RFD 1147 details: the scorer’s cost and its scale

Measured on the desk described in logbook-fourloops-first-runs.md, which holds the apparatus these figures come from. The entry says how the runs were made and this file says what they returned; re-running them needs the entry.

What it returns when the instruction matched

scale max consistency a second axis overall
10.0 9.2 2.0 4.29

The overall figure is 4.29 out of 10.0, and the spread behind it is the reason to carry both. Consistency scored 9.2 while the second axis scored 2.0, so a scorer reporting only the overall would say “middling” about a result that is excellent on one axis and poor on another. Which axis failed is the actionable half, and the overall alone discards it.

The negative control. A nonsense instruction returns 0.0 overall. That is what makes 4.29 a measurement rather than an impression: a scorer that cannot return 0.0 for nonsense is not discriminating, and every score it gives is consistent with it having read nothing. editscoreNonsenseOverall = 0.0 in fourloops-plan.usda carries that control beside the quantity it validates.

What it costs

configuration peak seconds
NF4, 512 x 512 6.7506 GiB 28 to 36
NF4, 1024 x 1024 8.6 GiB

6.7506 GiB is measured; the plan rounds it to 6.75 in editscorePeakGib512, and carries 8.6 as editscorePeakGib1024.

The seconds are a range rather than a median, and deliberately. The run-to-run spread at 512 is wider than the gap between the two OmniGen2 precisions RFD 1144 reports, so a single figure would imply a resolution these runs do not support.

Why this lives in its own document

Every loop in fourloops-plan.usda names EditScore as its score stage, and loop 1 pairs it with the Referee because an image scorer and a body scorer answer different questions. A figure that four loops depend on belongs beside none of them in particular.

It was in RFD 1144 first, on the reasoning that loop 2 is the smallest loop exercising the scorer and so its numbers were the plan’s. That reasoning holds for OmniGen2, which loops 2 and 3 use and loops 1 and 4 do not. It does not hold for a stage all four share.