researcherlisted
Install: claude install-skill ssubbotin/kaggle-research-skills
# Researcher: systematic invention for math, ML, and science
A model trained on physics up to 1900 cannot recombine its way to relativity. When the catalog of
known methods is spent, progress comes from operating on the structure of the problem itself: its
contradictions, its hidden assumptions, its unused resources. This skill is TRIZ (Altshuller)
translated to research work, extended with the recurring heuristics of major scientific inventions:
assumption excavation, symmetry arguments, representation change, analogy transplant.
Its main target is the unknown: questions nobody has answered, beyond the frontier of what is
published. Replicating a known frontier is a means to that end and never the goal (see "Two tracks:
the unknown first, the frontier absorbed").
## When the campaign needs RESULTS: the research loop, binding gap first
Why the loop is built this way: each design choice answers a measured failure. A loop that bans
training returns metadata only. Arms whose own predictions sit several times below their bar
forecast their own nulls: a batch of five arms, about 15 GPU-hours, gave **zero passes and three
harms** after every pre-registration had predicted +0.0001..+0.0012 against a 0.0036 bar. Stacks
of sub-bar cards harm and cannot be attributed. The largest harm was a loss-construction defect
visible in the smoke run. And an arm that changes a distilled student, whose measured weight on
anything but its teacher was 0.005-0.013, changes almost nothing that shi