v2: Add financial, events, curated docs; upgrade 604 DS10 files

PROVENANCE.md

@@ -1,145 +1,131 @@
 # Data Provenance
 
-This document describes the origin, processing method, and lineage of every table in the Epstein Document Archive.
+Per-table documentation of data sources, extraction methods, and quality notes.
 
-## Overview
+## documents
 
-The dataset combines two OCR pipelines and multiple community-curated layers:
+**1,424,673 rows.** One row per PDF file from the DOJ Epstein release.
 
-| Source | Files | Method | Quality |
-|--------|-------|--------|---------|
-| Gemini OCR | 848,228 | Gemini 2.5 Flash Lite | High (structured JSON output) |
-| Community Tesseract | 537,622 | Tesseract OCR via community repos | Variable |
-| Community curation | -- | Manual + automated analysis | Curated |
-| ML recovery | 39,588 pages | Redaction recovery model | Experimental |
+- **856,028 files** processed with **Gemini 2.5 Flash Lite** ($0.10/$0.40 per 1M tokens). Full structured extraction: document type classification, date parsing, entity extraction, handwriting/stamp detection, photo description.
+- **531,279 files** (DataSet 9) imported from the [rhowardstone/Epstein-research-data](https://github.com/rhowardstone/Epstein-research-data) community project using **Tesseract OCR**. Raw text only — no entity extraction or document classification. These have `ocr_source = 'tesseract-community'`.
+- **1,377 files** originally Tesseract, upgraded to Gemini in v2 (604 from DS10, 773 from other datasets).
+- **37,369 files** have `is_photo = true` (photos, stamps, blank pages).
+- `email_fields` column (new in v2): JSON-encoded email metadata (from, to, cc, subject, date) for email-type documents.
 
-**Success rate: 99.97%** (1,385,850 / 1,386,322 attempted).
+Distinguish OCR source: `ocr_source IS NULL` = Gemini, `ocr_source = 'tesseract-community'` = community.
 
-## Per-table provenance
+## entities
 
-### `documents` (1,413,765 rows)
+**10,629,198 rows.** Named entities extracted from documents.
 
-| Subset | Rows | Source | Method |
-|--------|------|--------|--------|
-| `ocr_source IS NULL` | 848,228 | DOJ FOIA PDFs | Gemini 2.5 Flash Lite (`gemini-2.5-flash-lite`) via Google AI API. Each PDF page sent as image, returns structured JSON with full text, document type, date, entity extraction, and metadata. |
-| `ocr_source = 'tesseract-community'` | 537,622 | Community repositories | Tesseract OCR. Primary source: [rhowardstone/Epstein-research-data](https://github.com/rhowardstone/Epstein-research-data) for DataSet 9 (531K files). Remaining community files fill gaps in DataSets 2-5, 12, FBIVault, and HouseOversightEstate. |
+- Gemini documents: entities come from the structured JSON extraction prompt (types: person, organization, location, date, reference_number, email_address, phone_number, monetary_amount).
+- Community documents: entities from post-processing NER pipeline.
+- Entity `normalized_value` provides cleaned/canonical forms where available.
 
-**Identifying OCR source**: Check the `ocr_source` column. NULL = Gemini, `"tesseract-community"` = community Tesseract.
+## chunks
 
-### `entities` (8,542,849 rows)
+**2,193,090 rows.** Text chunks for RAG (retrieval-augmented generation).
 
-Extracted entities (people, organizations, locations, dates, reference numbers) linked to their parent document via `file_key`.
+- ~800 token target per chunk with overlap at sentence boundaries.
+- `char_start` and `char_end` map back to the parent document's `full_text`.
 
-- **Gemini documents**: Entities extracted by Gemini as part of OCR (structured JSON output includes entity arrays).
-- **Community documents**: Entities extracted by post-processing NER on Tesseract OCR text.
-- **Entity types**: `person`, `organization`, `location`, `date`, `reference_number`.
+## embeddings_chunk
 
-### `chunks` (2,039,205 rows)
+**2,111,356 rows.** 768-dimensional float32 vectors from `gemini-embedding-001`.
 
-Text chunks derived from `documents.full_text` using ~800-token splitting with overlap. Used for retrieval-augmented generation (RAG).
+- ~96% coverage (documents with malformed text excluded).
+- Stored as `list<float32>` in Parquet.
+- `source_text_hash` links to the chunk text that was embedded.
 
-- **Source**: Derived from the full text of each document.
-- **Method**: Token-count-based splitting with character offset tracking.
-- **Fields**: `file_key`, `chunk_index`, `content`, `token_count`, `char_start`, `char_end`.
+## persons
 
-### `embeddings_chunk` (1,956,803 rows)
+**1,614 rows.** Curated person registry with canonical names, aliases, and categories.
 
-768-dimensional float32 embedding vectors for text chunks.
+- Categories: perpetrator, victim, associate, other.
+- `aliases` field: JSON array of known alternate names/spellings.
+- `search_terms`: additional search patterns for entity resolution.
+- Community-curated from multiple sources.
 
-- **Model**: `gemini-embedding-001` (Gemini Embedding v1).
-- **Input**: `chunks.content` (the chunk text).
-- **Coverage**: 96% of chunks (1,956,803 / 2,039,205). 1,249 malformed embeddings excluded.
-- **Format**: In Parquet, stored as `list<float32>` with fixed size 768. In SQLite, stored as raw `float32` BLOB (3,072 bytes per embedding).
-- **Note**: Summary embeddings (one per document) were removed from the published dataset. 92% of documents contain a single chunk, making summary and chunk embeddings identical. For multi-chunk documents, use the first chunk embedding or mean-pool across chunks for a document-level vector.
+## kg_entities / kg_relationships
 
-### `persons` (1,614 rows)
+**467 entities, 2,198 relationships.** Knowledge graph connecting people, organizations, and locations.
 
-Curated registry of people mentioned across the corpus.
+- Relationship types: traveled_with, associated_with, employed_by, legal_representative, financial_connection, etc.
+- `weight` indicates strength of connection (co-occurrence frequency).
+- `evidence` field links to source document file_keys.
 
-- **Source**: Community curation, combining automated entity extraction with manual verification.
-- **Fields**: `canonical_name`, `slug`, `category` (perpetrator/victim/associate/other), `aliases` (JSON array), `search_terms`, `sources`, `notes`.
-- **Upstream**: `release/persons_registry.json`.
+## recovered_redactions
 
-### `kg_entities` (467 rows)
+**37,870 rows.** Text recovered from under redaction bars using ML reconstruction.
 
-Knowledge graph entities with metadata.
+- `interest_score` (0-100) ranks significance of recovered content.
+- `names_found` lists person names detected in recovered text.
+- Experimental quality — treat as leads, not verified text.
 
-- **Source**: Community curation -- manually identified key entities with occupation, legal status, and mention counts.
-- **Fields**: `name`, `entity_type`, `description`, `metadata` (JSON with occupation, legal_status, mention counts).
-- **Upstream**: `release/knowledge_graph_entities.json`.
+## financial_transactions (NEW in v2)
 
-### `kg_relationships` (4,190 rows)
+**49,770 rows.** Credit card and bank transaction records extracted with DeepSeek.
 
-Knowledge graph relationships between entities.
+- Source: DOJ-released credit card statements and bank records (DataSet 10/11).
+- Extraction model: DeepSeek (per-page structured extraction).
+- 31% of raw extractions quarantined for quality issues (hallucinated amounts, duplicate entries, garbled OCR). Only clean records included.
+- Flight fields (flight_from, flight_to, flight_carrier, flight_passenger) populated for airline purchases.
+- Key cardholders: JEFFREY E EPSTEIN, GHISLAINE MAXWELL, KARYNA SHULIAK, HBRK ASSOCIATES, TERRAMAR PROJECT.
 
-- **Source**: Community curation -- relationships extracted from document analysis.
-- **Fields**: `source_name`, `target_name`, `relationship_type` (e.g., traveled_with, associated_with, employed_by), `weight`, `evidence`, `metadata`.
-- **Relationship types**: traveled_with, associated_with, employed_by, legal_representative, financial_connection, and others.
-- **Upstream**: `release/knowledge_graph_relationships.json`.
+## communication_records (NEW in v2)
 
-### `recovered_redactions` (39,588 rows)
+**128 rows.** Phone call and cell-site/CDR records.
 
-Text recovered from behind redaction bars in DOJ documents.
+- Source: DOJ-released phone records.
+- 99.8% of raw extractions quarantined (CDR data has very high hallucination rates). Only 128 high-confidence records included.
+- Fields: call date/time, duration, direction, location, number called, provider.
 
-- **Source**: Community ML analysis of redacted document pages.
-- **Method**: Machine learning model trained to reconstruct text obscured by redaction bars.
-- **Fields**: `file_key`, `page_number`, `reconstructed_text`, `interest_score`, `names_found`, `document_type`.
-- **Quality**: Experimental. Higher `interest_score` indicates more significant recovered content.
-- **Upstream**: `release/redacted_text_recovered.json.gz`.
+## investigative_records (NEW in v2)
 
-### `provenance` (pipeline audit trail)
+**143 rows.** Law enforcement reports, evidence recovery logs, and vehicle/property records.
 
-Three sub-tables from the OCR processing pipeline:
+- Source: FBI, PBSO, and other agency reports in DOJ release.
+- 87% of raw extractions quarantined. Only 143 verified records included.
+- Record types: law_enforcement_report, evidence_recovery_log, vehicle_property_record.
 
-| Sub-table | Rows | Description |
-|-----------|------|-------------|
-| `provenance/files` | 1,386,322 | Per-file processing record: SHA-256 hashes, status, tokens, latency, model |
-| `provenance/audit_log` | ~3.6M | Append-only event log: every processing step recorded |
-| `provenance/runs` | 120 | Pipeline run metadata: start/end times, success/failure counts |
+## derived_events (NEW in v2)
 
-- **Source**: `epstein_provenance.db` -- the forensic audit trail from the OCR pipeline.
-- **Key fields on `files`**: `file_key`, `pdf_sha256`, `output_sha256`, `status`, `input_tokens`, `output_tokens`, `api_latency_ms`, `model_used`.
+**3,038 events** with **5,751 participants** and **21,910 source document links.**
 
-## Dataset-level breakdown
+Three analysis tracks reconstruct Epstein's activities:
 
-### DOJ FOIA Releases
+- **Calendar track**: Meetings, dinners, appointments from Lesley Groff's daily schedule emails (2011-2016).
+- **Travel track**: Flights, hotel stays, ground transport from schedule transitions and AmEx bookings.
+- **Financial track**: Gift exchanges, institutional donations, major purchases from bank statements and emails.
 
-| Dataset | Total Files | Gemini | Community | Description |
-|---------|-------------|--------|-----------|-------------|
-| DataSet 1 | 3,158 | 3,158 | 0 | Initial release |
-| DataSet 2 | 574 | 49 | 525 | |
-| DataSet 3 | 67 | 49 | 18 | |
-| DataSet 4 | 152 | 49 | 103 | |
-| DataSet 5 | 120 | 49 | 71 | |
-| DataSet 6 | 13 | 13 | 0 | |
-| DataSet 7 | 17 | 17 | 0 | |
-| DataSet 8 | 10,595 | 10,595 | 0 | |
-| DataSet 9 | 531,279 | 0 | 531,279 | Entirely community-processed |
-| DataSet 10 | 503,154 | 502,548 | 606 | |
-| DataSet 11 | 331,655 | 331,651 | 4 | |
-| DataSet 12 | 152 | 50 | 102 | |
+Each event links to source EFTA documents via `event_sources` and participants via `event_participants`.
 
-### Non-DOJ Sources
+## curated_docs (NEW in v2)
 
-| Dataset | Files | Source | OCR |
-|---------|-------|--------|-----|
-| FBIVault | 22 | FBI Vault FOIA | Community Tesseract |
-| HouseOversightEstate | 4,892 | House Oversight Committee | Community Tesseract |
+**5,766 gold documents** across 5 investigation subjects: Hoffman (1,526), Gates (2,069), Summers (739), Clinton (765), Black (667).
 
-## Processing failures
+- Each document annotated with: tier (NUCLEAR/CRITICAL/HIGH/MEDIUM/SUPPORTING), category, date, sender, recipient, headline, key quote, investigative detail.
+- `also_appears_as`: JSON array of duplicate EFTA file_keys for the same document.
+- Only gold-status documents exported (24,706 rejected documents excluded).
 
-472 source PDFs could not be processed. These are documented in `release/epstein_problems.json` with:
-- `file_key` -- EFTA identifier
-- `doj_url` -- Original DOJ download URL
-- `error_message` -- Why processing failed
-- `category` -- Failure type (empty_source_pdf, corrupt_source_pdf, api_disconnect, doj_file_unavailable)
+## provenance/files
 
-## Reproducibility
+**1,387,775 rows.** One row per processed file with full pipeline metadata.
 
-The full processing pipeline is available at [github.com/kevinnbass/epstein](https://github.com/kevinnbass/epstein):
-- `process_pdfs_gemini.py` -- Gemini OCR pipeline
-- `import_community.py` -- Community data importer
-- `validate_dataset.py` -- 14-check validation suite
-- `epstein_audit.db` -- Complete audit trail with per-file SHA-256 checksums
+- `pdf_sha256`: SHA-256 hash of input PDF.
+- `output_sha256`: SHA-256 hash of output JSON.
+- `model_used`: OCR model (gemini-2.5-flash-lite, tesseract-community, etc.).
+- `input_tokens` / `output_tokens`: API token consumption.
+- `validation_score`: automated quality score (0-100).
 
-Every output JSON has a SHA-256 checksum recorded in the provenance database, enabling verification that published data matches pipeline output.
+## provenance/audit_log
+
+**3,711,609 rows.** Append-only forensic audit trail.
+
+- Every pipeline action (file processed, error, retry, fix applied) is logged.
+- `checksum` field provides tamper detection on critical operations.
+- Timestamps in ISO-8601 format.
+
+## provenance/runs
+
+**123 rows.** Pipeline execution records with timing, worker counts, and costs.

README.md

@@ -1,178 +1,196 @@
-# Epstein Document Archive
-
-**1.39 million OCR'd documents** from the DOJ Jeffrey Epstein document release, with extracted entities, text embeddings, a knowledge graph, and full pipeline provenance.
-
-This is the data behind [epstein.academy](https://epstein.academy).
-
-## What's in the dataset
-
-| Layer | Rows | Download Size | Description |
-|-------|------|---------------|-------------|
-| `documents` | 1,413,765 | ~800 MB | Full text of every document with metadata |
-| `entities` | 8,542,849 | ~200 MB | Extracted people, organizations, locations, dates |
-| `chunks` | 2,039,205 | ~1.5 GB | ~800-token text chunks for RAG |
-| `embeddings_chunk` | 1,956,803 | ~5 GB | 768-dim Gemini embeddings per chunk |
-| `provenance` | 4.9M rows | ~400 MB | Full pipeline audit trail |
-| `persons` | 1,614 | <1 MB | Curated person registry with aliases |
-| `kg_entities` | 467 | <1 MB | Knowledge graph entities |
-| `kg_relationships` | 4,190 | <1 MB | Knowledge graph relationships |
-| `recovered_redactions` | 39,588 | ~3 MB | Recovered text from redacted pages |
-
-**Total: ~8 GB compressed** (Parquet with zstd). Each layer is independent -- download only what you need.
-
-## Quick start
-
-### With HuggingFace `datasets`
+---
+license: cc-by-4.0
+task_categories:
+  - text-classification
+  - question-answering
+  - feature-extraction
+language:
+  - en
+tags:
+  - legal
+  - ocr
+  - documents
+  - foia
+  - knowledge-graph
+  - financial
+size_categories:
+  - 1M<n<10M
+pretty_name: "Epstein DOJ Document Archive (OCR + Structured Data)"
+---
+
+# Epstein DOJ Document Archive v2
+
+**1.42 million OCR'd documents** from the Department of Justice Jeffrey Epstein document release, with structured entity extraction, vector embeddings, financial transactions, communication records, and a forensic audit trail.
+
+Frontend: [epstein.academy](https://epstein.academy)
+
+## What's New in v2
+
+- **10.6M entities** (up from 8.5M) — expanded NER extraction
+- **2.1M chunk embeddings** (up from 1.96M) — more documents embedded
+- **49,770 financial transactions** — credit card and bank records (DeepSeek extraction)
+- **3,038 derived events** — reconstructed calendar, travel, and financial timeline
+- **5,766 curated gold documents** — expert-annotated research catalog across 5 subjects
+- **143 investigative records** — law enforcement reports and evidence logs
+- **128 communication records** — phone call and CDR data
+- **604 DS10 files upgraded** from Tesseract to Gemini OCR
+
+## Quick Start
+
+### HuggingFace Datasets (streaming)
 
 ```python
 from datasets import load_dataset
 
-# Stream documents (no full download needed)
-ds = load_dataset("kabasshouse/epstein-data", "documents", split="train", streaming=True)
-for doc in ds:
-    print(doc["file_key"], doc["dataset"], doc["document_type"])
-    print(doc["full_text"][:200])
+# Stream documents without downloading everything
+ds = load_dataset("kabasshouse/epstein-data", "documents", streaming=True)
+for doc in ds["train"]:
+    print(doc["file_key"], doc["document_type"], len(doc["full_text"] or ""))
     break
-
-# Load entities into memory
-entities = load_dataset("kabasshouse/epstein-data", "entities", split="train")
-print(f"{len(entities):,} entities loaded")
-
-# Filter to a specific dataset
-ds10 = load_dataset("kabasshouse/epstein-data", "documents", split="train")
-ds10 = ds10.filter(lambda x: x["dataset"] == "DataSet10")
 ```
 
-### With DuckDB (no download)
+### DuckDB (direct Parquet queries)
 
 ```sql
--- Query directly from HuggingFace (auto-downloads Parquet)
-SELECT file_key, dataset, document_type, date
+-- Query directly from HuggingFace without downloading
+SELECT file_key, document_type, date, char_count
 FROM 'hf://datasets/kabasshouse/epstein-data/data/documents/*.parquet'
-WHERE dataset = 'DataSet10' AND document_type = 'Email'
-LIMIT 10;
-
--- Count entities by type
-SELECT entity_type, COUNT(*) as cnt
-FROM 'hf://datasets/kabasshouse/epstein-data/data/entities/*.parquet'
-GROUP BY entity_type
-ORDER BY cnt DESC;
+WHERE document_type = 'Email'
+  AND date LIKE '2015%'
+ORDER BY date
+LIMIT 20;
+
+-- Financial transactions
+SELECT transaction_date, amount, merchant_name, cardholder
+FROM 'hf://datasets/kabasshouse/epstein-data/data/financial_transactions/*.parquet'
+WHERE cardholder LIKE '%EPSTEIN%'
+  AND amount > 1000
+ORDER BY amount DESC
+LIMIT 20;
+
+-- Curated gold documents
+SELECT file_key, subject, tier, headline, key_quote
+FROM 'hf://datasets/kabasshouse/epstein-data/data/curated_docs/*.parquet'
+WHERE tier = 'NUCLEAR'
+ORDER BY subject, doc_date;
 ```
 
-### With pandas
+### Pandas
 
 ```python
 import pandas as pd
 
-# Read a specific shard
-df = pd.read_parquet("hf://datasets/kabasshouse/epstein-data/data/documents/documents-00000.parquet")
-print(df.shape)
-print(df.columns.tolist())
+docs = pd.read_parquet("hf://datasets/kabasshouse/epstein-data/data/documents/")
+print(f"{len(docs):,} documents")
+print(docs.groupby("dataset").size().sort_values(ascending=False))
 ```
 
-### Assemble a local SQLite database
+## Data Layers
 
-```bash
-pip install pyarrow numpy tqdm huggingface_hub
+### Core Content
 
-# Core tables only (documents + entities, ~1 GB download)
-python assemble_db.py --layers core --output epstein.db
+| Layer | Rows | Description |
+|-------|------|-------------|
+| `documents` | 1,424,673 | Full OCR text, document type, date, photo flag |
+| `entities` | 10,629,198 | Named entities (person, org, location, date, etc.) |
+| `chunks` | 2,193,090 | ~800-token text chunks for RAG |
+| `embeddings_chunk` | 2,111,356 | 768-dim Gemini embeddings per chunk |
 
-# With text chunks (~2.5 GB download)
-python assemble_db.py --layers text --output epstein.db
+### Knowledge & Analysis
 
-# Full database with embeddings (~7.5 GB download)
-python assemble_db.py --layers full --output epstein.db
+| Layer | Rows | Description |
+|-------|------|-------------|
+| `persons` | 1,614 | Curated person registry (name, aliases, category) |
+| `kg_entities` | 467 | Knowledge graph nodes |
+| `kg_relationships` | 2,198 | Knowledge graph edges (traveled_with, associated_with, etc.) |
+| `recovered_redactions` | 37,870 | ML-recovered text from redacted pages |
+| `curated_docs` | 5,766 | Expert-annotated gold documents (5 subjects, tiered) |
 
-# Everything including provenance (~8 GB download)
-python assemble_db.py --layers all --output epstein.db
+### Structured Records (NEW in v2)
 
-# From a local Parquet export
-python assemble_db.py --local ./hf_export/ --layers core --output epstein.db
-```
+| Layer | Rows | Description |
+|-------|------|-------------|
+| `financial_transactions` | 49,770 | Credit card & bank transactions |
+| `derived_events` | 3,038 | Reconstructed calendar/travel/financial events |
+| `event_participants` | 5,751 | People linked to derived events |
+| `event_sources` | 21,910 | Source documents for derived events |
+| `investigative_records` | 143 | Law enforcement reports & evidence logs |
+| `communication_records` | 128 | Phone call & CDR records |
 
-## Source documents
+### Provenance
 
-The documents come from 12 DOJ FOIA dataset releases plus two community-sourced collections:
+| Layer | Rows | Description |
+|-------|------|-------------|
+| `provenance/files` | 1,387,775 | Per-file processing metadata + SHA-256 checksums |
+| `provenance/audit_log` | 3,711,609 | Append-only forensic audit trail |
+| `provenance/runs` | 123 | Pipeline execution records |
+
+## Datasets
 
 | Dataset | Files | Source |
 |---------|-------|--------|
-| DataSet 1 | 3,158 | DOJ FOIA |
-| DataSet 2 | 574 | DOJ FOIA |
-| DataSet 3 | 67 | DOJ FOIA |
-| DataSet 4 | 152 | DOJ FOIA |
-| DataSet 5 | 120 | DOJ FOIA |
-| DataSet 6 | 13 | DOJ FOIA |
-| DataSet 7 | 17 | DOJ FOIA |
-| DataSet 8 | 10,595 | DOJ FOIA |
-| DataSet 9 | 531,279 | DOJ FOIA |
-| DataSet 10 | 503,154 | DOJ FOIA |
-| DataSet 11 | 331,655 | DOJ FOIA |
-| DataSet 12 | 152 | DOJ FOIA |
-| FBIVault | 22 | FBI Vault FOIA |
+| DataSet 1 | 3,158 | DOJ EFTA release |
+| DataSet 2 | 574 | DOJ EFTA release |
+| DataSet 3 | 67 | DOJ EFTA release |
+| DataSet 4 | 152 | DOJ EFTA release |
+| DataSet 5 | 120 | DOJ EFTA release |
+| DataSet 6 | 13 | DOJ EFTA release |
+| DataSet 7 | 17 | DOJ EFTA release |
+| DataSet 8 | 10,595 | DOJ EFTA release |
+| DataSet 9 | 531,279 | DOJ EFTA release (community Tesseract OCR) |
+| DataSet 10 | 503,154 | DOJ EFTA release |
+| DataSet 11 | 331,655 | DOJ EFTA release |
+| DataSet 12 | 152 | DOJ EFTA release |
+| FBIVault | 22 | FBI Vault FOIA release |
 | HouseOversightEstate | 4,892 | House Oversight Committee |
 
-**Total: 1,385,850 successful** + 472 unrecoverable failures (documented in `release/epstein_problems.json`).
-
-## OCR provenance
-
-Two OCR sources were used:
+**468 unrecoverable failures** (corrupt/empty source PDFs). Full failure catalog in `release/epstein_problems.json`.
 
-- **Gemini 2.5 Flash Lite** (848,228 files): Primary OCR engine. These documents have `ocr_source` = NULL.
-- **Tesseract (community)** (537,622 files): Gap-fill from community repositories. These have `ocr_source` = `"tesseract-community"`.
+## OCR Sources
 
-See [PROVENANCE.md](PROVENANCE.md) for per-table source documentation.
+- **Gemini 2.5 Flash Lite**: 856,028 files — structured JSON output with entities, document classification, and metadata
+- **Tesseract (community)**: 531,279 files — raw text only (DataSet 9, community gap-fill imports)
+- **Upgraded**: 1,377 files originally processed with Tesseract, now re-processed with Gemini
 
-## Schema
+Distinguish OCR source via the `ocr_source` column: `NULL` = Gemini, `'tesseract-community'` = community Tesseract.
 
-Every document has a unique `file_key` (e.g., `EFTA00000001`) that serves as the primary identifier across all tables. The Parquet files use `file_key` everywhere -- no opaque integer IDs.
+## Curated Documents
 
-Key fields on `documents`:
-- `file_key` -- unique identifier (EFTA number)
-- `dataset` -- source dataset (e.g., "DataSet10")
-- `full_text` -- complete OCR text
-- `document_type` -- classified type (Email, Form, Letter, Photo, etc.)
-- `date` -- extracted date if available
-- `is_photo` -- whether the document is a photograph
-- `ocr_source` -- NULL for Gemini, "tesseract-community" for community OCR
+The `curated_docs` layer contains 5,766 expert-annotated gold documents across 5 investigation subjects:
 
-See `schema.sql` for the full SQLite schema used by `assemble_db.py`.
+| Subject | Gold Docs | Tiers |
+|---------|-----------|-------|
+| Hoffman | 1,526 | NUCLEAR / CRITICAL / HIGH / MEDIUM / SUPPORTING |
+| Gates | 2,069 | NUCLEAR / CRITICAL / HIGH / MEDIUM / SUPPORTING |
+| Summers | 739 | NUCLEAR / CRITICAL / HIGH / MEDIUM / SUPPORTING |
+| Clinton | 765 | NUCLEAR / CRITICAL / HIGH / MEDIUM / SUPPORTING |
+| Black | 667 | NUCLEAR / CRITICAL / HIGH / MEDIUM / SUPPORTING |
 
-## Known issues
+Each entry includes: tier, category, date, sender/recipient, headline, key quote, and investigative detail.
 
-- 472 source PDFs could not be processed (corrupt, empty, or unavailable). These are cataloged in `release/epstein_problems.json` with DOJ download URLs.
-- DataSet 9 (531K files) was entirely community-processed with Tesseract OCR, which has lower quality than Gemini.
-- Some documents are heavily redacted. `recovered_redactions` contains ML-recovered text from 39,588 redacted pages.
-- Embedding coverage is ~96% for chunks (1,249 malformed embeddings excluded). Summary embeddings were removed as redundant -- 92% of documents have a single chunk, making summary and chunk embeddings identical.
+## Financial Transactions
 
-## Release artifacts
+49,770 clean records extracted from credit card statements and bank records using DeepSeek. Includes:
 
-Small reference files are included directly in this repo under `release/`:
+- Transaction date, amount, currency, merchant
+- Cardholder name (Epstein, Maxwell, Shuliak, etc.)
+- Flight data (origin, destination, carrier, passenger) for airline purchases
+- Merchant category classification
 
-| File | Size | Description |
-|------|------|-------------|
-| `epstein_problems.json` | 280 KB | 472 processing failures with DOJ URLs |
-| `efta_dataset_mapping.json` | 4 KB | EFTA file key to DOJ URL mapping |
-| `persons_registry.json` | 436 KB | 1,614 curated person records |
-| `knowledge_graph_entities.json` | 172 KB | 467 KG entities |
-| `knowledge_graph_relationships.json` | 932 KB | 4,190 KG relationships |
-| `extracted_entities_filtered.json` | 1.9 MB | Filtered entity export |
-| `redacted_text_recovered.json.gz` | 2.5 MB | 39,588 recovered redacted pages |
-| `document_summary.csv.gz` | 1.8 MB | Document metadata summary |
-| `image_catalog.csv.gz` | 15 MB | Photo/image catalog |
+31% of raw extractions were quarantined for quality issues and excluded from this release.
 
 ## License
 
-This dataset is released under [CC-BY-4.0](LICENSE). The underlying documents are U.S. government records released under FOIA.
+CC-BY-4.0. Source documents are U.S. government public records.
 
 ## Citation
 
 ```bibtex
-@dataset{epstein_data_2026,
-  title={Epstein Document Archive},
-  author={Kevin Bass},
+@dataset{epstein_archive_2026,
+  title={Epstein DOJ Document Archive},
+  author={kabasshouse},
   year={2026},
   url={https://huggingface.co/datasets/kabasshouse/epstein-data},
-  note={1.39M OCR'd DOJ documents with entities, embeddings, and knowledge graph}
+  version={2.0}
 }
 ```

data/chunks/chunks-00000-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a981d273fec089a29a4974e3a2af050202679a1866d7d697838ff4b0b3a8a652
+size 42764173

data/chunks/chunks-00001-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d297fd634712e97d74b2cb6bf044896f5c4c663c8453c06aa33a580bfd86607
+size 33380833

data/chunks/chunks-00002-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e40c6aca5a9aea79a1def8a6df7bf46c4db51b5c950ba8e631c8f845ee88e87
+size 29118656

data/chunks/chunks-00003-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be7f48ed65262b15c339ebca9aa17643f7688cca33ec117ca2146645bb4ad904
+size 42233793

data/chunks/chunks-00004-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2df4e8a30985bfff388c710d5f1fd0f083e03dca874993c180bb47f6c13082e
+size 87261378

data/chunks/chunks-00005-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc38b2316707eb4a6b1352c2640345c876fbc3895aacee242caafcdc825ab543
+size 45014063

data/chunks/chunks-00006-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ca9ce7e668a7d0489caf7906828c35596dea52326e6820f5130b8193fb88d8b
+size 70822072

data/chunks/chunks-00007-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d93698d86e8b5651928f89230afbf2fc6f62c465cd6da99ecd0fb3f98b2ea65
+size 49084476

data/chunks/chunks-00008-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbad34373a91f59c77f05c055fe645ffe04e7741982c2773b6b780e1de7fb5a3
+size 89484554

data/chunks/chunks-00009-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06e5502ee0d7a23815b629163ab15c1c9da4ce91c46557e747a047a5c828c19c
+size 73717067

data/chunks/chunks-00010-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f649bc2bf8558e7cb069f157218e3e909471222dceebea3914038b597d7b89bd
+size 88033223

data/communication_records/communication_records-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a141c3718414c26778bd0042c3c91f4540b2e61ddc2a0b82ebabd43f1a0f15e9
+size 11056

data/curated_docs/curated_docs-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92dfd415fafda8467d6d1ca15834c4c451aec61f6e9e5453033ac06adaf110a1
+size 1629169

data/derived_events/derived_events-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e67707faf0c02e3ad8606a89ace4959fecfa060e27f0244e885ee537a2e44ca
+size 291804

data/documents/documents-00000-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7cbb617509ee7a80fba0131258a10c28540e88af6cc89f57d85cd2b4e5c582d
+size 32160053

data/documents/documents-00001-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de6b953b10b7c127441fe01f98c1932dca33c0a0cc1ae89ad43088a8a4a5c7a6
+size 17896030

data/documents/documents-00002-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20dcade7851269ecabfef1706bbd3c0d3b8d9b5d7c76de6127a384f096b4f1dc
+size 19282498

data/documents/documents-00003-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40283258fd792d74b30cfac709be1580252fb0d3b755f794c3abe5e5018ef5bc
+size 19455797

data/documents/documents-00004-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa27211d9f5730e8585fee6ec413aecf27f5fe88d343294c15568df991a76a9b
+size 17039369

data/documents/documents-00005-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:335f064da47bffeacaded9fe1021b26d2481e88fb71915b01285ec883b6876a8
+size 17440809

data/documents/documents-00006-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e25f3d8f31833bf103173dad70e52ca54d7945d397f5cecd23588c4c2c6a6db
+size 17857487

data/documents/documents-00007-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c950c8ab37e16d698d1fe5480977d0447831dad78957a15942ce75787ce3a10b
+size 109974347

data/documents/documents-00008-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:882668db59453182f3dc50560da56f1acbfcfc6b2a0e5ba09cd95faa1127c363
+size 25698673

data/documents/documents-00009-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:593199750fd761c17fe590d9c4e7bea71e23bf2e8d9c43f01a0b49cee4b57d23
+size 57062658

data/documents/documents-00010-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bcdd7b92050f8eac4e2f758b9489064df9265b6f6523f621965ae61461860f9
+size 45063376

data/documents/documents-00011-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a169499d8d62d68e80740b8e16418ca3d086c6d28314ab33e77af75b9a1f7cb9
+size 30585491

data/documents/documents-00012-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6e9112404eea9d15a08401dba574250bb51e396e63f1ae759eb66c05f66b730
+size 99331623

data/documents/documents-00013-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6cf89ef12831f9240d1b4747d8014cd4961edc0a963eced9ba7c20861c71371
+size 91335664

data/documents/documents-00014-of-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eade2f5a3e0b115bc1dc72a8c45e84e8dd2c03858b949bcf4728c479bcf3d5be
+size 34729689

data/embeddings_chunk/embeddings_chunk-00000-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86b02fb452416c4030ad2ddb88432fdc272492f89f645da7683abdf26960d4c9
+size 288483749

data/embeddings_chunk/embeddings_chunk-00001-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:148d386cd71cd4851ebbc4a635b059c33f9807b20dd362663bfbb1b801f618c2
+size 285222350

data/embeddings_chunk/embeddings_chunk-00002-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c2ef04c2ec3a7a06bc723d20e2e359a517a46aa18f6f17827b4f669c6c3a7c4
+size 289276438

data/embeddings_chunk/embeddings_chunk-00003-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cb3c5462ba3550dd12356b621fc69c31d3057a06acdeeae675cdfb4393a8cd0
+size 289338036

data/embeddings_chunk/embeddings_chunk-00004-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be1803d2be01c0d42bd5cca026226eb8b6b0c8d60bcc0eb08cbab620a7714393
+size 289401040

data/embeddings_chunk/embeddings_chunk-00005-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8367b73b4f07ad199e51f901114e18285d332ebe8b7f177a6769aa450b80ed6
+size 289392645

data/embeddings_chunk/embeddings_chunk-00006-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8708706541bad5ec2c4e2aa0f4f2c2ac0da13bffbd35ad65f2c2c1060a8c7c00
+size 289367235

data/embeddings_chunk/embeddings_chunk-00007-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce160de0b2308be6c77c8564afdd20df9031e3bbe5021c0f4567780816d0e6cd
+size 288539090

data/embeddings_chunk/embeddings_chunk-00008-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d038982c4f8a25368d4e6be209fd55bb6a18c1a751bf93f17cb223fe09627d4
+size 287019691

data/embeddings_chunk/embeddings_chunk-00009-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07ddf238792ec25ef0de6c02652036f2faf85f60ad7d4e59388460cb8f624061
+size 282093097

data/embeddings_chunk/embeddings_chunk-00010-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33daa455fce4d8f2003668f737452361775e8381511fe187b072eb751c396f66
+size 288806736

data/embeddings_chunk/embeddings_chunk-00011-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cca3204fb9e03c2ff5a372de2ffac3e4d5c1e0bd3db3b0ff2d1dc9e48ea60b90
+size 286994949

data/embeddings_chunk/embeddings_chunk-00012-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2561d7f17910d13d91a733b9eac39b9b8ea00f85a9bdee2c66c19cf320832d43
+size 288108603

data/embeddings_chunk/embeddings_chunk-00013-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4da100983a1f559f614686080c1e6cdf6b83fa165ebeffc9dc5954838e800be6
+size 289185371

data/embeddings_chunk/embeddings_chunk-00014-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51f74efd7e6751b179d7b4122eb5869a74029f4be2480eca38e8465e1ad096bb
+size 288657815

data/embeddings_chunk/embeddings_chunk-00015-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:437de34dd97ec77377c2200214343bc26b0d7802611805637439b82ae6a89188
+size 289344303

data/embeddings_chunk/embeddings_chunk-00016-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6551e7141e768bd62da4eb85f04be0b7136ba4ed16ecee55f73218acd503ac9f
+size 287325025

data/embeddings_chunk/embeddings_chunk-00017-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91e0bae0551f51fa8353c4c44db4e05ad301112e7f6eb31260452937bfedd6a5
+size 286707491

data/embeddings_chunk/embeddings_chunk-00018-of-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b6917c6d46af116184d9c95c1343467b0675a7df0736c921dacb8619ba250b7
+size 288816406