Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,150 @@
+# FDA Warning Letter Escalation Dataset
+**Version:** 1.0.0 | **Records:** 14,810 | **Price:** $2,000 | **Source:** FDA (public domain)
+
+## Dataset Summary
+**The dataset's core signal — whether an inspection escalated to a Warning Letter — is validated against FDA's own severity classifications: OAI inspections escalate at 64% versus 1.7% for NAI, a 39× relationship that confirms the data reflects real regulatory behavior.**
+
+The 14,810 inspections that escalated to enforcement — every inspection followed by a Warning Letter, Injunction, or Seizure — with full citation context and exact timing (days from inspection to action, and 90/180/365-day windows). Pure signal: the dataset of what FDA findings actually precede enforcement. Built for RA/QA risk modelling and enforcement-pattern analysis.
+
+Current through 2026 — the dataset includes FDA inspections from the most recent regulatory cycles
+(covering 5,713 facilities inspected in 2026 and 14,845 in 2025), not a historical snapshot.
+Quarterly refreshes keep it current.
+
+## What This Replaces
+Comparable facility-level FDA regulatory intelligence from incumbent enterprise
+platforms requires annual contracts and enterprise procurement. This dataset
+delivers the same enforcement-escalation intelligence as a one-time,
+analysis-ready download — no contract, no seat licenses, no platform lock-in.
+
+## Source
+All data derives from U.S. FDA public records:
+- **FDA Data Dashboard API** — inspections, inspection citations, inspection classifications,
+  compliance actions (Warning Letters / Injunctions / Seizures), and import refusals.
+- **openFDA** — 510(k) device clearances, establishment registrations, device & drug recall
+  enforcement reports, and device classification.
+- All sources are U.S. Government works in the **public domain (CC0)**. No PII is included
+  (no patient, reporter, or individual-inspector identities).
+
+## The Premium Fields
+The premium fields are the **escalation linkage** — computed by matching each inspection to the
+nearest subsequent enforcement action at the same facility:
+
+- **`days_inspection_to_warning_letter`** — exact day gap from the inspection's end to the nearest
+  subsequent Warning Letter / Injunction / Seizure. Always positive (directionality enforced);
+  null if no action followed. The raw day count is provided so you can set your own thresholds.
+- **`escalated_within_90d` / `_180d` / `_365d` / `escalated_ever`** — convenience booleans derived
+  from the day gap.
+- **`escalated_warning_letter_action_type`** — which action followed (Warning Letter, Injunction, or
+  Seizure), so you can restrict to Warning-Letter-only escalation if you wish.
+- **`citation_count` / `cfr_codes_cited` / `citation_descriptions`** — the inspection's full citation
+  context, so the finding and its consequence sit in one row.
+
+## Methodology
+- **Entity resolution.** Records are joined on the FDA **FEINumber** (Facility Establishment
+  Identifier), normalized to a trimmed string across every source. 510(k) clearances and import
+  refusals carry an FEI directly.
+- **Escalation matching.** For each inspection, the dataset finds compliance actions at the *same
+  FEI* whose action date is **strictly after** the inspection end date, and keeps the **nearest**
+  one (smallest positive day gap). Directionality is enforced in the join itself, so a Warning
+  Letter that predates an inspection can never be counted as escalation. Escalation actions are
+  Warning Letters (the canonical case) plus the stronger judicial actions Injunction and Seizure;
+  the action type is retained so you can restrict to Warning-Letter-only.
+- **Risk score.** Computed from each facility's history with a published formula:
+
+  ```
+  facility_risk_score = min(100,
+      oai_count                              × 15
+    + warning_letter_count                   × 20
+    + recall_count                           × 10
+    + min(import_refusal_count × 2, 20)           (import-refusal contribution capped at 20)
+    + (repeat_citation_flag ? 10 : 0)
+    + (most_recent_classification == OAI ? 15 : 0)
+    + recency_factor)
+
+  recency_factor (awarded only when the most recent inspection outcome was OAI or VAI):
+    +10 if that inspection was within 1 year, +5 if within 3 years, +2 if within 5 years
+    (evaluated as of the 2026-06-23 build date)
+  ```
+
+- **Recall attribution.** Recall enforcement reports carry no FEI, so recalls are matched to
+  facilities by **normalized firm name** (uppercased, punctuation and common legal suffixes
+  removed). The match is deliberately restricted to **unambiguous 1:1 names** — a normalized name
+  owned by exactly one facility — to prevent a multi-site chain's recalls from being falsely
+  attributed to every location that shares its name.
+
+## Validation
+- **Escalation signal validated.** OAI (Official Action Indicated) inspections escalate to Warning
+  Letters at **64%** versus **1.7%** for NAI (No Action Indicated) inspections — a monotone,
+  ~39× relationship confirming the computed escalation fields correlate correctly with inspection
+  severity. (If escalation were a join artifact, NAI and OAI would escalate at similar rates.)
+- **Directionality enforced.** Zero inspections have a negative inspection→action day gap; every
+  Warning Letter counted as escalation post-dates its inspection.
+- **No orphaned escalations.** Every recorded escalation action type exists in the source
+  compliance-action data.
+- **Recall attribution restricted** to unambiguous 1:1 facility-name matches to prevent false
+  attribution (see Methodology).
+- **Completeness.** Row counts reconcile exactly to the validated master tables.
+
+## Known Limitations
+- **Citation coverage.** Inspection citations cover electronically generated, finalized
+  inspections; some manually prepared Form 483s are not present in the FDA citations dataset.
+- **Recall attribution (~35%).** Only recalls with an unambiguous 1:1 firm-name match are
+  attributed; multi-site chains that share a normalized name are deliberately excluded to avoid
+  false attribution, so recall counts are conservative (under- rather than over-counted).
+- **Inspector-level detail not included.** The FDA citations API does not expose individual
+  inspector identity, so inspector-level fields are out of scope for this version.
+- **Risk score is a heuristic.** `facility_risk_score` is a transparent composite indicator, **not**
+  an official FDA designation or prediction. All underlying counts are included so you can recompute
+  it with your own weights.
+- **Warning Letter coverage.** Warning Letters are attributed to facilities that appear in the
+  inspection record; Warning Letters issued to firms without an inspection in the dataset are not
+  rolled into a facility row.
+- **Escalation timing always populated here.** Because this dataset is filtered to inspections that
+  escalated, `days_inspection_to_warning_letter` and `escalated_warning_letter_action_type` are
+  populated on every row. In the full inspection dataset these fields are null by design for the
+  large majority of inspections that never escalate, reflecting that most FDA inspections do not
+  lead to enforcement.
+- **Null-by-design citation context.** The citation fields (`citation_count`, `cfr_codes_cited`,
+  `citation_descriptions`) can be empty/null for escalated inspections with no posted citations on
+  record. This is accurate — some inspections result in enforcement without posted 483 citations,
+  and many inspections classified NAI carry no citations at all. The `posted_citations` flag is
+  included so this is transparent.
+
+## Field Definitions
+| Field | Type | Description |
+|-------|------|-------------|
+| `fei_number` | string | FDA Facility Establishment Identifier (FEI) of the inspected facility. |
+| `inspection_id` | string | FDA Data Dashboard inspection identifier (not unique per row — an inspection may span multiple product/program rows). |
+| `legal_name` | string | Inspected facility legal name. |
+| `inspection_end_date` | date | Date the inspection ended. |
+| `fiscal_year` | string | FDA fiscal year of the inspection. |
+| `classification` | string | Inspection outcome classification (full text). |
+| `classification_code` | string | Inspection outcome code: NAI (No Action), VAI (Voluntary Action), OAI (Official Action). |
+| `project_area` | string | FDA project area for the inspection. |
+| `product_type` | string | Product area covered by this inspection row (Drugs, Devices, Foods, etc.). |
+| `posted_citations` | string | FDA flag indicating whether citations were posted for this inspection. |
+| `city` | string | Facility city. |
+| `state` | string | Facility state/province. |
+| `state_code` | string | Facility state code. |
+| `country` | string | Facility country. |
+| `firm_profile` | string | URL to the facility's FDA FirmProfile page. |
+| `citation_count` | integer | Number of citations (483 observations) recorded for this inspection. |
+| `cfr_codes_cited` | list<string> | Distinct CFR codes cited in this inspection. |
+| `citation_descriptions` | list<string> | Short descriptions of the citations issued in this inspection. |
+| `days_inspection_to_warning_letter` | integer | Days from this inspection's end to the NEAREST subsequent escalation action at the same facility (Warning Letter / Injunction / Seizure). Populated on every row in this escalation-filtered dataset (null by design only in the full inspection dataset, when no escalation followed). Always positive — directionality is enforced (the action must post-date the inspection). |
+| `escalated_warning_letter_action_type` | string | ActionType of the nearest subsequent escalation (Warning Letter, Injunction, or Seizure); populated on every row in this escalation-filtered dataset. |
+| `escalated_within_90d` | boolean | True if an escalation action followed within 90 days of the inspection. |
+| `escalated_within_180d` | boolean | True if an escalation action followed within 180 days. |
+| `escalated_within_365d` | boolean | True if an escalation action followed within 365 days. |
+| `escalated_ever` | boolean | True if an escalation action followed at any time after the inspection. |
+
+## Update Cadence
+Quarterly, tracking FDA Data Dashboard refreshes.
+
+## License
+Derived entirely from U.S. federal government public records (FDA Data Dashboard, openFDA), which
+are in the public domain. This structured, enriched version is provided under **CC BY 4.0** —
+attribution required.
+
+## How to Access the Full Dataset
+Full dataset available at [Ruby Intelligence on Gumroad](https://rubyintelligence.gumroad.com/l/vvafiy).