Research tools forget across sessions, and they never notice when two sources disagree. Crosscheck is a small copilot on top of cogneethat does both: persistent memory of everything you feed it, and a hero feature that flags when sources contradict each other — e.g. "FooDB sustained 50,000 req/s" (2021) vs "only 10,000 req/s" (2024).
The obvious design — and why it wasn't enough
The first instinct is to make contradiction detection a pure graph query: build cognee's knowledge graph, then look for the same entity+attribute with different values across sources. Elegant, but it breaks on a real local stack (llama3.1:8b via Ollama). Two failures show up:
The knowledge graph flattens quantities. The extractor turns "50,000 requests per second" into a generic requests per second node and drops the number. The conflicting values never make it into the graph.
Entities get merged across sources. Every mention of FooDB collapses to one node, so the two throughput claims dedup into a single edge — nothing left to compare.
