[rusholme]

What and why

decision-reference-by-field is rewritten to the rule rapid PR 450 merged (d7e67597). The loop now builds a field's reference from the selected official L2 rows before the bound whose overlapfields list the field, of any SCA, as dev's get_overlapping_l2files does. loop.md restates the reference stage and the reference bound to match. The anchor name stays, so datasets.md's link still resolves.

Files

• system/decisions.md: decision-reference-by-field.
• system/loop.md: the reference stage's item and the reference-bound section.

Divergences

Membership parity, not full parity: dev's distance-from-tile-centre ordering is not ported, and the entry says so.

Measurements

	before	after
flight 1049 e4 reference	run 118, submitted by hand (job 23131: e1, e2, e3)	run 126, built by the loop (job 23193: e1, e2, e3), no hand job
e4 core / ring / usable / packets	+0.0013 / −0.0050 / 0.781 / 328	+0.0013 / −0.0050 / 0.781 / 328
HLWAS injected, detected of 600	545 (run 116)	546 (run 127)
false positives per epoch, mean / median / highest	213.7 / 228 / 233	213.3 / 230 / 233

The raw evidence is in brief ref-overlap's t1/ and in the runbook step.

Improvements seen

datasets.md's flight-1049 row still describes the earlier rule as current. Brief self-contained owns that page this weekend, so the row is left for the next datasets edit.

Session notes

none

🤖 Generated with Claude Code