Nurses are picketing an allocation model today. So here is an allocation model's output, measured: 92 consecutive days of payroll-verified nurse staffing, every certified nursing facility in America. The stripe is the weekend.
Every AI governance conversation runs on the models that suggest. The models that allocate came in through operations. No eval, no shadow-mode pilot, no abstention threshold — because it's just staffing.
There is no public dataset of nurse-scheduling model outputs. But there is one place where American healthcare publishes, day by day, exactly how many nursing hours it actually bought: the CMS Payroll Based Journal. Not a survey. Not a self-report. Payroll records, submitted quarterly, auditable.
Here is one quarter of it. Every square is one day, nationally. Darker is more registered-nurse time per resident.
Across those 92 days the average number of residents in America's nursing facilities on a weekday was 1,219,165. On a weekend it was 1,216,998 — two-tenths of one percent fewer. The people are still in the building. The RN hours are not.
The gap is 11.9% for registered nurses and 8.8% when you count every nursing role. RNs — the most licensed, most expensive, most assessment-capable staff — are the category that thins the most.
And look at July 4th, a Thursday. It staffs 9.1% below every other Thursday in the quarter. Whatever is driving this pattern is reading a calendar, not a chart.
Split the same quarter by facility size. These are the pooled weekday and weekend RN hours per resident-day for each size band:
Small facilities barely move. The largest ones cut nearly a fifth of their RN coverage every weekend. That is the opposite of what you'd expect if the weekend gap were about small buildings struggling to fill a shift — and exactly what you'd expect if it were about scheduling systems optimizing a labor budget at scale.
Ask the vendor of any staffing or census-projection tool one question: what is the ground truth you validated against? For a diagnostic model the answer is a biopsy. For a staffing model, the only available answer is what happened under the staffing the projection produced. The model helps create the outcome it is scored against. That isn't a bug in a particular product — it's a property of the whole category, and it is why “our model is 94% accurate” means something different here than it does in radiology.
These are nursing homes, not hospitals. PBJ covers CMS-certified skilled nursing facilities. The software the union named runs in acute-care hospitals, which have no equivalent daily public staffing file. This is the closest measurable analogue, not the same setting.
The national line hides an enormous distribution. Computed facility by facility (13,997 facilities with a full quarter of data), the weekend RN gap has a median of +8.2% but an interquartile range of −9.7% to +25.6%. Roughly 63% of facilities staff RNs lower on weekends; more than a third staff them higher. The clean 11.9% national figure is a pooled aggregate, not a description of any individual building.
Small facilities generate most of the noise. Median weekend gap is +2.1% at facilities under 50 residents and +13.4% at facilities over 100. At a 30-bed building, one RN shift moves the percentage by double digits; the tails of that facility-level distribution are mostly arithmetic, not policy. If you rank facilities by weekend gap and read the top of the list as “worst offenders,” you will mostly be ranking small denominators.
Hours are not care. RN hours per resident-day says nothing about skill mix, acuity, agency vs. employed staff, or what those hours were spent doing. And PBJ has no acuity adjustment at all: two facilities with identical hours can be carrying very different patients.
Correlation is not a scheduling model. Nothing in this data identifies why weekends are thinner. Nurse preference, weekend differentials, elective admission patterns and deliberate budget policy would all produce this stripe. The point is not that software caused it. The point is that this pattern is stable, national, decades old, invisible in any clinical governance process — and that it is precisely the kind of pattern a staffing model would learn from and reproduce.
Read the defaults in the three order sets you use most — not the alerts, the pre-filled quantities. Then read the default in your unit's weekend template. Same category of artifact. Only one of them has a governance committee.