When a doctor writes "rash possibly related to lamotrigine, per patient," the sentence records three things: an event (a rash), a causal hypothesis (the drug), and how much the writer vouches for it (a patient report, hedged). When a language model later compresses years of notes into a running summary, it tends to keep the event and drop the rest. After a few rounds of summarizing, the line may read "lamotrigine-induced rash," and a suspicion has become a fact. Nothing false was added in any single step. The same thing happens to agency, meaning who did what. "Medication was discontinued" is compatible with the patient stopping it, the physician holding it for toxicity, or insurance refusing it. A summarizer that fills in the missing agent is making a causal claim the record never made.
This matters because researchers who want to know how treatments work outside clinical trials increasingly mine exactly these narratives. Many of the variables that explain why real-world results fall short of trial results live only in free text: why a drug was stopped, whether an adverse event was attributed to it, whether the patient took it. We propose to measure and prevent this "stance laundering." Stance here means the evidence status of a claim (suspected, confirmed, or patient-reported) and its grammatical agent (who performed the action). Stance laundering is the silent upgrading of that stance and the silent filling-in of agents as clinical records are recursively summarized. We will show how much it distorts the treatment-discontinuation and adverse-event labels that real-world evidence studies depend on.
Core Claim
Recursive summarization of multi-encounter clinical records systematically upgrades hedged or reported claims to asserted ones and assigns agents to agentless events. The rate grows with the number of summarization steps. A per-event record of agent, reporter, and evidence status, carried alongside the summary and checked before each update, will cut this upgrading substantially without hurting summary quality. It will also change downstream discontinuation-reason labels, and it will increase appropriate abstention on audit tasks. The new object is neither a better summarizer nor a better fact checker. It is a measurement of whether a summary preserves how the source vouched for a claim, plus a mechanism that enforces that preservation across a long horizon.
Where This Sits
Three lines of work motivate the gap. Studies of news framing show that grammar alone relocates responsibility without asserting falsehoods. To Protect and To Serve? found conservative outlets used agentless passives 61% more often than liberal ones when describing police shootings. The FIGNEWS 2024 guideline work showed that passives and nominalizations can be annotated reliably, with a Cohen's kappa of 0.85. Modeling Factuality Judgments in Social Media Text showed that the reporting verb ("admit" versus "think") predicts reader certainty better than the source's identity does. Together these say that agency and evidence status are carried by recoverable surface features. None of this work examines clinical text or long documents.
Work on long-form narrative supplies the method and the failure mode. Re3 keeps an attribute dictionary of characters and issues local edits when a draft contradicts it. Its edits cover surface attributes like age or occupation, not evidence status. SNaC showed that span-level, category-specific annotation raises agreement from Krippendorff's alpha 0.19 to 0.69, and that overlap metrics cannot see discourse errors. We reuse Re3's "ledger plus local edit" design for a different attribute class. We reuse SNaC's annotation workflow for a new error category, the unsupported upgrade.
Clinical work defines the application and the closest competitors. CLIN-SUMM is the direct target. It builds a running summary by feeding the prior summary plus each new note into a "changes over time" update, compresses tokens by 85.3%, and is judged only by Likert ratings (4.84/5 correctness) and downstream dementia AUROC (0.79). Nothing in it checks whether hedges and agents survive. EHR-ReasonCon checks whether notes agree with structured tables and finds the best models reach F1 only in the 50s. It verifies propositions, not evidence status. Cite-While-You-Generate attaches source citations to summary sentences using decoder attention (76.33 Macro-F1). A citation shows where a claim came from, but not whether the summary kept the source's hedge. Cognitive bias in LLM reasoning found anchoring and confirmation errors in 23.1% of GPT-4 oncology-note interpretations, but it treats the notes themselves as neutral. Our hypothesis is that laundered summaries are one supply route for the anchors.
One result constrains the design. Dead or Murdered? found that models predict where a sentence puts its focus well (R² up to 0.66) but predict perceived blame poorly (R² about 0.24). We therefore do not try to model perceived blame or clinical fault. We track only the explicitly encoded agent, the reporter, and epistemic markers. These are closer to the surface-recoverable "focus" end of that finding.
Research Program
The first component is an annotation scheme and dataset. We draw multi-admission patients from MIMIC-IV, the population behind EHRNoteQA and CliniCARE-Bench, and restrict to patients with at least three discharge summaries. For every medication stop, dose change, and adverse event, annotators record four fields. The first is the grammatical agent (named, or absent through a passive, nominalization, or unaccusative verb such as "the rash developed"). The second is the reporter (clinician, patient, family, outside record). The third is the evidence status (confirmed, suspected, patient-reported, ruled out). The fourth is the stated cause, if any. Annotation follows SNaC's step-by-step, span-level procedure, with category definitions borrowed from the FIGNEWS guideline. We target a few hundred patients, a size comparable to the 105-note EHR-ReasonCon, and report agreement per field.
The second component is a measurement protocol. For each summary produced of a patient's record, we align every summary claim about an annotated event to its source records, using CWYG-style attention attribution as a candidate generator followed by human confirmation on a subset. We then compute an upgrade rate. Let $U$ be the number of aligned claims whose evidence status in the summary is stronger than the strongest status in any cited source, or whose agent is named when every source left it absent. Let $N$ be all aligned claims. The upgrade rate is $U/N$. We report it as a function of the number of recursive updates the claim has passed through, and that curve is the central drift measurement. A trained classifier, checked against the human labels, extends the metric to the full cohort.
The third component is the intervention, a stance ledger. For each event, the summarizer keeps a structured entry with the four annotated fields, extracted from each new note. Syntax-aware agent extraction in the style of the police-violence study handles passives and nominalizations, and reporting-verb features follow the factuality work. Before a running-summary update is accepted, a checker compares every summary sentence touching a ledger event against its entry. Any upgrade not licensed by a source triggers a local rewrite, as in Re3's edit step. When sources conflict, for example oncology attributing kidney injury to a drug while nephrology attributes it to sepsis, the ledger keeps both attributions and the summary must say so. It may not resolve the conflict.
Baselines
The naive reference point is the CLIN-SUMM two-prompt recursive pipeline, reimplemented on our MIMIC-IV cohort. We keep its Jaccard redundancy filter and sliding window and use the same underlying model as our method. It is the most directly comparable published longitudinal summarizer. Its reported numbers (4.84/5 correctness, 0.79 dementia AUROC) come from private Mass General Brigham data on a different task. We therefore report them only as context, and all comparisons use our reimplementation. No reported result was found on Lacuna for stance or agency preservation by any summarizer. Our baseline numbers will be the first.
The strongest cheap competitor is the same pipeline with an instruction to "preserve hedging, attribution, and who performed each action." It costs nothing and directly targets the failure. The strongest pipeline competitor is the CLIN-SUMM pipeline augmented with a standard clinical assertion-status tagger (present, negated, possible, hypothetical), whose tags are injected into the prompt. This baseline separates our contribution from what existing clinical NLP already captures. We also run a long-context condition with no summarization: all raw notes are fed to the model and it answers the downstream questions directly.
Every condition uses the same base model, the same patients, and the same number of development-set prompt iterations for tuning. Beating a baseline means a lower upgrade rate with overlapping confidence intervals excluded, at equal or better clinician-rated correctness and completeness.
Two ablations isolate the key ingredient. The first removes the agent and reporter fields, leaving only evidence status. This tests whether agency tracking adds anything beyond assertion status. The second keeps the ledger but drops the pre-update check. This tests whether explicit structure alone helps or whether enforcement is what matters.
Why not just…?
Why not just prompt the summarizer to preserve hedges? This is a real threat, and it may largely work for one or two updates. Our hypothesis is that instruction-following erodes as the running summary, not the original note, becomes the model's only view of the past. Re3's gains over rolling-window baselines (14% absolute on plot coherence) suggest that explicit external state beats implicit context over long horizons. If the prompt-only baseline matches the ledger across the full drift curve, we will report that. The contribution then shrinks to the measurement and the dataset, which would still be new.
Why not just use clinical assertion-status tagging? Assertion taggers mark whether a condition is present, negated, or possible. They do not say who stopped a drug or whether a claim came from the patient or the clinician. Nobody-named discontinuations are exactly where nonadherence and toxicity get confused. The factuality study shows the reporting verb carries certainty information that source identity does not replace. The assertion-injected baseline and the status-only ablation test this directly. If agent and reporter fields add nothing on the downstream discontinuation labels, the claim about agency fails.
Why not skip summarization and use a long-context model on raw notes? If no summary exists, nothing can be laundered in summarization. But EHRNoteQA reports that accuracy falls as more discharge summaries are added, and CliniCARE-Bench's best systems resolve only about 76% of longitudinal cases, dropping 4.8 to 14.8 points under defect-free scoring. Raw-context reading also has its own anchoring errors, per the oncology study. We include this condition on the downstream tasks. If it matches ledger-based summaries at acceptable cost, our intervention is only useful where summaries are required, for example in cohort-scale extraction or at care handoffs.
Why not just add citations with Cite-While-You-Generate? A citation lets a reader check a source, but the summary sentence still asserts the upgraded claim. At 76.33 Macro-F1 on short encounters, the citation itself is imperfect, and it has not been tested on multi-year records. We use CWYG as the alignment tool inside our metric. We also test "citations only" as a condition, to see whether attribution alone lowers the upgrade rate. We expect it not to.
Proposed Experiments
The first decisive check is whether laundering exists and grows. We plot the upgrade rate against the number of recursive updates for the CLIN-SUMM reimplementation. A flat, near-zero curve would falsify the premise, and we would stop.
The second check is whether the ledger lowers the curve relative to the prompt-only and assertion-tagged baselines. Clinician Likert ratings of correctness and completeness, collected on the CLIN-SUMM scale, confirm that the summaries did not become hedged into uselessness.
The third check is downstream consequence. From each condition's summaries we derive the label a real-world evidence study would use: discontinuation reason (toxicity, nonadherence, lack of efficacy, or indeterminate) and adverse-event attribution to a drug. We score these labels against the annotated sources. For a fixed drug class, we then report how the cohort-level proportion of "toxicity-driven discontinuation" shifts between conditions. This is a direct estimate of the bias laundering would inject into an effectiveness analysis.
The fourth check uses CliniCARE-Bench. We feed its longitudinal cases through each summarization condition and measure defect-free accuracy and the rate of correct "Indeterminate" verdicts. The benchmark reports that all systems over-commit rather than abstain. The prediction is that stance-preserving summaries raise appropriate abstention, and that laundered summaries lower it relative to raw context.
We also run a cross-check on CareTransition-Audit. There, clinicians marked 1.7% of items "Unclear" where models never did. We test whether ledger summaries move model "Unclear" rates toward the clinician rate without collapsing the κ agreement of about 0.49 reported for the best models. The main limitation carries over from all these benchmarks: every one is drawn from a single center. We will report that as a limit on generality, not claim it away.
References
- To Protect and To Serve? Analyzing Entity-Centric Framing of Police Violence
- The Guidelines Specialists at FIGNEWS 2024 Shared Task: An annotation guideline to Unravel Bias in News Media Narratives Using a Linguistic Approach
- Modeling Factuality Judgments in Social Media Text
- Re3: Generating Longer Stories With Recursive Reprompting and Revision
- SNaC: Coherence Error Detection for Narrative Summarization
- CLIN-SUMM: Temporal Summarization of Longitudinal Clinical Notes
- Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records
- Cite-While-You-Generate: Training-Free Evidence Attribution for Multimodal Clinical Summarization
- Cognitive bias in LLM reasoning compromises interpretation of clinical oncology notes
- Dead or Murdered? Predicting Responsibility Perception in Femicide News Reports
- EHRNoteQA: An LLM Benchmark for Real-World Clinical Practice Using Discharge Summaries
- CliniCARE-Bench: Clinical Calibrated Audit of Medical Reasoning in EHR
- CareTransition-Audit: A Benchmark to Audit Discharge Summaries for Efficient Care Transitions