Skip to content

research-ontology.md

Latest commit

 

History

History
113 lines (79 loc) · 4.4 KB

research-ontology.md

File metadata and controls

113 lines (79 loc) · 4.4 KB
Raw

Research Ontology

Claims, evidence, and provenance as first-class graph citizens.

Why research needs different schema

Standard knowledge graphs track facts. Research graphs track claims — assertions that may be supported, refuted, uncertain, or contested. The difference matters:

  • A fact: Person.name = "Alice"
  • A claim: "Alice's methodology increases throughput by 40% — based on one study, confidence 0.6"

A claim without source and confidence is just an assertion you can't evaluate. Every claim in a research graph must carry its provenance.

Start with ResearchNotes

start_here({ "template": "ResearchNotes" })

Seeds five classes:

  • Claim — the core unit. Every insight, finding, hypothesis, or assertion.
  • Document — papers, articles, transcripts, notes that contain claims
  • Concept — ideas, theories, domains, named abstractions
  • Person — authors, researchers, sources
  • Asset — datasets, codebases, outputs that claims are about

Claim is the core unit

Everything else supports claims. A Document exists because it contains claims. A Person exists because they authored or contested a claim. A Concept exists because claims are about it.

Required properties on every Claim:

add_property(owner="Claim", name="statement", datatype="string", required=true)
add_property(owner="Claim", name="source", datatype="string", required=true)
add_property(owner="Claim", name="confidence", datatype="float", required=true)

confidence is a float from 0.0 (speculation) to 1.0 (verified). Use it consistently — 0.9+ means well-supported, 0.5–0.8 means plausible, below 0.5 means worth noting but not acting on.

Key relations for research

Relation Domain → Range Use for
CITES Document → Document Citation chain
SUPPORTS Claim/Document → Claim Evidence for a claim
CONTRADICTS Claim/Document → Claim Evidence against
DERIVED_FROM Claim/Asset → Claim/Asset Lineage — derived findings
AUTHORED Person → Document/Claim Who wrote/made it
TAGGED_WITH Claim/Document → Concept Conceptual classification

Example: a contested claim

// Create the claim
create_entity("Claim", {
  "statement": "LLM agents with structured memory outperform baseline on long-horizon tasks",
  "source": "Smith et al. 2024, arXiv:2401.XXXXX",
  "confidence": 0.72
})  // → claim_id: "100"

// Create supporting evidence
create_entity("Document", {
  "title": "Smith et al. 2024",
  "url": "https://arxiv.org/abs/2401.XXXXX"
})  // → doc_id: "101"

create_relationship(doc_id="101", claim_id="100", "SUPPORTS")

// Create a contradicting claim
create_entity("Claim", {
  "statement": "Performance gains from structured memory are within noise margins",
  "source": "Jones 2024, replication study",
  "confidence": 0.55
})  // → counter_id: "102"

create_relationship(counter_id="102", claim_id="100", "CONTRADICTS")

Now you can query: "which claims contradict my main hypothesis?" — and each result carries the source and confidence of the contradiction.

Provenance chains

DERIVED_FROM tracks how findings build on each other:

Claim: "Agents need structured memory" (confidence: 0.9)
  ← DERIVED_FROM ← Claim: "Baseline agents forget context" (confidence: 0.95)
  ← DERIVED_FROM ← Claim: "Context loss correlates with task failure" (confidence: 0.85)

When a foundational claim's confidence drops (new replication fails), you can traverse the graph to find all derived claims that need re-evaluation.

Design principles for research graphs

  • Every Claim needs source and confidence. No exceptions. An unsourced claim is useless for research.
  • SUPPORTS and CONTRADICTS are the highest-value relations. Invest time in wiring them. They're what makes the graph useful for synthesis, not just storage.
  • Don't conflate Claim and Document. A Document contains Claims; it is not itself a Claim. A paper that argues X produces a Claim node for X, linked to the Document node via AUTHORED + SUPPORTS.
  • Confidence degrades through DERIVED_FROM chains. A claim derived from a 0.6-confidence finding should itself be no higher than 0.6, probably lower.
  • Use TAGGED_WITH for thematic navigation. Tag claims with Concept nodes for the research domains they belong to. This makes "find all claims about X" fast.