The rules-file comparator read: no rules file carries a half-split contrast¶

The 44 TIGRIS jobs pinned at 05465f1, every one completed (logs in datasets/logs/): 22 distillations, 229778 and 229779 the written guides and 229780 to 229799 the first ten admissible partitions of each organization (19.2 GH200-hours, 40 to 85 minutes a partition), and 22 ARMS=rules evaluations, 229800 to 229821, each queued behind the files it reads (12.0 GH200-hours). Read by retrieval_read.py --arms rules (rules-reading-openstack.json, rules-reading-wikimedia.json) with the retrieval foreign jobs as the base arm. The adapters' row reproduces the retrieval read's exactly, so both reads cover the same partitions and examples. Generated from the artifacts:

Superseded on 2026-10-08 for the reported contrasts. At the registered decoder (penalty 1.0) the read contrast is unchanged, but the reported ones are not: see "The rules files against no file and against each other, at penalty 1.0" below.

By the rule fixed before the run, distilled own minus sibling carries no contrast as large as the SESOI on either organization: both intervals sit inside the SESOI band (0.01), and they are narrower than the retrieval arm's or the adapters'. Reported, not read: neither a distilled file nor the written guide lifts exact match over the base model (OpenStack's estimates are near minus the SESOI, intervals touching zero); a file from the organization's other half does no better than the other organization's; and the distilled and written files score alike. The one interval clear of zero is Wikimedia's written own minus written foreign, +0.0056, below the SESOI. The adapters' own minus sibling stays inconclusive, so the plan's question (does any file carry a contrast the adapters carry) has no adapter contrast to compare against on the development window; what the read does show is that a rules file in the system turn moves the 7B model's revisions by less than the SESOI. Exploratory: nothing here binds a verdict, and the test window stays sealed.

Nineteen logs carry one tokenizer notice, a 45,388-token sequence against the model's 32,768: every Wikimedia partition's distillation and evaluation but partition 5's, and OpenStack's foreign evaluation, which builds Wikimedia's pools. build_supervised tokenizes each example whole before refusing any over 2,048 tokens, so that example is left out of the pools as it is out of the adapters' training sets, and no model reads it. Some retrieval jobs' logs carry the same notice.