Precedent, warm and cold

Do the passes most similar to this one, by what is knowable before the waterfall is opened, carry the same recorded outcome more often than chance?

Agreement at 5, per arm

739 decisively labelled passes from 116 stations and 295 satellites. Chance is 0.531, which is the label mix alone.

Each line is one arm. The filled mark is the warm condition, which allows any neighbour; the hollow mark is cold, which forbids the query's own station and satellite. The dashed rule is the label mix alone. The length of a line is what the entity restriction costs that arm.

Four retrieval arms over the same candidate pool, under two conditions. Warm allows any other observation. Cold requires a different ground station, a different physical site and a different satellite, because in this corpus the outcome is partly a property of who recorded it. Agreement at 5 is the mean over queries of the share of retrieved neighbours carrying the query's own label.

Arm	Warm	Cold
Granite embedding of the card	0.618	0.554
Standardised numbers, Euclidean	0.592	0.537
The station's own recent passes	0.605	not applicable
Uniform draw from the same pool	0.530	0.528

The cold condition excludes candidates from the query's own station, so this arm has no definition here. Reported as null rather than as zero agreement, which would be a measurement nobody made.

What survives correction

Every margin is a paired difference over the same queries, resampled by ground station and widened by Bonferroni over the comparisons this study makes.

Comparison	Condition	Margin	95% interval	Corrected	Survives
Granite against chance	warm	0.0880	[0.0465, 0.1344]	[0.0338, 0.1558]	yes
Numbers against chance	warm	0.0620	[0.0279, 0.1046]	[0.0184, 0.1217]	yes
Station history against chance	warm	0.0800	[0.0290, 0.1409]	[0.0112, 0.1672]	yes
Granite against the numbers	warm	0.0260	[-0.0044, 0.0557]	[-0.0169, 0.0660]	no
Granite against chance	cold	0.0260	[-0.0196, 0.0763]	[-0.0353, 0.0998]	no
Numbers against chance	cold	0.0092	[-0.0107, 0.0321]	[-0.0185, 0.0432]	no
Granite against the numbers	cold	0.0168	[-0.0261, 0.0601]	[-0.0406, 0.0783]	no

Warm, all three retrievers beat chance and survive the correction. Cold, none of them does: the interval contains zero. The embedding also does not beat seven standardised numbers under either condition. That is the finding, and it is the reason the queue is not ranked by this.

The index

The measurement above is exact cosine search. This is what a real vector index returned for the same queries.

Backend

chromadb 1.5.9, cosine, in memory

Recall at 5, warm

0.9992

over 739 queries

Recall at 5, cold

0.9995

over 739 queries

Embedding

granite-embedding:278m

277.45M, F16

The share of the exact top-5 that the index also returned, per query, averaged.

The cold condition is answered by a metadata filter on station, site and satellite inside the index rather than by filtering afterwards, so this also checks that the filter and the exclusion rule agree.

When the exclusion rule gained the site and the filter did not, this number fell from 0.94 to 0.77 on the cold condition, which is what it is here to do.

What this does not measure

Whether a reviewer shown these neighbours decides faster or better. That is kill gate 4's territory. Gate 4 has been answered on the decidability of its sample and not on whether a note or a neighbour helps, so this question is still open. Its verdict word lives in its own receipt and is not restated here, because a literal in this file cannot follow it.
Whether the label a neighbour carries is correct. The network's own verdict on an observation is a silver label, and agreement with it is agreement with the network rather than with the sky.
Anything about the images. Every arm here sees only what is knowable before the waterfall is opened, which is the point, and it means none of these arms is a detector.