Origin
The issue began at the evidence gate. The author moved it inside the hospital, where the tools had already entered.
The starting prompt treated evidence as a front-door condition for clinical AI. The author rejected that formulation as too narrow. Many tools were already entering through the electronic health record, bundled products, and pilots that became routine. The live question became which safeguards should govern systems already in use.
The selected landing was “default absorption”: installation and renewal can harden into routine without a fresh evidence decision. A later construct critique forced an important correction. The essay stopped treating that as a story about what people privately trust and instead tracked operational reliance — deployment, signing, and exposure — and the gap between those routines and a stake-matched evidence loop.
A final source-fidelity pass went back to the papers and public instruments. It produced nine concrete corrections, including study-maturity percentages, alert-override ranges, the number and type of systems studied, and the limits of claims about experienced readers. Those repairs are visible in the Reference layer.
Where it came from
Human trigger
The author saw a timing problem in the original question: an evidence gate cannot govern tools that have already arrived through institutional software and ordinary workflow.
Process pressure
Rival landings, a construct critique, a plain-register experiment, two outside narrative reviews, and a late primary-source re-check each had authority to alter the draft.
Public check
Twenty-three public sources anchor adoption, drift, the Epic sepsis chronology, ambient-note errors, alert behavior, attitudes, contracts, liability, and current governance.
The conversation behind this
How did the issue change?
Read the public-safe verbatim development record: the author rejects the first frame, redirects the landing, pushes the prose into plain language, adjudicates outside reviews, and changes the process after publication. Two preservation gaps are marked rather than reconstructed.
What shaped it
Changed
- The essay moved from “what evidence before entry?” to “who has to keep checking after entry?” because clinical AI had already entered through software channels and defaults.
- “Trust” was narrowed from an inferred psychological state to observable operational reliance: hospitals deploy and renew, clinicians use or sign, and patients inherit exposure.
- The first readable draft was rebuilt in a two-layer form: a general-reader body with exact receipts underneath, rather than specialist language carried through the whole page.
- The deep source check corrected nine mismatches or overstatements before publication, and the author moved that check earlier in the standing process for future issues.
Prevented
- A universal claim that no institution was checking locally. The public record often cannot show whether local checks occurred.
- Unvetted claims about clinicians signing without reading, patients “accepting” systems, or private motives behind adoption and renewal.
- Treating alert overrides, an unchanged-note rate, or adoption counts as direct measures of trust.
- A one-size-fits-all checking regime for a drafting tool, a risk score, and an action-triggering system.
Preserved
- The core distinction: installed is not the same as earned.
- The quiet-failure problem: drift, calibration failure, invented chart details, and alert habituation can hide behind familiar interfaces.
- The three-party split — institution, clinician, patient — without pretending they make the same decision.
- The first-line assignment: the hospital owns local checking because it chooses, deploys, and renews the system and holds stop authority at least on paper; vendor contracts and integration can make a feature hard to separate, and vendors remain answerable for what they build and disclose.
What this page is not
This Origin page is context, not certification. It shows the question, pressures, and changes behind the issue. It does not prove the published argument true, safe, or complete.
The source logs preserve the beginning and the final development phase, but not a continuous middle. The development record labels both gaps and does not reconstruct missing dialogue from the edit log or finished artifacts. The author’s sealed pre-process diagnostic remains unpublished by standing rule.