Community Health Market Watch · Datable Services
These queues are open to read — what the readers detected, what entity resolution is unsure about, and what the document reader pulled out of each filing with its quote. Deciding now has a screen: the analyst console — sign in, act, and every decision is stamped with your identity and written where it cannot be edited afterwards. Enrollment is deliberate: an account alone decides nothing. Detections sit at REPORTED_SECONDARY or lower until an analyst acts — machine-read from public filings, not yet checked, and some of them wrong.
Each milestone carries its definition of done from the build plan, what actually shipped, and a link to where the result lives on the platform. The critical path (M0→M1→M2→M3→M5→M6) is warm through M4; M5's depth waits on the contract by design. Statuses as of August 24, 2026.
Monorepo, Supabase (Postgres + content-addressed artifact storage), the append-only signal ledger with its dedupe key and mutation-rejecting triggers, the ingestion-run ledger, and the RLS deny-by-default boundary. Proven by a clean-room rebuild: the cached Madera decision artifact survived and re-verified by hash.
The entity graph (facility, organization, time-stamped ownership edges, geography, service lines) plus Madera's full ledger — 328 provable days of warning through closure, reopening, and the maternity unit that never returned. days_before_outcome matches the backtest row for row.
1,215 hospitals (CA + MA), 93.6% of open CA hospitals with full ownership chains from PECOS, birth volumes from HCAI 2019–2024 including the post-2021 form-change proxy, honestly labeled. Skilled nursing joined the registry on Aug 10 — 1,303 CA facilities from the iQIES file — because the AG docket's live transactions are nursing-home deals and the reader had nothing to resolve them to. SNF ownership and cost reports are separate CMS files, not yet loaded, and the explorer says so on every SNF record rather than showing an empty chart. The registry_* views are the deliberate public read layer.
Live: WARN (time-travel replay proven), HCRIS financials, the AG docket watcher, CMS ownership/CHOW, CMS iQIES for skilled nursing, HCAI utilization + public funding — and the document reader, graded against curated ground truth before being trusted with new documents. Stored extractions make re-reading free: --reresolve replays every document against the current registry with no model calls, which turned 40 held events into ledger rows the day nursing facilities landed. Honestly stubbed: SNF enrollments/owners, 990s (needs EINs), EMMA/ratings (needs the issuer watch list), AHA (needs the license). Remaining for full DoD: scheduled runs (cron) instead of manual.
Deterministic-first resolution with a human queue for everything fuzzy — 101 proposals queued, zero auto-applied, decisions reversible and ledgered. The Steward stress test reconstructed hospital → Steward entities → MPT-as-landlord entirely from public files, including the SEC-only sale-leaseback layer PECOS never records (diagram below).
Gated by design, not by engineering: deep tranche work is contract-scoped per the working understanding. Atlanta Medical Center is first after Madera — Linda will test the system against the case she knows. Tranche completion also triggers the demo-database split.
The dedicated predictive milestone, reframed per the Aug 10 1:1: closures are the pilot, and the artifact is the scored backtest beside the tranche's lead-time chart. Explainable-first: regularized logistic or gradient-boosted trees, every score decomposing into named reasons ("flagged because births fell below the level operators say forces closure"). Features are point-in-time honest — indexed by when a fact became publicly knowable, not when it happened, because cost reports land a year late and a backtest that ignores that lag would flatter itself. Ten cases makes the first validation directional, not publishable — the write-up will say so. The threshold library's engineered features double as the viral-card candidate generators (doc 09).
The seed is working: the three queues are on this page and open to read — resolution proposals with their evidence, ledger events with their workflow state, and reader extractions with verbatim quotes. Deciding runs through the pipeline tooling, because the gate is Postgres itself: SECURITY DEFINER functions that stamp the analyst's identity from their token, append-only writes, and anon denied on every write path. Remaining: a decision surface that doesn't put a password box on a public page, confidence-upgrade (supersede) flow, multi-analyst roles.
Writer drafts from APPROVED events only, every sentence carrying a source_id. Role taxonomy now includes union + employer segments and names CC Rapid Response as the tier-3 receiving channel; the shareable-card channel (receipt on the card + platform permalink) lands here too (doc 09).
The three views exist and are live (explorer, progress, internal) with facility permalinks throughout. The merchandising decision — map homepage, ticker, ranked watch list, cards, chat reader — deliberately waits for the stakeholder interviews; the candidate list is recorded as their stimulus set (doc 09).
ownership_chain(Nashoba Valley Medical Center, 2023-06-01). Today's PECOS holds only Steward's post-collapse residue — the chain lives in the January 2023 vintage, and the landlord layer lives in MPT's SEC-filed master lease. PECOS never records landlords.
| Adapter | Source | Yield so far | Status |
|---|---|---|---|
| edd/warn-notices | CA EDD WARN report (Tue/Thu) | first hospital WARN caught Aug 17: Sharp Memorial (San Diego), filed Aug 11, 20 employees — in the analyst queue | live |
| cms/hcris-cost-reports | HCRIS vintages 2017→latest | 74,746 metric rows · 2,814 facility-years | live |
| ca-dockets/ag-charities-watch | oag.ca.gov nonprofit hospital docket | 177 documents tracked · page-diff per run | live |
| cms/pos-hospitals + enrollments + owners + chow | CMS POS / PECOS (incl. historical vintages) | 1,570 hospitals (CA · GA · MA) · operators, parents, CHOW per state | live |
| cms/pos-snfs | CMS iQIES POS — long-term care | 1,303 CA nursing facilities · 1,440 status rows · 325 chains | live |
| cms/snf-enrollments + snf-owners | CMS SNF Enrollments / All Owners | 1,128 of 1,166 open SNFs with operators (97%) · 3,665 parent-owner edges · 54MB national file cached gzipped | live |
| hcai/annual-utilization + public-funding | CHHS open data + hcai.ca.gov | 1,862 volume rows · 20 DHLP awards | live |
| readers/ag-docket | AG PDFs — the hard 20% | 8 documents read · 2/2 vs curated ground truth · 40 events on re-resolve, $0 | live |
| courtlistener/bankruptcy-watch | Federal bankruptcy dockets via RECAP | built Aug 24 — conservative name-match on operators of open facilities; anonymous runs paced, token raises limits | built |
| cms/designation-watch | CAH status from POS snapshots | 1,569 baseline rows · 84 CAH facilities under watch (44 CA · 40 GA) · diffs clean | live |
| irs/form990 | ProPublica by EIN | blocked on EIN acquisition (M4 queue) | stub |
| emma + ratings + gdelt | bond disclosures · agencies · news | blocked on issuer watch list / reader | stub |
| aha/annual-survey | AHA (licensed) | activates with Community Catalyst's license | gated |
Everything the platform runs on, and what it actually bills. The distinction that matters: almost every data source is free, because the thesis is that the warning signs are already public. The money goes to infrastructure and to reading the hard 20% — the PDFs no structured feed covers. Figures are actuals where a bill exists and marked as estimates where one doesn't.
| Tool | What it does here | Basis | Cost |
|---|---|---|---|
| Infrastructure — recurring | |||
| Supabase (Pro) | Postgres, the append-only ledger, artifact storage, the API layer that enforces the approval gate, and analyst auth | $25/mo + usage over the included 8 GB database / 100 GB storage | $25/mo 39 MB db, 160 MB artifacts — far inside the allowance |
| Vercel | Hosts the three surfaces and the marketwatch.datableservices.com subdomain | Shared with the existing Datable Services account; static pages, no serverless functions | no incremental cost |
| Domain | marketwatch.datableservices.com | Subdomain of a domain already owned | $0 |
| Language-model reading — metered, the only variable line | |||
| Claude Opus 5 — readers/ag-docket | Reads AG docket PDFs and proposes dated, quoted events. The hard 20% no structured feed covers | $5 / M input, $25 / M output. Prompt caching on the system prompt; stored extractions make re-reads free | $9.89 to date 1.83M in / 29.7K out, 8 documents |
| — docket backlog | 169 documents still unread on the AG docket | Estimate at $0.74–1.24 per document, the observed range (long PDFs cost more) | $125–210 est. |
| — steady state | New filings as the docket updates, a handful per week | Same per-document rate; re-resolution against a changed registry costs nothing | ~$5–15/mo est. |
| Data sources — public records, free | |||
| CMS | Provider of Services (hospitals + iQIES long-term care), PECOS enrollment & ownership, CHOW, HCRIS cost reports | data.cms.gov open data API | $0 |
| CA HCAI | Hospital annual utilization, birth volumes, Distressed Hospital Loan Program awards | CHHS open data + hcai.ca.gov | $0 |
| CA DOJ / Attorney General | The nonprofit healthcare transaction docket — the earliest public signal we have | oag.ca.gov, scraped and cached | $0 |
| CA EDD | WARN layoff notices | Published report, twice weekly | $0 |
| SEC EDGAR | REIT master leases and sale-leasebacks — the ownership layer PECOS never records | Public filings | $0 |
| US Census | County geography, FIPS keys | Public files | $0 |
| IRS 990 via ProPublica | Nonprofit financials (blocked on EIN acquisition) | Free API | $0 |
| MSRB EMMA | Municipal bond disclosures — covenant breaches and rating actions | Free; blocked on building the issuer watch list | $0 |
| Licensed / not yet acquired | |||
| AHA Annual Survey | Service lines and system membership. The loader is scaffolded, not built | Institutional licence — Community Catalyst's to hold, not ours | not priced here |
| Ratings agencies | Moody's / Fitch / S&P actions on hospital debt | Press releases are free; full feeds are enterprise-priced | free tier only |
What would change the run-rate: a curated demo database (a second Supabase project, roughly doubling the base) once tranche 1 lands; and scheduled ingestion, which adds compute but not licence cost. Scoring and dispatch add no new vendors — they are computation over data already held.