Get sweep progress and results
Returns incremental sweep progress. The default ranked view sorts and may truncate the
display leaderboard. order=natural returns every available row, untruncated, ordered by
deterministic runIx; use that view when materialising durable trial rows.
The ranked view is ordered by plateau score by default, not by the raw objective. A
plateau score is the objective of the worst run in a parameter point's immediate
neighbourhood, so a point scores well only if the region around it also does — the highest
raw score is frequently a spike that does not survive the parameters moving slightly. Pass
ranking=raw for the unadjusted objective order.
Rows in the ranked view carry plateauScore and neighbourCount when plateau ranking
applied. Read them together: neighbourCount: 0 means the point had no neighbours to
compare against, so its plateau score is unevidenced rather than confirmed. Sweeps
submitted before plateau ranking existed have no stored parameter grid to rebuild a
neighbourhood from and are always ranked raw; the response's ranking field says which
ordering was actually used.
A sweep submitted with walkForward answers in a different shape, and the walkForward
field on the response is what tells the two apart — it appears as soon as the sweep is
accepted, before any fold has finished, so it is safe to branch on while polling. There
the leaderboard is one row per completed fold: that fold's winner as it scored
out-of-sample, with runIx carrying the fold index rather than a grid position. The
in-sample runs behind those winners are not retained — they are an optimization's working
set, and only the winner survives its fold. ranking is always raw and no plateau, DSR
or PBO figure is reported: the out-of-sample scores are already the honest number, and
layering a certification computed over F observations on top of them would overstate what
was measured.
Get sweep progress and results Returns incremental sweep progress. The default
rankedview sorts and may truncate the display leaderboard.order=naturalreturns every available row, untruncated, ordered by deterministicrunIx; use that view when materialising durable trial rows.The
rankedview is ordered by plateau score by default, not by the raw objective. A plateau score is the objective of the worst run in a parameter point's immediate neighbourhood, so a point scores well only if the region around it also does — the highest raw score is frequently a spike that does not survive the parameters moving slightly. Passranking=rawfor the unadjusted objective order.Rows in the
rankedview carryplateauScoreandneighbourCountwhen plateau ranking applied. Read them together:neighbourCount: 0means the point had no neighbours to compare against, so its plateau score is unevidenced rather than confirmed. Sweeps submitted before plateau ranking existed have no stored parameter grid to rebuild a neighbourhood from and are always ranked raw; the response'srankingfield says which ordering was actually used.A sweep submitted with
walkForwardanswers in a different shape, and thewalkForwardfield on the response is what tells the two apart — it appears as soon as the sweep is accepted, before any fold has finished, so it is safe to branch on while polling. There the leaderboard is one row per completed fold: that fold's winner as it scored out-of-sample, withrunIxcarrying the fold index rather than a grid position. The in-sample runs behind those winners are not retained — they are an optimization's working set, and only the winner survives its fold.rankingis alwaysrawand no plateau, DSR or PBO figure is reported: the out-of-sample scores are already the honest number, and layering a certification computed over F observations on top of them would overstate what was measured.