[rusholme]

What and why

socsims-1021 task 5: the team page. A log post, log/2026-10-27-socsims-1021.md, sets run 158 (Russ Laher's 10/21/2026 socsims window, 1,998 images, and its 10/20 pilot of 72) beside his write-up: the selection and configuration, jobs and exit codes, timings, the catalog-count table with his column, the per-image comparison (medians, quartiles, the ten most different), the alert-side table, the gate's held sets, the checks, the gallery verdicts with the panels' paths, and what differs from dev. system/datasets.md's socsims section points at it.

Files

• log/2026-10-27-socsims-1021.md: new.
• system/datasets.md: one sentence linking the post.

Divergences from the brief

• The page is a log post, not a system page: AGENTS.md says dated reviews go in the log, and the site has no results or testing section.
• No decisions.md line: ruling 2 was not taken (SFFT published under the socsims block, no Singular matrix), and the overlay's one divergence from dev the page names, the reference tie-break, is the loop's and is already recorded on datasets.md. The cap and the gate are existing rulings.
• Gallery paths are given relative to the outputs root (158/qa/full/), with no bucket name, as datasets.md does for the inputs bucket.
• Missing numbers, written as missing: SMDC's per-step SFFT and PSF-fit seconds (the difference job does not record them); whether Russ's references carry the same dark pits (not checked).

Measurements

Every number is from the brief's t2/ (pilot NUMBERS, VERDICTS, CHECK-READING), t3/ (stats, passes.log, PR 500 and 501 bodies) and t4/ (CHECK, NUMBERS, VERDICTS, compare-russ.md), and Russ's from the brief's extraction of his socsim_tests.rst. Headlines: 1,998 of 1,998 published, 0 failed; SFFT+ PSF-fit median per image 11,856 vs 10,396 (ratio median 1.154); SFFT+ SExtractor 146,053,130 vs 145,903,083; held 1,997 of 1,998 sets (core 1,993, median +0.066); panels 13 of 15 scene removed, 2 partly (one SCA 2 pointing, the same in Russ's difference).

Improvements seen

• The difference job records no SFFT or PSF-fit catalog seconds, so the timing table cannot set those steps beside Russ's, which were 83% of his instance time. Seen, not built.

Session notes

none

What the next task must know

Task 6's runbook housekeeping line is independent of this page.

Anything left

none

🤖 Generated with Claude Code