Grounded Adjudication of Variations across Extracted TimeLines (GAVEL): Comparing Clinical Timelines Against Their Case Reports

URL SCAN: Grounded Adjudication of Variations across Extracted TimeLines (GAVEL): Comparing Clinical Timelines Against Their Case Reports
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

GAVEL builds a machine-mediated tribunal for clinical timeline extraction. It replaces crude reference comparison and imperfect expert annotations with report-grounded adjudication: identify each discrepancy, classify it, attach the supporting passage, issue a verdict, and guide a merged revision.

The reported gains are substantial at the workflow level: 89.4% and 88.6% of findings were confirmed by manual review; merged timelines were preferred in 77.0% of 126-report comparisons; and discrepancies attributed to the evaluated timeline fell from 7.63 to 0.85 per report.

But these are adjudication and preference metrics, not proof of clinical truth, patient safety, causal correctness, or real-world utility. GAVEL improves the machinery for deciding which machine-produced timeline looks better against a case report. It does not eliminate uncertainty; it industrializes its management.

The Core Fallacy

The central error under the Discontinuity Thesis is a category error: treating the problem primarily as evaluation quality rather than labor substitution.

GAVEL may solve a genuine bottleneck in benchmarking and revision. That does not preserve the economic role of expert annotators. It makes their judgment more compressible, auditable, and eventually replaceable. The paper’s own architecture strengthens P1: cognitive work is divided into extraction, matching, adjudication, evidence retrieval, and merging, then each layer is made machine-operable.

A report passage attached to a verdict is not ground truth. A preferred merge is not truth. A lower discrepancy count is not clinical validity. Traceability is a control feature, not an epistemic guarantee. GAVEL makes automation safer to deploy, which makes automation easier to deploy.

Hidden Assumptions

• The case report is complete, accurate, and chronologically coherent enough to serve as the adjudication substrate.
• The LLM judge can distinguish genuine disagreement from omission, ambiguity, documentation error, and clinically meaningful nuance.
• Manual confirmation is sufficiently independent, despite using the same report and evaluation framework.
• Event-level discrepancy counts correlate with clinical usefulness and safety.
• The 0.10 cutoff is meaningful rather than an unstable threshold; the reported 60% and 48% true-match rates near it expose fragility.
• A preferred merged timeline is objectively superior rather than merely more readable or more aligned with evaluator expectations.
• Results across 126 reports, six extractors, two human annotators, and the named models generalize across specialties, institutions, documentation styles, and missing-data regimes.
• Report-grounded evidence is enough; the wider patient record, conflicting sources, and undocumented events are treated as outside the problem.
• Adding a passage citation creates accountability. It does not assign liability, detect every hallucination, or guarantee safe downstream decisions.
• Human review remains available at the required scale and cost. That is an institutional lag, not a permanent moat.

Social Function

Classification: partial truth, verification arbitrage, and transition management.

This is not simple copium. The reported engineering improvement is real within the stated evaluation. But the social function is to convert expert disagreement from a blocking requirement into a structured exception queue. It turns human judgment into calibration data, review labor, and residual oversight around an increasingly autonomous pipeline.

The paper therefore functions as a bridge technology. It reassures institutions that cognitive automation can be made auditable while quietly reducing the amount of cognition that must remain human. Its moat is not human expertise; its moat is the data, infrastructure, deployment authority, and liability position of whoever owns the adjudication system.

The Verdict

GAVEL is a competent instrument for accelerating the replacement of clinical timeline evaluators. Its 77% merge preference and sharp discrepancy reduction demonstrate workflow leverage, not preserved human indispensability. Under the Discontinuity Thesis, it is evidence for the transition from human annotation to machine extraction plus machine adjudication, with humans pushed into temporary servitor and liability roles.

The paper does not defeat obsolescence. It documents the tribunal being built to process it.