A hospital's sepsis software fires and somebody starts the bundle 79.7% of the time. Somebody finishes it 62.9% of the time. The 16.8 points in between are the escalation path — and they are the only part of it anyone measures.
In late September 2024, patients at Adventist Health in Bakersfield noticed a travel nurse behaving strangely. She was taking their IV opioids from a secured cabinet and charting them as given. The hospital was running machine-learning software built to catch exactly that pattern. It fired. Hospital managers ignored the alerts, auditors found.
There is no dataset of ignored diversion alerts. Hospitals are not required to disclose that they run this software, or to report when it fails, so no denominator exists. But there is one place where CMS does publish, hospital by hospital, what happened after a clinical trigger fired — and it tells the same story with 3,084 data points.
Each dot is one hospital. Horizontal: how often the first move landed. Vertical: how often the whole thing got finished. The dashed navy line is y = x — where a hospital would sit if starting and finishing were the same act. Hover any dot. Then drag the sample‑size floor.
The dots that look most alarming — the hospital at 0%, the hospital at 100% — are almost all small ones. Below 30 reported sepsis cases the standard deviation of completion is 22.8 points. At 300 cases or more it is 14.0. Nothing about care quality changed between those two groups. The arithmetic got quieter.
| Denominator | Hospitals | Mean SEP_1 | SD | Scoring 0 or 100 |
|---|---|---|---|---|
| Under 30 cases | 275 | 58.6 | 22.8 | 3.3% |
| 30–99 | 867 | 63.3 | 16.5 | 0.5% |
| 100–299 | 1,613 | 62.9 | 14.8 | 0.1% |
| 300 or more | 329 | 65.6 | 14.0 | 0.0% |
The gap itself, though, does not evaporate. It narrows from 19.0 points to 15.4 and stops. Every subgroup you can build out of this file — nonprofit, for-profit, government, tiny, enormous — finishes worse than it starts, by between fifteen and nineteen points. That is not a small-hospital problem. That is the shape of the thing.
1. A third of hospitals aren't here at all. The CMS file contains 4,660 hospitals. Only 3,106 have a SEP_1 score; 1,554 are blank, footnoted “too few cases,” “not available,” or “suppressed.” The hospitals most likely to have an escalation problem are the ones least likely to generate a publishable number.
2. This is chart abstraction, not behaviour. SEP_1 asks whether an element was documented within a window. It does not know whether a clinician read an alert, dismissed it, or never saw it. It is the closest public proxy for “did anyone act” that exists, and it is still several inferential steps away.
3. The measure that would matter isn't collected anywhere. There is no CMS measure — not one, out of the 24 in this file — for time‑to‑acknowledge, close‑without‑comment rate, or alert override volume. Every number on this page describes the outcome of an escalation path. None of them describe the path.
4. The correlations here are weak on purpose. Switch the vertical axis to median ED stay and you get r ≈ −0.19 — slower departments finish bundles slightly less often. Set the denominator floor to 100 and it strengthens to −0.25. That is a real signal and a thin one. It is exactly the size of effect that becomes a slide deck claiming crowding causes sepsis mortality, which this data cannot support.
The detector is the easy half. Sepsis surveillance is a solved detection problem — the trigger logic is twenty years old and it fires. Diversion surveillance is a solved detection problem too; at Bakersfield it fired for weeks. What both stories have in common is that everything downstream of the firing was unowned, uninstrumented, and invisible to the vendor.
If your product can be fully ignored and never tell you, you don't have telemetry. You have a dashboard.Builder's Briefing, August 26, 2026
The instrumentation nobody ships: who received it, what their queue looked like when it landed, what the false‑positive rate did to their trust by week six, whether closing it required a reason, and whether anyone ever audited the closures. None of that is in a model card. All of it decides whether the model does anything at all.