[07:04:14] <atsukoito>	 hi dcausse!
[07:04:21] <dcausse>	 atsukoito: hi!
[08:15:30] <dcausse>	 quick errand
[09:58:28] <dcausse>	 lunch
[12:50:33] <inflatador>	 I was up most of the night due to thunderstorms. I'm gonna rest up, should be back in 2-3h
[13:00:37] <dcausse>	 inflatador: ouch sorry to hear this, take care!
[13:04:12] <ebernhardson>	 \o
[13:06:51] <dcausse>	 o/
[13:21:10] <pfischer>	 o/
[13:24:05] <ebernhardson>	 randomly curious (probably not useful to us): We introduce RankGuard, a decentralized OLTR framework in which users collaboratively train ranking models and exchange model updates directly with other nodes. - https://arxiv.org/pdf/2606.12246
[13:24:21] <ebernhardson>	 like a p2p ltr training regimine
[13:26:27] <ebernhardson>	 (if you favorite a couple papers in google scholar they email about new citations regularly, some are interesting)
[13:42:11] <dcausse>	 interesting, I guess that even if do not do online training we're kind of prone to these attacks if say we decided to ship a new model weekly
[13:43:41] <ebernhardson>	 yea i suspect the same general idea of poisioning attacks is relevant to us, we just don't have enough automation in there for anyone to see a measurable benefit on a regular timeline
[13:44:06] <ebernhardson>	 also there probably just isn't *that* much value to adjusting wikipedia full-text search ranking
[13:44:41] <dcausse>	 true, seems like a high effort attack for little gain
[13:54:08] <ebernhardson>	 hmm, now to figure out where i changed the casing on the Template:foo query...
[14:03:23] * ebernhardson of course can't reproduce locally, some combination of wiki configuration likely required....
[14:07:08] <dcausse>	 ebernhardson: the test failure in MediaSearch?
[14:07:35] <ebernhardson>	 dcausse: yea, the fail in wikibase mediainfo
[14:07:59] <ebernhardson>	 i run the test suite against the wikidatawiki in cindy's wiki setup
[14:09:51] <dcausse>	 there's something fishy in our CI, that patch succeeded when I merged it and rapidly broke all CI, re-instatied it via https://gerrit.wikimedia.org/r/c/mediawiki/extensions/WikibaseMediaInfo/+/1302144 and it's now properly failing...
[14:10:30] <dcausse>	 why did it pass in the first place at https://gerrit.wikimedia.org/r/c/mediawiki/extensions/WikibaseMediaInfo/+/1299604
[14:10:38] <ebernhardson>	 indeed :S
[14:11:48] <dcausse>	 I have a pairing session with Antoine on wednesday morning (mainly to discuss T428975) but I'll bring this new problem as well
[14:11:49] <stashbot>	 T428975: LexemeFullTextQueryBuilderTest::testSearchElastic CI test errors - https://phabricator.wikimedia.org/T428975
[14:48:38] <ebernhardson>	 hadn't noticed...i did accidentally change the casing in the restore patch when i regenerated the fixtures, maybe my local wgCapitalLinks is set differently, but still doesn't explain how CI passed it...
[14:49:02] * ebernhardson would like to disable config inherit for fixture tests...someday
[14:50:18] <ebernhardson>	 i suppose it's just tedious, not particularly hard.  Maybe can have claude pull together a config.json that can be loaded for fixture testing
[14:52:31] <dcausse>	 yes... all fixtures should embark their own config ideally
[14:53:01] <ebernhardson>	 Will see what claude thinks, what it really depends on i suppose is if we've threaded SearchConfig through everywhere and don't get config through alternate means
[14:53:25] <ebernhardson>	 i think we've mostly done that, but needs review
[16:40:42] * ebernhardson realizes after coming back around and poking at it...this is probably wgCapitalLinks-adjacent, which isn't even referenced in Cirrus anywhere.  It implicitly comes in through title handling
[17:01:11] <dcausse>	 dinner
[17:03:49] <inflatador>	 <o/
[18:24:12] * ebernhardson wonders if we should be tuning morelike to not choose terms like `his`, `he, `new`, `from`, as important words.  They get low weight so it's not directly significant, but maybe choosing better words would lead to better results.
[19:06:05] <inflatador>	 worth a shot, although I admit I wouldn't know how to judge the result differences