RFD 1147: What EditScore costs and returns

Problem

EditScore is the score stage of every loop in fourloops-plan.usda. Loops 1 through 4 each name it, and loop 1 pairs it with the Referee.

Its measurements were written into RFD 1144 because loop 2 is the smallest loop that exercises it. That put a figure belonging to four loops inside the document for one of them, so a reader of loop 3 had to know to look in loop 2 for what the scorer costs.

Decision

Give the scorer its own RFD, and let the loops cite it.

DETAILS.md holds what it costs and what it returns: peak memory at 512 and at 1024 square, the overall score for an instruction the edit matched, and the two-axis spread behind that overall. Both fourloops-plan.usda and fourloops-etnf.usda name it as a source, so a quantity in either layer resolves against the scorer’s own document.

Its negative control is what makes the number a measurement. A nonsense instruction returns 0.0 overall. A scorer that returns a middling figure for nonsense has not discriminated, and its score for a real edit would mean nothing.

logbook-fourloops-first-runs.md keeps the apparatus, and this RFD keeps the result. The entry says how the runs were made, which is what lets somebody make them again.

Details

The measurements and the retractions