dsh-deep-research

Most AI research tools share the same blind spot: a URL pasted beside a claim is treated as evidence, and twelve outlets are treated as twelve independent witnesses. Nobody checks whether the cited sentence actually appears on the page, and genuine refutations drown in a sea of syndicated support.

Kestrel closes those gaps with code, not prompts. Every citation must survive a literal substring (or normalized, or fuzzy) match against the stored source text before it is admitted, so a hallucinated quote simply fails the gate. MinHash-LSH similarity feeds a lineage DAG that Tarjan SCC condenses, so twelve outlets republishing one wire article count as a single origin. An adversarial tribunal pairs a prosecutor hunting refutation with a defender seeking support; one verified contradiction hard-caps confidence regardless of how many 'supporting' sources pile up. Verified claims persist in a local evidence graph with per-claim volatility-based freshness: five years for theorems, twenty-four hours for stock prices.

If you need an auditable research trail, Kestrel fits: analysts, researchers writing citation-heavy work, or any team skeptical of AI-generated conclusions. It ships with zero runtime dependencies, embeds as a standalone Node library, and degrades to a deterministic lexical judge when no LLM is present, so a research run never silently returns an empty result.