Substrate & order sets: real data
(GREEN)
70,563 real Phase-0 orders · 3 hoods · 184 days
(2026-04-05→10-05) · 334 signatures: actual Need_By, priority, BUD, and
9.7% cancellation rate
all completed original preps, full facility -
no hood extrapolation (scenario world)
two fully-simulated worlds on the SAME orders,
technicians, and days: Eagle IV (real-time alerts) vs a no-real-time-alerting
passive baseline (cameras capture, no alerts; pharmacists verify from the
photos). 184/184 days each pass the invariant suite + B=projection(A)
SCENARIO dataset: scripted degrading-tech illustration
+ STAT day-boundary dispatch fix: canonical data & dashboards
untouched
every technician action logged with timestamps,
images, and CV confidence: the value is read from that observed stream,
never from ground truth (structurally unreachable)
How errors &
detection are modeled (YELLOW: the generative assumptions)
errors EMERGE per prep: per-technician propensity ×
fatigue × interruption over the real orders: distributions reported, never
blanket rates tuned to an outcome
volumetric: out-of-tolerance at |dev| > 10% (ISMP);
Eagle optical alert at 8%; counted as COMMITTED errors only (a tech's own
glance-and-redraw is not an error): the ±10% line is load-bearing and
drug-flat here (a candidate for per-drug-class tolerance)
high-alert biologics error at half propensity
(extra bench scrutiny / double-checks)
documentation-capture failure 0.00195 per critical
presentation (MEDMARX ~30% doc share); recalled-lot = one scripted event
per study
EAGLE detection: materials scan (identity/expiry/lot),
3-camera temporal technique flag, step-sequence omission catch, Syringe
Vision: all prevented in-hood; only septum-coring / touch-contamination
can escape
passive-baseline detection: photo-review is
MAGNITUDE-GATED: 95% at the ±10% floor → 99% by 20% deviation · expiry 50% ·
recalled lot 0% (no human lot cross-check) · process/motion 5% (a still can't
show a missed swab)
How dollars &
hours are valued (the dashboard's own method)
materials valued ORDER-BY-ORDER: each remade prep at
ITS own bill of materials (real vial counts × acquisition cost); a rework
wastes $3 consumables only (the drug is retained): NOT a flat $/event
labor from MEASURED in-hood hands-on cycle
times
acquisition costs (ref_price_list.csv) are per-vial
ESTIMATES anchored to public price points (pip-tazo $3.13, AmBisome $328,
micafungin $47, voriconazole $25/vial; IVIG $80/g): REPLACE with the site
wholesaler export
wages locked: technician $38/h · pharmacist $138/h ·
annualization ×2 (184 d → yr)
biologic mix-shift (Lever A): reweighted to the
non-hazardous $4.53M/hood benchmark at CONSTANT volume · cancellation
intraday TIMING is synthetic (uniform): not modeled
Scenario
note: built from the SCENARIO worlds (days_scenario vs
days_baseline_scenario): a scripted degrading-technician illustration plus
the STAT day-boundary dispatch fix. Illustrative scenario: one
technician's simulated performance decline, shown from observed Eagle
data.
EagleIV
by NeoRx