Type Alias SweepRunRow

Same units as cagr on the single-run result — a ratio, not a percent.

Probability that this run's Sharpe reflects real edge rather than the best draw from however many parameter vectors were tried. Above ~0.95 the result survives the multiple-testing correction; near 0.5 or below it is indistinguishable from the best of a pile of coin flips. Absent on aborted runs, on sweeps with too few trials to establish any dispersion to deflate against, on runs with fewer than 3 period returns, and on a degenerate (near-constant) return series — all cases where the underlying statistic isn't meaningfully computable, rather than genuinely zero. A present value is the computed probability, however small.

Present whenever this trial actually has a curve — which is any completed run of a sweep that requested curves at all (equityCurve.mode topN/topPct), not only the ranked winners. Absent, not null, when there is genuinely nothing: the run aborted, it made no trades, or the sweep never retained curves (mode auto/none). Selection decides how the curve travels, not whether it exists. A row may carry url alone, or url together with the points inline. Read points/equities to tell whether the curve is inline — never the presence of this object, and never the absence of url. Both mislead, and in opposite directions: this object is present on rows that carry only a pointer, and url stays present alongside an inline curve so a caller can still re-request a different transform. An inline curve is emitted on the materialisation view (?order=natural), where each row is read once; the ranked view keeps the pointer, being polled. When only a pointer is present, GET the url for the curve. Its own meta there is the real, possibly size-guarded outcome. This outer meta may be a declaration rather than a measurement — for a row whose curve was not promoted, the point count is derived from the trade count (one equity point per closed trade) instead of being read from the store, to keep a polled response from doing per-row work.

Same units as maxDrawdownPercent on the single-run result — percent (0-100 scale).

How many neighbouring parameter points backed the plateauScore. Zero means the point had no neighbours in the grid, so its score is unevidenced rather than confirmed — the value alone cannot be distinguished from a genuinely robust one.

The objective of the worst run in this point's immediate neighbourhood — how well the region around it holds up, not how well it scored itself. Present only in the ranked view when plateau ranking applied. Always read together with neighbourCount.

Absolute net PnL in the output currency.

Same units as pnlTotalPercent on the single-run result — percent (0-100 scale).

Present only in the ranked view.

Deterministic zero-based expansion index, stable across shards and ranking.

Same units as winRate on the single-run result — a fraction, 0.0-1.0 (a rate, not a percent).