Digital pilot · regenerated 10 November 2026

Charlie Owl on a synthetic 274‑bed hospital

We run the real platform engines against Control General Hospital, a fully synthetic 274-bed facility, and publish what comes out. No real patients, no real hospital, and nothing hand-typed: every figure on this page is read from the machine’s own output, which reproduces byte for byte from the repository with two commands.

What Owl Found

Nine programs, 110 requirements, 52 gaps,
a 28-day path to close them.

Owl assessed the hospital’s nine accreditation, certification and designation programs as of 10 November 2026: which requirements are met, which are partial or not met, who owns each fix, and what closing them would take.

52
Requirements not fully met (22 not met, 30 partial) out of 110
Counted from the hospital’s own evidence
28 days
Critical path through the remediation roadmap
Every step has an owner role and a dependency
279 h
Modeled staff hours to close the gaps
Planning estimate, not measured performance
$32.7K–$48.9K
Modeled cost to close, including one external mock survey
BLS May 2025 wages; one finding refused a price rather than guessed

Every attestation and decision in the run is written as a receipt: 154 in all, and every one of the 49 on hash-chained record types verifies end to end.

Readiness Drifts

The same hospital, ten weeks later,
with nothing changed.

We ran the same hospital, with the same evidence, as of 2 September and again as of 10 November. Nothing new went wrong. Evidence simply aged, and surveys got closer. That drift is exactly what a binder and a pre-survey scramble miss, and what continuous readiness is for.

ProgramMet / partial, 2 SepMet / partial, 10 NovNext survey, 2 SepNext survey, 10 Nov
Joint Commission14 / 511 / 8120 days51 days
ACS Trauma Level II10 / 47 / 7560 days491 days
Primary Stroke Center9 / 47 / 6300 days231 days

Across all nine programs, requirements not fully met rose from 44 to 52 while the count outright not met stayed at 22. The modeled cost to close rose with them, from $31.2K–$44.4K to $32.7K–$48.9K.

What We Retracted

We audited our own July model
and published the corrections.

The first version of this pilot, published in July 2026, came from a hand-maintained record no code read. We replaced it with the machine runs above, and when the engines disagreed with what we had published, we printed both.

The avoided-harm estimate: $264,870 became $112,557

The July model priced an all-falls rate at the cost of a fall with injury, over-counting avoided injuries by about 3.8×. The falls line alone fell from $176,400 to $39,841. The engines print the published and corrected arithmetic side by side.

Faults injected on purpose

The run deliberately breaks things a real hospital would break: an unreadable evidence source, a role with no published wage, a step that may need capital. Owl refuses to guess, and says why, instead of filling the gap with a number.

Honest Limits

A control run, not a hospital deployment.

  • Every patient, record and staff member is synthetic. No real facility was assessed and no real PHI was used.
  • Hours and dollars are modeled planning estimates, not measured performance or an audited financial. Cost bands are summed independently; they are not confidence intervals.
  • Clinical models in the codebase that are not yet validated are labeled UNVALIDATED in their own code. A real pilot will measure evidence retrieval time, preparation effort and corrective-action completion.
The Full Paper

Read the digital pilot white paper.

Revision 3.1 (November 2026) sets out the method, every figure, every retraction and the commands that reproduce them. It is available on request.