Overview
Most AI research collapses into one search and a confident summary. This harness refuses that shape: it splits the question, searches in parallel, reads the actual sources, and then tries to disprove what it found before writing anything down.
Claims that survive get cited. Claims that do not, never reach the report.
How it works
The run moves through five phases.
Scope — the question is decomposed into five distinct search angles, so one framing cannot dominate the result. If the question is underspecified — "what car to buy", with no budget, use case, or region — it asks two or three clarifying questions first instead of researching the wrong thing.
Search — five agents search in parallel, one per angle, each blind to what the others surface.
Fetch — URLs are deduplicated across all angles, the top fifteen sources are fetched, and each is reduced to falsifiable claims rather than impressions.
Verify — every claim faces three independent skeptics, each prompted to refute it. A claim dies on two votes out of three. This is the step that separates the harness from a summarizer.
Synthesize — semantically duplicate claims are merged, what remains is ranked by confidence, and the report ships with its sources attached.
Usage
/deep-research <your question>
The more specific the question, the better the angles. Give it a budget, a region, a timeframe — whatever narrows it.