findings can be captured before readers see one another's findings. Record later discussion as shared information, not independent discovery. Measure additional correct findings, shared misses, false objections, findings lost during filtering and adjudication effort. More disagreement or more citations do not establish independence or usefulness. Remove the multi-model procedure if it makes no worthwhile additional contribution.

Two studies motivate these choices but do not establish their effectiveness for handover records. Zöller and colleagues combined physician and model diagnoses on selected clinical vignettes. Physician–model pairs were reported to have greater error diversity than model–model pairs in that setting [17]. This supports testing a hybrid comparison; it does not establish that clinician and model errors are independent.

Farajiamiri and colleagues compared baseline (zero-shot) answers with answers produced using shared evidence across 34 models and 169 radiology questions. They reported lower disagreement among models and a rise in the share of models answering correctly, alongside questions on which most models converged on the same wrong answer [18]. The supplement records apparent inconsistencies in Table 2 that could not be reconciled from its stated per-question definitions. No effect-size estimate from that study is used to plan the proposed experiment.

The limited implication is to record shared evidence and assess correctness, complementary catches and shared misses separately from agreement.

### 6.3 Study C: a bounded hospital pathway

#### Stages, triggers and clinical responsibility

A clinical study would begin with retrospective feasibility, then prospective silent evaluation, in which the workflow's outputs are recorded but not used to change care except through the findings policy below, followed by an assisted-workflow comparison if justified. It should cover one service and one class of explicit, non-emergency follow-up recommendation, with a named owner and the capacity to act.

Retrospective feasibility should first characterise the local problem and existing response. Define an eligible case denominator and sampling rule in advance, rather than selecting only known failures. Trace material amendments, their availability to later teams, subsequent decisions, ownership and evidence of appropriate follow-up. Distinguish stale or inaccessible information from interpretation errors, unclear responsibility, missing outside records, patient choices and practical access barriers. Record current safeguards and staff effort. This assessment requires the applicable governance and has not been performed. If no worthwhile remediable gap is found, or the main barrier requires a different intervention, the added correction procedure may not be justified.

Before evaluation, the protocol would distinguish routine record updating from additional review. Additional review would be triggered by changes, conflicts or unresolved responsibilities that could materially affect a consequential claim or care action.

Before silent evaluation, clinical and research governance would approve a policy for all potentially material findings, including non-urgent unresolved follow-up. Each finding would reach a named qualified clinical owner for verification and triage. Define urgency categories, response windows and escalation when the owner is unavailable or does not respond in time. Existing urgent-care routes must continue.

An unverified model alert would not authorise patient contact or care. The clinical owner would determine the response. Record findings, verification, decisions and actions, including any departure from purely silent observation and what that means for interpreting the study.

#### A functioning comparator

The comparator should have functioning tracking and human coordination, including applicable amended-result and responsibility safeguards such as those described by SAFER [44]. Its functioning would be audited against those safeguards before the comparison, with the audit reported. Comparing a staffed intervention with an abandoned inbox would confuse the effect of the proposed method with the effect of providing a basic service. Randomisation and any grouping of cases for assignment or analysis must reflect how staff share cases and information. A statistician and the clinical partner should set the sample size, minimum worthwhile benefit and acceptable harm and burden boundaries before confirmation.

#### Establishing what care actually occurred

One prospective outcome measure could be appropriate resolution within the clinically specified window among eligible cases. Report completed appropriate care, appropriate non-intervention and informed refusal separately.

Both groups would use the same prespecified evidence hierarchy, observation windows and retrieval/contact effort to establish the outcome. This assessment would be separate from the workflow's own status labels. For completion, verified records of performed care would be sought; an order or booking alone would not suffice. Non-intervention would require the relevant clinical decision and its grounds, and informed refusal its documented decision process, each established in both groups by the same scheduled clinician attestation or patient-contact procedure after the observation window closes, not from records generated by either workflow. Records written by the tested workflow, including disposition reasons, would be removed from assessor dossiers, and the treatment of unknown outcomes in the analysis would be fixed in advance, with sensitivity analyses treating them as unresolved and as appropriately resolved.

Prespecified corroboration would address missing or conflicting evidence, including records from outside the system. If the evidence remains insufficient, the relevant care assessment would remain unknown.

Qualified outcome assessors who are independent of intervention delivery would apply these rules to evidence dossiers in a common format. They would not receive group labels or workflow closure labels. They would assess both whether an event occurred and whether it was appropriate. Masking would be attempted, with residual clues and failures reported rather than assumed absent.

Patient contact or additional retrieval can itself change care. The same scheduled attempts and escalation rules would apply to both groups, with resulting contacts, decisions and care changes recorded. Measurement would therefore be comparable while allowing the study to detect actual differences in care.

#### Burden, patient outcomes and research governance

Additional outcomes would include serious misses, erroneous changes, unnecessary investigations, patient burden, clinically relevant quality of life, staff time, access and differences across groups. Report alerts and repeat alerts per eligible case, triage and review time, workflow delays, and material findings missed by the trigger rules. Limits for unacceptable differences across groups and for departures from documented patient decisions would be set with the clinical partner before confirmation, so that the stopping rules in Section 6.4 can apply them. Better documentation or more appropriate resolution of follow-up tasks would not by itself establish net benefit to patients.

DECIDE-AI addresses early live clinical evaluation [19]. A later randomised trial should use current protocol and reporting guidance, including SPIRIT 2025 and CONSORT 2025, alongside applicable AI extensions such as CONSORT-AI [20,21,22]. These references guide development of a future protocol.

Registration and ethical review must follow the applicable research requirements. The Declaration of Helsinki calls for participant engagement and public reporting of negative and inconclusive findings [23]. This outline is not a completed trial protocol or ethics approval.

### 6.4 What would count against the proposal?

The programme should continue only if the prespecified added benefit is established within acceptable burden and harm, including every part of Study A's joint success rule.

It should simplify or stop if competent record repair, a tracker or one model performs as well within the prespecified decision bounds. A breached safety constraint, unacceptable review burden or unacceptable inequitable effects would also warrant simplification or stopping. A favourable result on one outcome cannot override a failed safety constraint.

The result should remain inconclusive when the sample, reference standard or uncertainty intervals cannot distinguish these possibilities. Failure to show a difference is not automatically evidence that two procedures perform equally well.

Faster correction is valuable only if correctness, patient agency and care quality are preserved. These initial studies cannot establish effects beyond their prespecified endpoints. Clinical benefit, economic value and performance in other settings require their own evidence.

## 7. Responsibility, evidence and incentives

The work needs a clinical home, professional competence, patient participation and resources for follow-up. A first implementation could be a programme within an existing hospital, involving clinical, quality, informatics and research staff. This paper does not prescribe founding a new hospital or legal entity.

Responsibility for developing the tool, using it, assessing incidents and publishing evaluations should be explicit. Common control and financial interests must be disclosed. Separate boxes on an organisation chart do not establish independent judgement. Evaluation teams need the right to publish unfavourable findings. Patients need accessible ways to challenge errors and understand decisions; staff who raise concerns need protection from retaliation and a named route to resolution.

Incentives should support justified correction and appropriate care, rather than reward the number of objections or automatic defence of the model. Clinical teams need authority to suspend the procedure when its burden or consequences become unacceptable. Authority to admit evidence, change records, set AI permissions and revise institutional rules also needs stated limits, recorded reasons and effective appeal, including challenges to the decisions of those who lead or fund the programme. Keeping a record is insufficient if people cannot contest the authority acting on it.

Financing requires the same scrutiny. Appropriate review, coordination and access need resources even when an algorithm is inexpensive. Rewards tied only to testing volume, referrals, closed tasks or a proprietary score may conflict with appropriate care. Payment based on outcomes can also encourage selection of easier cases or distort measurement. The organisation operating the programme must keep its methods and financing open to examination.

A Cochrane review found an association between manufacturer sponsorship and more favourable efficacy results and conclusions in drug and device studies [24]. This does not make every sponsored result false or establish a single motive for an industry. Scrutiny should apply to pharmaceuticals, testing businesses and AI services, including the institution's own commercial interests.

Health records require appropriate confidentiality, access, correction and retention arrangements. Study A begins with no patient data. Before any real record is used, the protocol would state where the model runs, whether any patient text leaves the hospital's controlled environment, and under which agreement; this is a condition of partnership. Retaining a record for its lawful period does not mean making it public or permanently immutable. In the European Union, software qualification and classification depend on intended purpose and function [25]. The EU Artificial Intelligence Act may also apply; its obligations depend on classification and on application dates still phasing in, and the protocol would record the applicable assessment. Dataset-specific access conditions also matter: PhysioNet restricts use of credentialed data with external AI services unless the applicable requirements are met [26]. Neither source substitutes for assessment of the actual study, jurisdiction and data arrangements. No system or study described here has been legally classified or authorised.

## 8. What remains unknown

The main limitation is that the proposed workflow has not been evaluated. No reduction in clinical error, improvement in quality of life, cost saving or ability to operate at scale has been demonstrated. Existing systems may meet these requirements with little additional machinery. The proposed comparisons test that possibility.

The literature search is selective. Relevant work may have been missed, and the proposal's distinctness remains open to examination. A consistent record is also different from a correct diagnosis, and neither establishes patient benefit.

Reviewers can make mistakes, especially when the same people design, label and assess cases. Several models may repeat a shared error. Under workload, human approval may become a routine click. Performance in a simulated handover sequence or one hospital cannot establish general effectiveness.

AI could shorten the time between new evidence, correction and appropriate care. It could also spread an error faster, make unsupported claims more persuasive or add work to already stretched services. Learning requires adverse findings to change the procedure, rather than merely enter its archive.

The purpose of this research is to improve patient care and the institution's ability to learn. This paper offers a way to test one part of that ambition: whether a specified correction workflow adds enough to competent simpler practice to justify using it. The result may support the proposal, call for a simpler approach or remain inconclusive.

A hospital that wants to test this would need to do three things: choose one service and one class of non-emergency follow-up recommendation with a named owner; let its clinicians construct and assess the simulated cases for Study A; and, if Study A justifies it, run the retrospective feasibility stage of Study C on its own records under its own governance. In return it would get a measured account of where corrections are lost in that pathway and whether an added procedure earns its cost, including the finding that its existing process already suffices.

## Author declarations

**AI assistance.** AI tools supported literature discovery, bounded source inspection, study-design development, drafting and revision. Codex prepared and revised the manuscript; Claude contributed critique and a scope-cut draft; Kimi, DeepSeek and Grok supplied model critiques at recorded earlier versions. Qwen subsequently supplied a critique of v0.8 and a supplement-informed reassessment. A later user-carried Claude critique assessed v0.11; Codex checked its concerns and revised the specification in v0.12. The v0.13 revision added clinical motivation, AI-documentation evidence and a more explicit account of predecessors and local problem assessment, using internal source-research assistance. The v0.14 revision reworded the practical questions in Section 1.3 at the author's request. The v0.15 revision, following a further round of model critiques recorded in the supplement, clarified the scoring and reference-assessment rules in Section 6.1, the outcome ascertainment in Section 6.3, the governance passage in Section 7, and the release labelling. These critiques were not journal peer review or independent clinical validation. The author supplied the institutional research question and programme context and selected the model's role in the proposed workflow. Responsibility for the argument, contents, licence and decision to publish remains with the human author. The accompanying supplement records the preparation history, source-access limits and the distinction between initial critique and directed confirmation.

**Competing interests.** The author has a commercial interest in CounterProof, a practice of Clavestra Capital Limited, whose review and construction activities could benefit from adoption of the proposed methods. This manuscript provides no independent evaluation of CounterProof's effectiveness.

**Funding.** No external funding supported this work.

**Data and ethics.** No patient records or participant data were processed for this manuscript. No participant study was performed. The proposed study scenarios are hypothetical; cited clinical observations come from published sources. Future clinical work requires its own governance and approvals; none are claimed here.

**Licence.** Creative Commons Attribution 4.0 International (CC BY 4.0).

## References

Numbered in order of addition to the manuscript, not of first citation.

1. Soons LJ. *Agentic Epistemics: An Introduction*. CounterProof Research. 31 October 2026, revised 1 November 2026. [Programme introduction](https://research.counterproof.io/intro/agentic-epistemics/). Accessed 4 November 2026.
2. Soons LJ. *Agentic Institutions: Organisational Memory, Correction and Continuity Across Successive AI Agents*. Preprint v0.3. 28 October 2026. [DOI](https://doi.org/10.5281/zenodo.23727633).
3. Soons LJ. *Agentic Representation: Evidence, Selection and Warrant in AI-Assisted Reports*. Conceptual preprint v0.3. 28 October 2026. [DOI](https://doi.org/10.5281/zenodo.23727617).
4. Soons LJ. *Agentic Stemmatics: Collation, Provenance, and Conservation for an Emerging Machine Textual Tradition*. Preprint v1.41. 18 October 2026. [DOI](https://doi.org/10.5281/zenodo.23458161).
5. AHRQ. About Learning Health Systems. [Official framework](https://www.ahrq.gov/learning-health-systems/about.html). Accessed 4 November 2026.
6. Faden RR et al. An ethics framework for a learning health care system: a departure from traditional research ethics and clinical ethics. *Hastings Center Report*. 2013;43(s1):S16–S27. [DOI](https://doi.org/10.1002/hast.134).
7. Gourishankar A, Shalhout SZ. Learning Health Systems for Chronic Disease: Embedded Artificial Intelligence Tools That Continuously Improve Care. *Preventing Chronic Disease*. 2026;23:260103. [Original essay](https://www.cdc.gov/pcd/issues/2026/26_0103.htm), [DOI](https://doi.org/10.5888/pcd23.260103).
8. Lin G, Jiang R, Elhadad N, Xu X. A framework for longitudinal health AI agents. *Nature Health*. 2026;1:796–805. [Author manuscript](https://pmc.ncbi.nlm.nih.gov/articles/PMC13367785/), [DOI](https://doi.org/10.1038/s44360-026-00141-x).
9. Imani MH, Imani AH. Keeping clinical evidence current as medical artificial intelligence systems change. *Discover Artificial Intelligence*. 2026;6:1654. [DOI](https://doi.org/10.1007/s44163-026-02636-8).
10. HL7. FHIR: Task, CarePlan and Provenance. R5, v5.0.0. [Task](https://hl7.org/fhir/R5/task.html), [CarePlan](https://hl7.org/fhir/R5/careplan.html), [Provenance](https://hl7.org/fhir/R5/provenance.html). Accessed 4 November 2026.
11. NIST. HL7 AI Transparency on Fast Healthcare Interoperability Resources. Updated 15 September 2026. [Project description](https://www.nist.gov/programs-projects/hl7-ai-transparency-fast-healthcare-interoperability-resources-fhir). Accessed 4 November 2026.
12. Wang L et al. Large Language Model–Assisted Radiology Report Quality Assurance: Real-World Multicenter Evaluation. *Journal of Medical Internet Research*. 2026;28:e106250. [Original article](https://www.jmir.org/2026/1/e106250), [DOI](https://doi.org/10.2196/106250).
13. Flanders AE et al. A multi-agent large language model framework to automatically assess performance of a clinical AI Triage tool. *npj Health Systems*. 2026;3:35. [Original article](https://www.nature.com/articles/s44401-026-00100-4), [DOI](https://doi.org/10.1038/s44401-026-00100-4).
14. Zaki-Metias KM et al. The FIND Program: Improving Follow-up of Incidental Imaging Findings. *Journal of Digital Imaging*. 2023;36:804–811. [Original article](https://link.springer.com/article/10.1007/s10278-023-00780-6), [DOI](https://doi.org/10.1007/s10278-023-00780-6). Read with the [published correction](https://doi.org/10.1007/s10278-023-00898-7), 15 September 2023;36:2662–2663, concerning study counts and Table 1.
15. Agweyu A et al. Generative AI-enabled clinical decision support system in primary care: a pragmatic, cluster-randomized trial. *Nature Medicine*. 2026;32:3032–3039. [DOI](https://doi.org/10.1038/s41591-026-04503-6).
16. Sackett DL et al. Evidence based medicine: what it is and what it isn't. *BMJ*. 1996;312:71–72. [DOI](https://doi.org/10.1136/bmj.312.7023.71).
17. Zöller N et al. Human–AI collectives most accurately diagnose clinical vignettes. *Proceedings of the National Academy of Sciences*. 2025;122(24):e2426153122. [DOI](https://doi.org/10.1073/pnas.2426153122).
18. Farajiamiri M et al. Retrieval-augmented reasoning reshapes collective reliability in text-based radiology question answering. *Patterns*. 2026;7(10):101639. [DOI](https://doi.org/10.1016/j.patter.2026.101639). See the supplement for source-version and numerical-consistency qualifications.
19. Vasey B et al. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. *BMJ*. 2022;377:e070904. [DOI](https://doi.org/10.1136/bmj-2022-070904).
20. Chan AW, Boutron I, Hopewell S, et al. SPIRIT 2025 statement: updated guideline for protocols of randomised trials. *BMJ*. 2025;389:e081477. [DOI](https://doi.org/10.1136/bmj-2024-081477).
21. Hopewell S, Chan AW, Collins GS, et al. CONSORT 2025 statement: updated guideline for reporting randomised trials. *BMJ*. 2025;389:e081123. [DOI](https://doi.org/10.1136/bmj-2024-081123).
22. Liu X et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI Extension. *BMJ*. 2020;370:m3164. [DOI](https://doi.org/10.1136/bmj.m3164).
23. World Medical Association. Declaration of Helsinki, 2024 revision. [Current official text](https://www.wma.net/policies-post/wma-declaration-of-helsinki/).
24. Lundh A, Lexchin J, Mintzes B, Schroll JB, Bero L. Industry sponsorship and research outcome. *Cochrane Database of Systematic Reviews*. 2017;2:MR000033. [DOI](https://doi.org/10.1002/14651858.MR000033.pub3).
25. Medical Device Coordination Group. *Guidance on Qualification and Classification of Software in Regulation (EU) 2017/745 – MDR and Regulation (EU) 2017/746 – IVDR*. MDCG 2019-11 Rev.1. June 2025. [Official guidance](https://health.ec.europa.eu/document/download/b45335c5-1679-4c71-a91c-fc7a4d37f12b_en?filename=mdcg_2019_11_en.pdf).
26. PhysioNet. *Use of MIMIC Data with Large Language Models and Online Services*. 24 September 2025. [Dataset-use policy](https://physionet.org/news/post/llm-responsible-use/).
27. Starmer AJ et al. Changes in Medical Errors after Implementation of a Handoff Program. *New England Journal of Medicine*. 2014;371:1803–1812. [Primary article](https://www.nejm.org/doi/full/10.1056/NEJMsa1405556).
28. Bradford A et al. Evaluation of Measure Dx, a Resource to Accelerate Diagnostic Safety Learning and Improvement. *Journal of General Internal Medicine*. 2025;40(4):782–789; first published 22 October 2024. [Primary article](https://pmc.ncbi.nlm.nih.gov/articles/PMC11914432/). DOI: 10.1007/s11606-024-09132-8.
29. Institute for Healthcare Improvement. Plan-Do-Study-Act (PDSA) Worksheet. [Official method guidance](https://www.ihi.org/library/tools/plan-do-study-act-pdsa-worksheet). Accessed 4 November 2026.
30. Zai AH, Adibuzzaman M, McManus DD, Walkey A. The HALO Model: A Learning Health System Framework for Artificial Intelligence. *Learning Health Systems*. 2026;10(3):e70100. [Primary article](https://onlinelibrary.wiley.com/doi/10.1002/lrh2.70100).
31. Pugh SL et al. Detecting Clinical Discrepancies in Health Coaching Agents: A Dual-Stream Memory and Reconciliation Architecture. *Proceedings of Machine Learning for Healthcare*. 2026;PMLR 340:1581–1612. [Proceedings record](https://proceedings.mlr.press/v340/pugh26a.html), [inspected preprint v1](https://arxiv.org/html/2604.27045v1). Final-PDF/preprint equivalence was not checked.
32. Wang H, Zhao Z, Zhao Y. ClinTraceBench: Source-Verifiable Longitudinal Clinical Reasoning over EHR-Derived Dialogues. *Findings of EMNLP*. 2026:31524–31549. [Official publication record](https://aclanthology.org/2026.findings-emnlp.1480/).
33. Lu Z, Su J. Dialectic-Med: Mitigating Diagnostic Hallucinations via Counterfactual Adversarial Multi-Agent Debate. *Findings of ACL*. 2026:36877–36894. [Publisher abstract and record](https://aclanthology.org/2026.findings-acl.1837/). DOI: 10.18653/v1/2026.findings-acl.1837.
34. Liu Z et al. AIGS: Generating Science from AI-Powered Automated Falsification. Preprint, arXiv:2411.11910v2, 24 November 2024. [Inspected version](https://arxiv.org/html/2411.11910v2).
35. Li Y et al. Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation. Preprint, arXiv:2610.17763v1, 13 October 2026. [Inspected version](https://arxiv.org/html/2610.17763v1).
36. Wang Y et al. From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents. Preprint, arXiv:2608.10502v2, 22 October 2026. [Inspected version](https://arxiv.org/html/2608.10502v2).
37. Sebin B, Sebin IK. From model failure to system harm: operationalizing a sociotechnical pathway for healthcare AI safety. *Frontiers in Artificial Intelligence*. 2026;9:1956533. Perspective. [Primary article](https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1956533/full).
38. Semmelweis IP. *Die Aetiologie, der Begriff und die Prophylaxis des Kindbettfiebers*. 1861, pp.55–56. [Digital transcription](https://textgridrep.de/browse/4bn26.0).
39. Holmes OW. The Contagiousness of Puerperal Fever. First published 1843. [Primary essay transcription](https://biotech.law.lsu.edu/cphl/history/articles/pf_holmes.htm).
40. Snow J. *On the Mode of Communication of Cholera*. Second edition, 1855, pp.51–52,74–79. [Scan-derived digital edition](https://www.gutenberg.org/cache/epub/72894/pg72894-images.html).
41. Soons LJ. *Agentic Settlements and Payments: A Federated Settlement Layer Beneath Machine-Payment Protocols*. Author-supplied conceptual preprint, draft v0.15, 4 November 2026. Inspected PDF, especially §§7.2,11.4,13; no separate public deposit or DOI was verified for this draft.
42. Son AY et al. Patient recalls associated with resident-to-attending radiology report discrepancies: predictive factors for risky discrepancies. *Insights into Imaging*. 2022;13:97. [Primary article](https://link.springer.com/article/10.1186/s13244-022-01233-4).
43. Dalal AK et al. The Impact of Automated Notification on Follow-up of Actionable Tests Pending at Discharge: a Cluster-Randomized Controlled Trial. *Journal of General Internal Medicine*. 2018;33:1043–1051. [Primary article](https://pmc.ncbi.nlm.nih.gov/articles/PMC6025668/), [DOI](https://doi.org/10.1007/s11606-018-4393-y).
44. Office of the National Coordinator for Health Information Technology. *SAFER Self-Assessment: Test Results Reporting and Follow-Up*. Recommended practices 2.2–2.4, pp.13–15. [Official guide](https://www.healthit.gov/sites/default/files/playbook/pdf/8-test-results-reporting-final.pdf). Retrieved 4 November 2026; the document refers to the 2025 guides and carries August 2024 footers.
45. Williams CYK et al. Physician- and Large Language Model–Generated Hospital Discharge Summaries. *JAMA Internal Medicine*. 2025;185:818–825. [Primary article](https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2833228), [DOI](https://doi.org/10.1001/jamainternmed.2025.0821).
46. Mohta V. Amended Lab Results: Communication Slip. *PSNet WebM&M*. Agency for Healthcare Research and Quality. 1 February 2012. [Reported case and commentary](https://psnet.ahrq.gov/web-mm/amended-lab-results-communication-slip).