PA.claims Technical Architecture & Data Moat · Quest Public Insurance Adjusters (#W573247)
Court-Verified Estimate Auditing & Level-0 Immutable WORM Data Capture Platform
How PA.claims ingests, slices, classifies, and cent-reconciles 312,100 public court & statutory PDFs (121,072 scanned paths, 87,058 distinct SHA-256 documents, 275,182 classified spans, 1,644,614 EAV tags) into an immutable PostgreSQL WORM vault, preserves thin carrier patch scopes (<25 lines), and projects evidence through 6 pluggable SQL Value Lenses.
Distinct SHA-256 Documents
87,058
121,072 Scanned Paths · 312,100 Disk PDFs
Classified Page Spans
275,182
Across 24 Canonical Classes (CLASS_01..24)
EAV Dimension Tags
1,644,614
Asserter, Posture, Epistemic Status & Trade
Opposing Estimate Pairs
106 Stored
75 G-C Pairs (39 Both ≥25 + 36 Thin Carrier)
Core Pipeline · Deterministic ECF/ROA Slicer & 7-Stage Conjunction Gate
1. Deterministic Document Slicing (S1_SLICE) & Estimate Classification (S2_CLASSIFY)
Omnibus federal litigation dockets and state appellate Records on Appeal (ROAs) pack 500 to 3,000 pages—mixing carrier estimates, public adjuster estimates, contractor invoices, settlement checks, policy endorsements, engineer reports, and appraisal awards inside a single PDF. All ingestion runs through a PostgreSQL job queue (pipeline.document and pipeline.stage_run using SELECT ... FOR UPDATE SKIP LOCKED):
| Stage | Gate Name | Deterministic Rule & Anti-Contamination Guardrail |
|---|---|---|
| Stage S1 | Deterministic ECF Header + ROA Index Slicer (S1_SLICE) |
Before any OCR or LLM classification runs on an omnibus PDF, S1_SLICE parses federal blue ECF headers (Case X:XX-cv-XXXXX Document N Filed MM/DD/YY Page P of Q) and state Record-on-Appeal (ROA) index tables deterministically in Python (PyMuPDF), falling back to header-continuity scoring only when neither exists. |
| Stage S2 | 7-Stage Conjunction Gate (S2_CLASSIFY) |
Admits a slice as a CARRIER_ESTIMATE or CLAIMANT_ESTIMATE only when it passes pikepdf XMP producer checks + structural table column headers (Description, Qty, Unit Price, Tax, O&P, RCV, Deprec., ACV) + summary math block (RCV - Dep = ACV) + negative contamination gates (excluding contractor invoices, orders, CRNs, and checks) + author_class joined against OIR / DFS PA / DBPR license rosters. |
| Stage S3 | Cent-Exact Line Parser (S3_PARSE_LINES) |
Extracts line items in Python Decimal (ROUND_HALF_UP) via xact_gate. Unparseable tokens remain NULL with parse_state = 'UNRESOLVED' (never coerced to $0.00), and observed_cat_code is strictly separated from inferred_cat_code. |
| Stage S4 & S5 | Case Pairing & Outcome Extraction (S4_PAIR_CASE / S5_EXTRACT_OUTCOME) |
Links opposing estimates and outcome instruments (appraisal awards, summary judgment orders, settlement stipulations) using multi-anchor docket, claim number, and situs resolution, plus cross-linking Florida DFS PIITIL pre-suit notices against federal removal ECF 1 filings. |
| Stage S6 | Tier Hierarchy (G-A..G-D) & Thin-Carrier Preservation |
• G-A (TIER_A_ITEMIZED_OUTCOME_TRIPLE): Carrier Scope + Claimant Scope + Itemized Award/Order + Policy (3 complete cases currently reconciled end-to-end: Walters v. Wilshire/IAT, Windsor Village, Contour Buck/Call/Carl).• G-B (TIER_B_LUMP_SUM_OUTCOME_TRIPLE): Opposing scopes + lump-sum judgment/settlement (Villamar v. Lexington, etc.).• G-C (TIER_C_ADVERSARIAL_PAIR_NO_OUTCOME): 75 pairs (39 with ≥25 lines on both sides + 36 where the PA has ≥25 lines and the Carrier wrote a thin <25-line scope—thin carrier scopes are never rejected, because carrier omission is the core signal we measure).• G-D (TIER_D_SINGLE_SIDE_PRIOR): 31 incomplete pairs + single-side PIITIL demand estimates.
|
Upstream Data Capture · Compliant Public-Record Ingestion
2. Six Upstream Data Capture Feeds & Strict Rate-Limit Governance
| Upstream Source | Ingestion Method & Rate Limit | Primary Data Captured | Downstream Product Use |
|---|---|---|---|
| 1. CourtListener / Free Law Project (RECAP) | Offline S3 Bulk CSVs (com-courtlistener-storage) + REST API (125 calls/day (5/min, 50/hr, 125/day); hard-capped ≤100 calls/hr on authenticated REST) |
Federal property insurance & commercial receivership dockets, complaints, estimate exhibits, and appraisal awards. | Populates CLASS_01..24 spans, opposing estimate pairs, and commercial receiver operating reports. |
| 2. Florida DFS PIITIL & Ch. 119 Exports | Polite 2 → 4 worker attachment-HTML pre-filtering + formal Ch. 119 bulk SQL/CSV export |
Pre-suit Property Insurance Intent to Initiate Litigation notices (2+ attachments with carrier/denial/estimate hints). |
Early-warning carrier dispute radar and PIITIL ↔ Federal Removal ECF 1 cross-linking. |
| 3. FL Supreme Court AOSC24-65 Bulk Clerk Feeds | Notarized bulk attorney/party access (0 CAPTCHA scraper workers) |
State Circuit Civil property insurance dockets and attached carrier/PA estimate exhibits (Palm Beach, Broward, Miami-Dade). | Captures in-state domestic carriers (Citizens, Universal, HCI/TypTap, Heritage, Slide). |
| 4. FL OIR IRFS / QUASR / ISU Telemetry | Quarterly regulatory filings & market-share tables | County-level policies in force, written premium, non-renewals, and claims closed with/without payment. | Normalizes litigation counts against actual market share in terminal.regulator_metric. |
| 5. OpenFEMA NFIP & NOAA/NWS ASOS | Bulk redacted claims API + ASOS weather station swaths | Historical flood losses, peak wind gusts (KNEW, KMSY, KPBI), and hail diameter swaths. |
Corroborates date-of-loss weather triggers for roof/envelope audits and renewal briefs. |
| 6. SEC EDGAR 10-K / 10-Q / 6-K API | Official XBRL & full-text API (≤10 req/sec) |
Public insurer loss ratios/reserves and Sunbelt multifamily REIT (MAA, IRT, FSV) per-door operating expenses. |
Provides institutional $/door insurance, tax, and utility benchmarks for commercial audits. |