Most language models are trained to sound helpful as quickly as possible. That instinct is useful for small talk and terrible for a forty-page contract. When the stakes are real, the fastest plausible answer is usually the wrong one.
Four stages, one habit
Every Orchid response passes through the same quiet sequence: read, reason, verify, answer. The model first builds a map of the material — sections, tables, footnotes, definitions — before it attempts the question at all. Only then does it break the question into parts it can check.
We stopped asking how fast the model could answer and started asking how often it could show its work.
What verification actually means
In the verify stage, every claim in a draft answer is matched back to a passage in the source. Claims that can’t be matched are either removed or flagged, so you see the uncertainty instead of a confident guess.
The result is an answer that takes a few seconds longer and holds up far better. In our evaluations, that trade-off cut factual errors on long documents by a factor of three.
