Connect claims, evidence and context so people and agents can use knowledge with its sources in view.
A set of documents can become a connected layer of claims, evidence, conditions and relationships that people and agents can navigate.
Ketone-body mitochondrial oxidation in mushroom body neurons is required for memory formation during starvation.
This research example follows claims in one source document; the same links can connect claims and evidence across a document set.
Statements, claims, methods, controls and outcomes remain traceable to the document rather than becoming isolated passages.
The extracted claims join a larger network of findings, controls and supporting context.
Inspect claims, evidence, methods and measurements as typed objects.
Follow support and qualification links back to the source passage.
Compare objects, check prerequisites and return versioned proposals linked to evidence.
Evidence set → comparison → versioned result. Access through API, MCP or an application.
Share of benchmark questions for which retrieval found source evidence marked necessary to answer. This measures retrieval coverage, not final answer accuracy.
Useful retrieved-context precision is the share of returned context judged relevant to the question.
The curves show reference information recovered as the maximum context budget grows.
Both approaches receive the same maximum text allowance at each step. The results measure retrieval into context.
Sample standard deviation across model-level average scores, in percentage points.
Lower spread means more stable scores between models. This summarizes model-level averages, not question-level uncertainty.
Separate scientific-content retrieval evaluations. The equal-budget comparison uses a distinct question set from the required-evidence and useful-context measures.