veridian.

A corpus-grounded research cognition engine: retrieve literature on a topic, discover its interpretive terrain, and reflect an entered understanding against what the corpus actually supports.

Why this exists

Entering an unfamiliar research domain means facing hundreds of papers with no obvious map — no visible concept hierarchy, no sense of which interpretations are dominant versus emerging, and no easy way to tell whether a claim you're forming is actually supported by the literature. Most tools either summarize, compressing away the interpretive landscape, or generate, producing claims ungrounded in any specific corpus. veridian does neither: it's an orientation and reflection system, not a summary engine or a shortcut to expertise.

Approach

A single context-engineering pipeline: retrieve abstracts from PubMed, embed and cluster them, generate a grounded thematic summary per cluster, extract atomic claims from a user's own free-form reasoning, align those claims to the cluster(s) they're closest to, extract and resolve named entities into a typed knowledge graph, and reflect which claims are well-supported, weakly supported, or ungrounded — without issuing a grade.

Key results

A working structural prototype with the full pipeline implemented end-to-end: PubMed retrieval, embedding and clustering, LLM cluster summarization, claim extraction and grounding, entity resolution into a queryable knowledge graph, and both a static portfolio demo mode and a live backend mode from the same frontend. The open question this prototype is built to eventually answer — how well corpus-relative grounding catches unsupported claims in practice — hasn't been formally evaluated yet.

Status

Early prototype. Core pipeline is implemented and runnable; formal evaluation and an agentic reasoning layer on top of retrieval are planned next.