Get sweep sensitivity surfaces
How the objective moves as each parameter moves — the question a leaderboard cannot answer.
A leaderboard says which point won; a sweep can spend its entire budget on an axis that
never moved the objective at all, and showing only the top rows hides that completely.
A marginal takes one axis and collapses every other one: for each value of that axis,
it aggregates every run that used it, whatever the rest of the parameters were. A flat
marginal means the axis is irrelevant over the range swept. best, mean and worst are
all reported because them disagreeing is itself the signal — a value with a high best and
a poor mean works only in specific company, which is an interaction between parameters
and would be invisible behind a single number.
A heatmap does the same over a pair of axes, where that interaction becomes visible
directly.
Served from the sweep's stored rows: no re-run, no engine call, and it works on a sweep
still in flight — the aggregates then describe the runs finished so far. Aborted runs are
excluded throughout, since a run that threw measured nothing and counting it as a bad
outcome would invent evidence against a parameter value that was never really tested.
This is a separate endpoint rather than extra fields on the result view because the
two-dimensional half is quadratic in the axis count (N axes give N(N-1)/2 surfaces, each
the product of two axes' value counts) and is not wanted on the poll that drives progress.
Get sweep sensitivity surfaces How the objective moves as each parameter moves — the question a leaderboard cannot answer. A leaderboard says which point won; a sweep can spend its entire budget on an axis that never moved the objective at all, and showing only the top rows hides that completely.
A marginal takes one axis and collapses every other one: for each value of that axis, it aggregates every run that used it, whatever the rest of the parameters were. A flat marginal means the axis is irrelevant over the range swept.
best,meanandworstare all reported because them disagreeing is itself the signal — a value with a highbestand a poormeanworks only in specific company, which is an interaction between parameters and would be invisible behind a single number.A heatmap does the same over a pair of axes, where that interaction becomes visible directly.
Served from the sweep's stored rows: no re-run, no engine call, and it works on a sweep still in flight — the aggregates then describe the runs finished so far. Aborted runs are excluded throughout, since a run that threw measured nothing and counting it as a bad outcome would invent evidence against a parameter value that was never really tested.
This is a separate endpoint rather than extra fields on the result view because the two-dimensional half is quadratic in the axis count (N axes give N(N-1)/2 surfaces, each the product of two axes' value counts) and is not wanted on the poll that drives progress.