Skip to content

SOURCE_AND_FACT_MODEL.md

Latest commit

 

History

History
143 lines (113 loc) · 5.14 KB

SOURCE_AND_FACT_MODEL.md

File metadata and controls

143 lines (113 loc) · 5.14 KB
Raw

Source and Fact Model

FGIP separates citations, extracted facts, and graph edges into distinct layers. Each layer has a different purpose and a different consumer.

The three layers

sources.jsonl        → Who said it, when, where (human-readable citation)
facts.jsonl          → What was said, structured (machine-readable claim)
proposed_edges.jsonl → How it connects to the graph (candidate relationship)

Layer 1: Sources (data/sources/sources.jsonl)

A source is a citable document. It has a publisher, a URL, a publication date, and a tier.

{
  "source_id": "congress_ndaa_fy2025_hr8070",
  "tier": 0,
  "source_type": "legislation",
  "publisher": "Congress.gov",
  "title": "National Defense Authorization Act for Fiscal Year 2025",
  "url": "https://congress.gov/bill/118th-congress/house-bill/8070",
  "published_at": "2024-12-23",
  "accessed_at": "2026-05-09"
}

Purpose: Audit trail. Anyone can click the URL and verify the document exists.

Tier assignment:

  • Tier 0: Government records (Congress.gov, EDGAR, USASpending, Federal Register, NRC, FERC, state PUCs)
  • Tier 1: Professional sources (Reuters, company press releases, earnings reports, industry associations)
  • Tier 2: Commentary (news articles, podcasts, social media)
  • Tier 3: Hypothesis (AI analysis, user observations, cross-reference inferences)

See EVIDENCE_TIERS.md for full tier definitions.

Layer 2: Extracted facts (data/extracted/facts.jsonl)

A fact is a structured claim extracted from a source. One source can produce multiple facts.

{
  "fact_id": "fact_ndaa_fy2025_enacted",
  "source_id": "congress_ndaa_fy2025_hr8070",
  "fact_type": "legislation",
  "subject": "Congress",
  "predicate": "ENACTED",
  "object": "NDAA FY2025",
  "value_usd": 895200000000,
  "date": "2024-12-23",
  "confidence": 1.0,
  "note": "$895.2B defense authorization. 5.2% military pay raise."
}

Purpose: Machine-readable claims. The graph doesn't read articles — it reads facts.

Fields:

Field Type Required Description
fact_id string yes Unique identifier
source_id string yes Links back to source citation
fact_type string yes Category (legislation, procurement, earnings, supply_demand, etc.)
subject string yes Who/what is acting
predicate string yes The action or relationship
object string yes Who/what is acted upon
quantity number no Numeric value if applicable
unit string no Unit of quantity
value_usd number no Dollar value if applicable
date string yes When the fact applies
confidence float yes 0.0 to 1.0
note string no Context for human readers

Layer 3: Proposed edges (data/edges/proposed_edges_examples.jsonl)

A proposed edge is a candidate relationship for the graph. It references a fact and maps it to graph node IDs.

{
  "edge_id": "edge_ndaa_fy2025_lmt",
  "fact_id": "fact_ndaa_fy2025_enacted",
  "from_node": "ndaa-fy2025",
  "to_node": "lockheed-martin",
  "relationship": "AUTHORIZES_FUNDING",
  "confidence": 0.95,
  "agent_name": "congress",
  "tier": 0,
  "note": "$895.2B authorization. F-35, HIMARS, ATACMS, hypersonic programs."
}

Purpose: Staging area for graph insertion. Edges sit here until reviewed and promoted.

Promotion rules:

  • Tier 0 agent edges self-certify (the data IS the evidence)
  • Tier 1 agent edges require artifact evidence
  • Tier 2+ edges require manual review or triangulation

How the layers connect

Source (sources.jsonl)
  ↓ source_id
Fact (facts.jsonl)
  ↓ fact_id
Proposed Edge (proposed_edges_examples.jsonl)
  ↓ promotion
Graph Edge (fgip.db edges table)
  ↓ conviction engine
Thesis Receipt (fgip_receipts/*.json)

One source can produce many facts. One fact can produce many edges. The conviction engine queries edges (both promoted and proposed) to score theses.

Example: Defense funding chain

Source: Congress.gov, H.R.8070, NDAA FY2025

Facts extracted:

  • Congress ENACTED NDAA FY2025 ($895.2B)
  • 5.2% military pay raise authorized

Edges proposed:

  • ndaa-fy2025 → lockheed-martin (AUTHORIZES_FUNDING, 0.95)
  • ndaa-fy2025 → raytheon (AUTHORIZES_FUNDING, 0.95)
  • ndaa-fy2025 → northrop-grumman (AUTHORIZES_FUNDING, 0.95)
  • ndaa-fy2025 → general-dynamics (AUTHORIZES_FUNDING, 0.95)
  • ndaa-fy2025 → huntington-ingalls (AUTHORIZES_FUNDING, 0.95)
  • ndaa-fy2025 → bwxt (AUTHORIZES_FUNDING, 0.95)

Thesis receipt: thesis-defense-primes — 148 confirming signals, 115 Tier 0, conviction level 5.

Design principles

  1. Citations are for humans. A source entry lets anyone verify the original document.
  2. Extracted facts are for the graph. Structured claims that agents can query.
  3. Proposed edges are for staging. Claims wait for review before becoming graph edges.
  4. Receipts are for trust. A receipt records what was believed, why, and what would falsify it.

Each layer has a different audience and a different standard of proof. Mixing them produces documents that are neither verifiable nor queryable.