Community Health Market Watch · prepared by Datable Services
The system has four layers: the graph (who owns and operates every facility, over time), the signal ledger (dated, sourced events per facility), the dataset (the retrospective case corpus the model will learn from), and dispatch (routing what matters to the people who can act). The graph and ledger are live. The dataset has its first case fully loaded. The model and dispatch come after the case tranche.
The work of the past week, newest at the top. Every item is verifiable in the repository history and the ingestion-run ledger.
What's moving now comes first, then what's finished (newest first), then what's queued. Each phase states what it delivered (or will), what it makes possible, and — where a phase is waiting — exactly which decision it waits on. Nothing here waits on engineering that isn't named.
Layoff notices, ownership changes, and the Attorney General's transaction docket are watched continuously. The document reader — a language-model component that reads docket filings and proposes dated, quoted events — is graded against the hand-transcribed Madera record before being trusted with new documents, and it recovers those events exactly. Its first pass over the live docket surfaced signals months earlier than news coverage, every one held for analyst review before it can reach anyone. Extractions are kept, so when the registry grows the filings can be re-read against it at no cost — that is how 40 previously unresolvable events landed the day nursing facilities arrived.
The review workflow is now a working screen: an analyst signs in and approves or rejects matches, moves detected events through review, and reads the document reader's extractions with their verbatim quotes. Every decision is recorded with the analyst's identity, permanently. This is the gate between everything the system detects and anything a partner or the public ever sees.
Every California hospital as a node in the graph — ownership chains from the government's own enrollment files (private-equity and REIT relationships flagged), birth-volume history from state utilization reports, five years of Medicare cost-report financials — and now every California skilled-nursing facility beside them, because that is where the transactions under review actually are. Every facility has a permanent address on the explorer that any report or card can link back to.
The same hospital appears under different names in every data source. Resolution is deterministic where the government provides keys, and everything uncertain waits for a human — 101 uncertain matches are queued for review right now, none applied automatically. As a stress test, the graph reconstructed the Steward Health Care pattern — hospital → operator → REIT landlord — entirely from public files.
The graph's data model, the append-only signal ledger, and the provenance rules: every ingested document is cached and hash-verified at intake, corrections never erase history, and the approval gate that keeps anything unreviewed off public surfaces is enforced in the data layer itself — not in anyone's memory.
Ten fully-documented retrospective cases — Madera's treatment applied to closures, service-line losses, and the survivors that didn't close (the controls that make the model's error rates honest). Atlanta Medical Center is next. The tranche produces the lead-time analysis: how much warning the public record actually gives, case by case.
The dedicated predictive milestone: facility closures are the pilot, per the Aug 10 working session. The scoring layer trains on the dataset and ships with published operating characteristics — sensitivity, specificity, calibration on held-out cases — and its artifact is a scored backtest beside the tranche's lead-time chart: the facilities that closed flagged early, the survivors correctly left alone. Scores use only what was publicly knowable on the scoring date, so the lead-time claim is honest by construction. With ten cases the first validation is directional, not publishable — and the write-up says so. Every score decomposes into named reasons an organizer can say at a hearing and an epidemiologist can audit. No score reaches anyone without analyst approval.
Role-based routing of approved intelligence: partner alerts, funder briefs, policymaker one-pagers, union and employer briefings, and Community Catalyst's Rapid Response as the receiving channel for the most urgent signals. Every packet drafts from approved events only, every sentence carrying its source — and every artifact links back to the live record on the platform.
Medicare cost-report history for every California and Georgia hospital is loaded and queryable. For Madera, the graph now shows what was visible all along: cash fell from $19.6M to $8.7M in two fiscal years, and the FY2022 loss was $11.6M — from the filings themselves, not from news coverage.
Every open California hospital resolves to its operator and owners; private-equity and REIT relationships are flagged from the government's own ownership records. As a stress test, the graph reconstructed the Steward pattern — hospital → Steward entities → Medical Properties Trust as landlord — entirely from public files, including the sale-leaseback layer that Medicare ownership records never capture (it was recovered from SEC filings).
The Attorney General's docket has moved to nursing homes: the deals under review right now are skilled-nursing acquisitions, chain roll-ups, and REIT sale-leasebacks. California's 1,303 nursing facilities are now in the graph alongside its hospitals, with the chain operator each one reports — Ensign, PACS, and 300-odd others — so those filings land on real facilities instead of going unread. Ownership chains for nursing facilities landed August 24 — 97% of open facilities now resolve to their operator, with the parent-owner layer above them. SNF cost reports remain unloaded, and the platform says so wherever that data would appear.
Madera Community Hospital is Case 01 in the dataset: 19 dated events from the first public warning through closure, reopening, and the maternity unit that never returned — every event citing a cached primary source, verified against the original document. The same treatment showed 814 → 694 births in the state's own utilization files before the count went to zero.
Layoff notices (WARN), Medicare ownership changes, and the Attorney General's transaction docket are monitored continuously, with every retrieved document cached at ingestion — insurance against public files disappearing. The document reader turns those filings into dated, quoted events, and it is graded against the hand-transcribed Madera record before it is trusted with anything new. The channel is catching real activity: the August 17 run picked up a WARN notice filed August 11 by Sharp Memorial Hospital (San Diego — 20 employees, per the state's own report), now in the analyst queue awaiting review. A fourth channel came online August 17: Critical Access designation is now tracked over time, so a hospital losing its enhanced-payment status — a revenue cliff that precedes distress by quarters — becomes a detectable event.
The signal ledger for Madera Community Hospital, counting down. Approximate dates (known only to the month) are marked with ~.