data-archivist

Data Archivist Skill

Overview

The Data Archivist transforms raw OSINT (PDFs, CSVs, scraped web records, court filings, FOIA dumps) into queryable, reproducible databases that serve as the foundation for investigative analysis.

This is NOT data analysis (that's downstream). This IS:

• Ingestion standardization: Converting diverse formats (PDF tables, hand-typed CSVs, HTML scrapes, XML feeds) into uniform schema
• Schema design: Star schema patterns for investigative queries (facts, dimensions, timelines)
• Data provenance tracking: Chain of custody for evidence (where did it come from? when? how was it extracted? by whom?)
• Reproducible queries: Saved SQL patterns, not ad-hoc summaries
• Resource optimization: Free tier awareness (BigQuery free tier limits, PostgreSQL disk constraints, SQLite for local-first)

Binding protocol: This desk operates under skills/investigative-journalist/evidence-preservation-protocol.md. Originals in the write-once evidence/ directory are never modified — all transformation outputs (OCR text, cleaned CSVs, database loads) are derived copies in working/ that cite their source DOC- ID and hash. Every ingest, transform, and export is an entry in the master file's append-only collection_log[]. Nothing is deleted while legal_hold is true (the default).

---

Data Ingestion Standardization

Supported Input Formats

Format	Typical Source	Challenges	Standardization Approach
PDF (tabular)	Government reports, court filings, financial statements	Text extraction loses formatting; columns mis-aligned	OCR + manual table mapping; extract to CSV; validate row counts
PDF (unstructured)	News articles, regulations, correspondence	OCR errors; no structured fields	Extract text; keyword search for extraction; manual review
CSV	FOIA dumps, export from legacy systems, aggregated data	Missing headers; inconsistent delimiters; encoding issues (UTF-8 vs. Latin-1)	Validate headers; detect delimiter; convert encoding; document source URL
XLS/XLSX	Spreadsheets, financial records, surveys	Multiple sheets; merged cells; formulas (not values)	Extract all sheets separately; flatten merged cells; export formulas as values
JSON	API exports, web scrapes (JavaScript-heavy sites), config files	Nested structures; missing fields; null vs. empty string inconsistency	Flatten nested JSON; standardize nulls; document schema version
HTML Tables	Scraped web pages, cached legislative records	Cells span multiple columns/rows; colspan/rowspan confusion	Parse colspan/rowspan; extract as CSV; validate row integrity
XML/RSS	News feeds, structured government data, legal databases	Namespace confusion; CDATA sections; encoding issues	Strip namespaces; unescape CDATA; standardize date formats
SQL Dumps	Database exports from legacy systems, web platform backups	Encoding issues; proprietary SQL syntax; no schema documentation	Parse generic SQL (CREATE TABLE, INSERT); document assumed data types

Standardization Workflow

Phase 1: Source Documentation (required before ingestion)

For every data source, create a source-manifest.yaml:

source:
  name: "Property Tax Assessment 2024"
  url: "https://assessor.county.gov/export/property-assessment-2024.csv"
  date_accessed: "2025-03-15"
  date_published: "2025-02-01"

format:
  type: "CSV"
  encoding: "UTF-8"
  delimiter: ","
  headers_row: 1

extraction:
  method: "direct_download"
  by: "[email protected]"
  date: "2025-03-15"

fields:
  - name: "parcel_id"
    description: "Unique parcel identifier (APN)"
    type: "string"
    sample_values: ["123-456-789", "987-654-321"]
  - name: "owner_name"
    description: "Property owner name (may be LLC)"
    type: "string"
  - name: "assessed_value"
    description: "2024 assessed value in USD"
    type: "decimal"
  - name: "last_sale_date"
    description: "Date of last recorded sale (YYYY-MM-DD)"
    type: "date"
    null_handling: "Unknown sales marked as 1900-01-01"

row_count: 45829
sample_rows:
  - parcel_id: "123-456-789"
    owner_name: "John Smith"
    assessed_value: "250000"
    last_sale_date: "2020-06-15"

limitations:
  - "Data may be 1–3 months stale (assessed annually)"
  - "Some LLCs have nominees; true owner unknown"
  - "Exemptions (non-profit, government) may not be listed"

chain_of_custody:
  - event: "Downloaded from county website"
    date: "2025-03-15"
    by: "[email protected]"
    tool: "wget"
    hash_md5: "abc123def456..."
  - event: "Uploaded to newsroom database"
    date: "2025-03-15"
    by: "[email protected]"
    table: "property_tax.assessment_2024"

---

Platform-Agnostic Schema Design

The Investigative Star Schema

Most investigative databases follow a star schema pattern:

                    ┌─────────────┐
                    │   TIMELINE  │
                    │  (dim_date) │
                    └─────────────┘
                           ▲
                           │
     ┌─────────────────────┼─────────────────────┐
     │                     │                     │
┌─────────┐          ┌──────────┐          ┌──────────┐
│ PEOPLE  │          │ FACTS    │          │ ENTITIES │
│(dim_    │          │(fact_    │          │(dim_     │
│people)  │──────────│events)   │──────────│entities) │
└─────────┘          └──────────┘          └──────────┘
     │                     │                     │
     └─────────────────────┼─────────────────────┘
                           │
                    ┌─────────────┐
                    │  DOCUMENTS  │
                    │  (dim_docs) │
                    └─────────────┘

Table Definitions

Fact Table: fact_events (central)

CREATE TABLE fact_events (
  event_id          BIGINT PRIMARY KEY,
  person_id         BIGINT REFERENCES dim_people(person_id),
  entity_id         BIGINT REFERENCES dim_entities(entity_id),
  document_id       BIGINT REFERENCES dim_documents(document_id),
  event_date        DATE,
  event_type        VARCHAR(50),  -- "meeting", "payment", "filing", "communication"
  amount_usd        DECIMAL(14,2),
  description       TEXT,
  source_url        VARCHAR(2048),
  source_page       INT,
  confidence_level  VARCHAR(20),  -- "primary_source", "secondary", "inferred"
  FOREIGN KEY (person_id) REFERENCES dim_people(person_id),
  FOREIGN KEY (entity_id) REFERENCES dim_entities(entity_id),
  FOREIGN KEY (document_id) REFERENCES dim_documents(document_id),
  FOREIGN KEY (event_date) REFERENCES dim_date(date_key)
);

Dimension Table: dim_people

CREATE TABLE dim_people (
  person_id       BIGINT PRIMARY KEY,
  full_name       VARCHAR(255) NOT NULL,
  last_name       VARCHAR(100),
  first_name      VARCHAR(100),
  aliases         TEXT,  -- JSON array of known aliases
  titles          TEXT,  -- JSON array: "Mayor", "CEO", etc.
  birth_date      DATE,
  dob_confidence  VARCHAR(20),  -- "confirmed", "estimated", "unknown"
  gender          VARCHAR(20),
  address         VARCHAR(255),
  phone           VARCHAR(20),
  email           VARCHAR(255),
  linkedin_url    VARCHAR(255),
  notes           TEXT,
  source_urls     TEXT,  -- JSON array of URLs where info confirmed
  created_date    TIMESTAMP,
  updated_date    TIMESTAMP
);

Dimension Table: dim_entities

CREATE TABLE dim_entities (
  entity_id           BIGINT PRIMARY KEY,
  entity_name         VARCHAR(255) NOT NULL,
  entity_type         VARCHAR(50),  -- "LLC", "Corporation", "Government Agency", "Non-profit"
  formed_date         DATE,
  dissolved_date      DATE,
  jurisdiction        VARCHAR(100),  -- "Delaware", "New York", etc.
  address             VARCHAR(255),
  registered_agent    VARCHAR(255),
  ucc_number          VARCHAR(50),
  tax_id              VARCHAR(50),
  notes               TEXT,
  source_urls         TEXT,  -- JSON array
  created_date        TIMESTAMP,
  updated_date        TIMESTAMP
);

Dimension Table: dim_documents

CREATE TABLE dim_documents (
  document_id     BIGINT PRIMARY KEY,
  title           VARCHAR(255),
  document_type   VARCHAR(50),  -- "PDF", "CSV", "Court Filing", "Article"
  source_url      VARCHAR(2048),
  date_accessed   DATE,
  date_published  DATE,
  hash_md5        VARCHAR(32),  -- For version control
  file_path       VARCHAR(500),  -- Local archive path
  page_count      INT,
  ocr_confidence  DECIMAL(3,2),  -- 0.0–1.0 for OCR documents
  notes           TEXT,
  created_date    TIMESTAMP,
  updated_date    TIMESTAMP
);

Dimension Table: dim_date

CREATE TABLE dim_date (
  date_key        DATE PRIMARY KEY,
  year            INT,
  quarter         INT,
  month           INT,
  day             INT,
  week            INT,
  day_of_week     VARCHAR(10),
  is_weekend      BOOLEAN,
  day_name        VARCHAR(10)
);

---

Reproducible Query Generation

Query Library (not ad-hoc analysis)

Every common question gets a saved, documented SQL pattern:

Query 1: "Show all payments from Person X to Entity Y"

-- query_id: q001
-- purpose: Trace financial relationships between individuals and entities
-- created: 2025-03-15
-- author: [email protected]

SELECT
  fe.event_date,
  fe.amount_usd,
  fe.description,
  doc.title AS source_document,
  CONCAT(
    'https://source-archive.newsroom.org/',
    doc.file_path
  ) AS primary_source
FROM fact_events fe
LEFT JOIN dim_documents doc ON fe.document_id = doc.document_id
WHERE
  fe.person_id = ?  -- parameter: person_id
  AND fe.entity_id = ?  -- parameter: entity_id
  AND fe.event_type = 'payment'
ORDER BY fe.event_date DESC;

Query 2: "Timeline of Person X's roles/titles"

-- query_id: q002
-- purpose: Establish professional/political timeline
-- created: 2025-03-15
-- author: [email protected]

SELECT
  dd.date_key AS event_date,
  fe.event_type,
  fe.description,
  COALESCE(dp.titles, 'N/A') AS titles_at_time,
  doc.title AS source,
  doc.source_url
FROM fact_events fe
LEFT JOIN dim_date dd ON fe.event_date = dd.date_key
LEFT JOIN dim_people dp ON fe.person_id = dp.person_id
LEFT JOIN dim_documents doc ON fe.document_id = doc.document_id
WHERE fe.person_id = ?
ORDER BY fe.event_date ASC;

Query 3: "Web of connections: Entity X connected to Persons A–Z"

-- query_id: q003
-- purpose: Network analysis; identify co-conspirators or related parties
-- created: 2025-03-15
-- author: [email protected]

WITH entity_connections AS (
  SELECT
    fe.entity_id,
    fe.person_id,
    COUNT(*) AS event_count,
    MIN(fe.event_date) AS first_event,
    MAX(fe.event_date) AS last_event,
    SUM(fe.amount_usd) AS total_usd
  FROM fact_events fe
  GROUP BY fe.entity_id, fe.person_id
)
SELECT
  de.entity_name,
  dp.full_name,
  ec.event_count,
  ec.first_event,
  ec.last_event,
  ec.total_usd
FROM entity_connections ec
LEFT JOIN dim_entities de ON ec.entity_id = de.entity_id
LEFT JOIN dim_people dp ON ec.person_id = dp.person_id
WHERE ec.entity_id = ?
ORDER BY ec.total_usd DESC;

Query Library Maintenance

Store all queries in a version-controlled file (queries.sql or directory):

queries:
  q001:
    name: "Payments between person and entity"
    file: "queries/q001_payments_person_entity.sql"
    parameters: [person_id, entity_id]
    outputs: [event_date, amount_usd, description, source_document]

  q002:
    name: "Timeline of person's roles"
    file: "queries/q002_timeline_person_roles.sql"
    parameters: [person_id]
    outputs: [event_date, event_type, description, titles, source]

  q003:
    name: "Entity's connections to persons"
    file: "queries/q003_entity_connections.sql"
    parameters: [entity_id]
    outputs: [entity_name, full_name, event_count, first_event, last_event, total_usd]

---

Data Provenance & Chain of Custody

Evidence Tracking for Legal Defense

Every row in fact_events must be traceable to its source. This protects the newsroom if a claim is challenged:

Field	Purpose	Example
document_id	Foreign key to source document	847 (links to court filing PDF)
source_url	URL where data came from	https://courts.state.gov/filings/case-2024-001234
source_page	Page number (if PDF or scanned)	5
confidence_level	How confident in the extraction	"primary_source", "secondary", "inferred"

Chain of Custody Log (required for sensitive claims):

claim: "ABC Corp paid $50,000 to Mayor Jane Smith"
evidence:
  - step: 1
    date: "2025-03-10"
    action: "Downloaded campaign finance disclosure (Form XYZ) from state website"
    url: "https://elections.state.gov/disclosures/2024/mayor-smith-jane.pdf"
    by: "[email protected]"
    hash_md5: "abc123def456..."

  - step: 2
    date: "2025-03-10"
    action: "OCR extraction of page 3, table 'Contributions Over $1,000'"
    ocr_tool: "AWS Textract"
    ocr_confidence: 0.98
    result: "Extracted row: date=2024-06-15, contributor=ABC Corp (LLC), amount=$50,000"
    by: "[email protected]"

  - step: 3
    date: "2025-03-10"
    action: "Manual verification: cross-referenced against ABC Corp registered agent in Secretary of State database"
    verified: true
    by: "[email protected]"

  - step: 4
    date: "2025-03-15"
    action: "Added to database table: fact_events, event_id=9847"
    database: "newsroom_investigations"
    document_id: 512
    by: "[email protected]"

  - step: 5
    date: "2025-04-01"
    action: "Publication in article 'Mayor Smith's Undisclosed Corporate Donations'"
    article_url: "https://newsroom.org/articles/2025-04-01-mayor-smith-donations.html"
    by: "[email protected]"

If a legal challenge arises: every step from source to publication is documented and repeatable.

---

Platform Selection & Implementation

BigQuery (GCP)

Best for: Large datasets (>1M rows), frequent queries, collaborative analysis, free tier (1 TB/month query)

Pros:

• Handles investigative star schemas elegantly
• SQL pricing: pay for bytes scanned, not compute
• Built-in ML (anomaly detection for suspicious patterns)
• Shareable query results; audit log for FOIA

Cons:

• Needs GCP account; learning curve
• Costs accumulate if queries are inefficient
• Data stored in cloud (privacy considerations)

Setup:

# Install gcloud CLI
gcloud auth login
gcloud config set project my-newsroom-project

# Create dataset
bq mk newsroom_investigations

# Load CSV
bq load \
  newsroom_investigations.fact_events \
  data/fact_events.csv \
  --autodetect \
  --source_format=CSV

# Run saved query
bq query --use_legacy_sql=false < queries/q001_payments_person_entity.sql

PostgreSQL (Self-Hosted)

Best for: Medium datasets (10K–100M rows), local-first, full control, no cloud dependencies

Pros:

• Free and open-source
• Full-text search, JSON support, advanced indexing
• Data stays on your server (better privacy)
• Complex queries (CTEs, window functions) well-supported

Cons:

• Requires database admin; backup responsibilities
• Performance depends on hardware
• Requires VPN/SSH for remote access

Setup:

# Install PostgreSQL
sudo apt-get install postgresql postgresql-contrib

# Create database
createdb newsroom_investigations

# Connect
psql newsroom_investigations

# Run schema from schema.sql
\i schema.sql

# Load data
\COPY fact_events FROM 'data/fact_events.csv' CSV HEADER;

SQLite (Local, Zero Infrastructure)

Best for: Smaller datasets (<10M rows), single reporter, no server dependencies, prototyping

Pros:

• Single file; no setup required
• Perfect for local version control (commit .db to git)
• Queryable via CLI, Python, web interface
• Zero cost, zero admin

Cons:

• Doesn't scale to enterprise size
• No remote access; file-based (disk I/O bottleneck)
• Concurrent writes limited

Setup:

# Create database
sqlite3 newsroom_investigations.db < schema.sql

# Load data
sqlite3 newsroom_investigations.db
> .mode csv
> .import data/fact_events.csv fact_events

# Query
sqlite3 newsroom_investigations.db "SELECT * FROM fact_events LIMIT 10;"

---

Resource Constraint Awareness

Free Tier Limits (as of 2025)

Platform	Monthly Free Allowance	Hard Stop	Overflow Cost
BigQuery	1 TB query scan, 10 GB storage	Yes, at 1 TB	$6.25 per TB scanned
PostgreSQL (RDS)	None; pay-as-you-go starting at ~$10/month	No; scales with usage	Varies by instance size
SQLite	Unlimited; single file on disk	Limited by disk space	None (self-hosted)

Recommendation for startups/small newsrooms:

1. Prototype with SQLite (zero cost, instant setup)
2. Scale to PostgreSQL when dataset >5M rows or multiple reporters
3. Migrate to BigQuery only if you need shared access + no infrastructure burden

---

Common Mistakes & Red Flags

Ingestion Mistakes

Mistake	Why It Fails	Fix
No source manifest	Can't verify data origin; chain of custody broken	Create source-manifest.yaml for EVERY source before ingestion
CSV uploaded directly without validation	Missing headers; encoding issues; silent data loss	Run validation script: check row count, detect delimiter, test encoding
PDF table extracted as text (no structure)	Can't query; numbers stuck in paragraphs	Use OCR + table detection (AWS Textract, Google Vision); extract to structured CSV
Null values mixed: NULL, "N/A", empty string, "unknown"	Queries fail; counts off	Standardize: NULL for missing, "unknown" for intentional non-disclosure
Duplicate rows (same fact ingested twice)	Inflated counts; double-counted payments	Add UNIQUE constraints to database; run deduplication check post-ingestion
No hash verification	Can't detect if data was corrupted in transit	Generate MD5 hash of source file; store in source_manifest; re-verify post-upload

Schema Mistakes

Mistake	Why It Fails	Fix
No date_key in fact table	Can't join with dim_date; temporal analysis impossible	Always include event_date in fact table; join to dim_date for calendar calculations
Foreign keys not enforced	Orphaned rows; referential integrity breaks	Add FOREIGN KEY constraints; PostgreSQL/BigQuery will prevent invalid inserts
Mixing of data types (amount_usd as VARCHAR)	Sorting/summing fails; string "100" != numeric 100	Use DECIMAL/NUMERIC for amounts; DATE for dates; INT for IDs
No confidence_level field	Can't distinguish primary source from inferred	Add confidence_level to every fact; use during analysis ("primary_source" only)
Overly wide tables (100+ columns)	Slow queries; hard to maintain; unclear what fields mean	Break into normalized tables; use lookups (dim_people, dim_entities)

Query Mistakes

Mistake	Why It Fails	Fix
Ad-hoc queries; no saved pattern	Analysis not repeatable; other reporters can't replicate	Document every query pattern; save to version control; name logically (q001, q002...)
Hardcoded IDs ("WHERE person_id = 123")	Not portable; brittle; other reporters can't use	Use parameters ("WHERE person_id = ?") or variables
No LIMIT on exploratory queries	Scans entire table; costs skyrocket on BigQuery	Always test with LIMIT 100 first; understand row count before full scans
LEFT JOIN without clarifying missing data	Includes null rows; double-counted facts	Use INNER JOIN for core facts; LEFT JOIN only when absence is meaningful
No ORDER BY	Results appear random; hard to verify accuracy	Always ORDER BY primary key or date; makes results deterministic

Provenance Mistakes

Mistake	Why It Fails	Fix
No source_url in fact_events	If claim is challenged, can't show original source	Every row must link to document_id, which links to source_url
"OCR'ed from PDF" with no confidence score	Can't distinguish high-confidence vs. low-confidence extractions	Include ocr_confidence (0.0–1.0) and flag low-confidence rows for manual review
Deleting intermediate work (raw PDFs, extraction scripts)	Archive destroyed; chain of custody lost	Never delete raw data or extraction scripts; version control all SQL schemas
No date_accessed or date_published	Can't determine if data is stale or superseded	Every source_manifest must include date_accessed, date_published, and publication date

---

Checklist for Publication

Before publishing claims based on archived data:

• Source manifest complete for every data source (URL, date accessed, schema documented)
• Chain of custody log created (if high-stakes claim)
• Data validation passed: row count matches source, encoding correct, no silent truncation
• Foreign key constraints enforced; no orphaned rows
• Confidence levels assigned (primary_source, secondary, inferred); only primary_source used for definitive claims
• Queries saved and documented (not ad-hoc)
• Query results repeatable (others can run same query, get same results)
• Hash verification done (MD5 of source files stored and re-verified)
• No hardcoded IDs in production queries (use parameters)
• Null handling standardized (NULL vs. "unknown")
• All intermediate work archived (raw CSVs, extraction scripts, OCR output)

---

Example: Full Workflow

Scenario: Investigative reporter discovers that municipal contract awards are going disproportionately to one vendor. Wants to prove it with data.

Step 1: Source Collection

• Download municipal contracts CSV from city procurement website
• Scrape vendor bids from city council meeting minutes (HTML)
• Extract vendor payment records from annual audit (PDF table)

Step 2: Source Documentation Create source-manifest.yaml for each:

source:
  name: "Municipal Contracts 2020–2024"
  url: "https://procurement.city.gov/contracts/export"
  date_accessed: "2025-03-15"
  ...

Step 3: Ingestion

• Validate CSV: check headers, detect delimiter, verify encoding
• Scrape HTML → parse tables → export to CSV
• OCR PDF → extract table → validate against audit report

Step 4: Load into Database

• Create dim_entities (vendors), dim_people (decision-makers), fact_events (contracts)
• Load cleaned CSVs
• Run integrity checks: no orphaned rows, all amount_usd are numeric

Step 5: Query & Analyze

-- Which vendor got the most contracts?
SELECT
  de.entity_name,
  COUNT(*) AS contract_count,
  SUM(fe.amount_usd) AS total_awarded
FROM fact_events fe
LEFT JOIN dim_entities de ON fe.entity_id = de.entity_id
WHERE fe.event_type = 'contract_award'
  AND YEAR(fe.event_date) BETWEEN 2020 AND 2024
GROUP BY de.entity_name
ORDER BY total_awarded DESC;

Step 6: Document Findings

• Save query as q_contracts_by_vendor.sql
• Link results to source documents
• Create chain of custody log

Step 7: Publish

• Editor verifies using saved queries (reproducible)
• Archive all intermediate work
• Article published with links to data sources

---

Tools & Resources

Tool	Purpose	Cost	Learning Curve
BigQuery	Cloud data warehouse; scalable analysis	Free tier: 1 TB/month query	Medium
PostgreSQL	Self-hosted relational DB; full-featured	Free	Medium–High
SQLite	Lightweight local DB; no setup	Free	Low
AWS Textract	PDF/image OCR; table extraction	~$1.50 per page	Low–Medium
Google Sheets	Quick CSV import, pivot tables, sharing	Free; 5M cells limit	Very Low
DuckDB	In-process SQL; analyze CSVs directly	Free	Low
Apache Airflow	Automated data pipeline scheduling	Free (self-hosted)	High

---

Escalation

If during ingestion you discover conflicting data (same entity listed twice with different details):

1. Document both versions with sources
2. Flag for editorial review
3. Do NOT arbitrate; let editor decide which version to use
4. Update confidence_level accordingly ("confirmed" vs. "disputed")