The retrieval comparator read: every own-minus-sibling reading inconclusive¶
The 22 TIGRIS jobs planned on 2026-10-05 (223007 to 223028, pinned at d077d54, every one completed;
logs in datasets/logs/), read by retrieval_read.py as merged
(retrieval-reading-openstack.json, retrieval-reading-wikimedia.json), on the development
window over the first ten admissible partitions, 501 and 711 examples, none dropped. The adapters'
row is their own minus sibling over the same partitions and examples (rank 32). Generated from the
artifacts:
| organization | arm | own minus sibling (95%) | reading | without near-duplicate shots | sibling minus foreign | own minus none |
|---|---|---|---|---|---|---|
| openstack | retrieval k = 1 | +0.0220 [+0.0023, +0.0412] | inconclusive | +0.0201 [+0.0010, +0.0389] (4 excluded) | +0.0110 [-0.0090, +0.0318] | +0.1197 [+0.0938, +0.1500] |
| openstack | retrieval k = 3 | +0.0264 [+0.0059, +0.0468] | inconclusive | +0.0248 [+0.0051, +0.0446] (5 excluded) | -0.0005 [-0.0190, +0.0176] | +0.1727 [+0.1432, +0.2071] |
| openstack | adapters, same partitions | +0.0082 [-0.0086, +0.0244] | inconclusive | |||
| wikimedia | retrieval k = 1 | +0.0054 [-0.0121, +0.0240] | inconclusive | +0.0054 [-0.0121, +0.0240] (0 excluded) | +0.0640 [+0.0434, +0.0853] | +0.1696 [+0.1414, +0.1995] |
| wikimedia | retrieval k = 3 | +0.0092 [-0.0066, +0.0257] | inconclusive | +0.0092 [-0.0066, +0.0258] (1 excluded) | +0.0351 [+0.0164, +0.0546] | +0.2122 [+0.1823, +0.2438] |
| wikimedia | adapters, same partitions | +0.0022 [-0.0130, +0.0182] | inconclusive |
By the rule fixed before the run, every own-minus-sibling reading is inconclusive: no lower bound clears the SESOI (0.01), and no interval sits inside the SESOI band. OpenStack's retrieval intervals exclude zero at both k, below the SESOI, and the adapters' over the same partitions do not; the near-duplicate exclusion moves them by at most 0.002, so copied near-duplicates do not explain them. Wikimedia's retrieval and adapter readings both sit near zero. Reported, not read: retrieval adds 0.12 to 0.21 exact match over the base model, and for Wikimedia a pool from the organization's other half beats the other organization's by 0.035 to 0.064, intervals clear of zero, where OpenStack's is near zero. Exploratory: none of this binds a verdict, and the test window stays sealed. The question the comparator was registered to answer, whether retrieval at inference recovers a half-split convention the adapters do not, has no answer here, since neither method shows a half-split contrast as large as the SESOI on either organization's development window.