storm
Goal: take any problem statement — technical or not, from a user who may know nothing
about the domain — reach shared understanding on one decision through an interview,
then autonomously build deep, sourced knowledge across research rounds and land on a
pick, recommending only when the evidence is strong enough, stating what is still unknown
first, and handing the remaining fog to wayfinder.
Method from Stanford OVAL's STORM (multi-perspective question asking, NAACL 2024) and
Co-STORM (a moderator mining uncited sources for new directions, EMNLP 2024). Bias
controls from open science (pre-registration and Popper's falsifier), Klein's
premortem, Wald's survivorship lesson (the graveyard), and the council pattern
(fresh voices that see only the question). Companion to compass (breadth across named
alternatives) and oracle (vertical unknown-unknowns); storm is the heavier sibling.
The interview is the only gate (plus one skippable checkpoint, Step 7). After the user confirms shared understanding, every remaining step runs autonomously through to the recommendation.