Parallel model engine
ML Lab
Open anomalies
0
data contradicting peers or own record
Open connections
0
implied by the ontology, stated nowhere
Last scan
—
weekly · Mondays 09:40 UTC · pg_cron
Learned models
2
score refresh weekly · retrain monthly
Engine runs
| Kind | Started | Finished | OK | Stats |
|---|---|---|---|---|
| model p_adverse_action_12m | 2026-07-29 | v0.1 | default | R23/W4-3. Deterministic logistic p(Medicare termination within 12m), SNF only. Label = first POS termination date (any code: voluntary closure, involuntary termination, merger exit); facilities already terminated by the panel date are excluded, not counted as safe. Trained on panels 2024-07/2024-10/2025-01 (n=46,466, 283 positives), evaluated on the held-out later panel 2025-04 BEFORE shipping: AUC 0.971, top-decile lift 9.5x, top-decile capture 95.4% of the 65 terminations that followed. Separability is high because closures announce themselves in the data: collapsing occupancy is the strongest signal, and the learned NEGATIVE weight on 12m deficiency counts is the surveys-stop signature of a facility winding down, not a claim that clean inspections are dangerous. Same 15 as-of-date features and anti-leakage chassis as p_chow_12m; SFF entry rejected as a label (only one month of SFF history exists, so as-of transitions cannot be reconstructed). Score = probability x 100; grade NULL. Not in any composite. |
| model p_chow_12m | 2026-07-29 | v0.1 | default | R22/W4-2. Deterministic logistic p(CHOW within 12m), SNF only. Trained on panels 2024-07/2024-10/2025-01 (n=50,700, 1,207 positives), evaluated on the held-out later panel 2025-04 BEFORE shipping: AUC 0.874, top-decile lift 5.0x (top 10% of predictions captured 49.6% of the CHOWs that actually closed in the next 12 months). Shipped coefficients are the evaluated ones; no post-holdout refit. Caveat: holdout positives (133) run below the training base rate because recent CHOWs surface in public records with a lag, so measured lift is conservative. Features are as-of-date public records: PBJ staffing level/trend/contract share (quarterly history to 2022Q1), deficiency and CMP counts (12m), ownership tenure and churn from CHOW-evidenced transactions, beds, occupancy, ownership type, chain size. Chain membership and ownership type are current-state (history not published). SFF and star ratings deliberately excluded (no as-of history). Notable learned direction: long ownership tenure RAISES sale odds and a recent prior CHOW lowers them. Score = probability x 100; grade is NULL because a probability is not a quality letter. Not in any composite. |
scans execute inside Postgres on pg_cron; score refreshes and retrains run in the API process, triggered over pg_net — the cadence survives sandbox loss
Anomalies — 0 open · Penalty spike
| Facility | Type | Signal | Setting | State | City | First seen |
|---|
trailing-12-month fines ≥$50k and ≥3x the facility's own prior 3-year annual average
Connections — 0 open
| Subject | Relation | Strength | Evidence | First seen |
|---|
relationships the ontology implies but no single source states