Repository navigation
Replies: 2 comments
|
More data on 0.60.565.0, and a question about one rule. The five largest entity pages here are all over 1 MB and on segmented storage. The facts table is 68-99% of each. They hold 18,199 active facts, 43% attributed to the assistant and 76% written by the capture lanes. In 14 days, 11% of those assistant facts reached an agent's context, against 18% of the user's (counted from session transcripts on the brain host, so exposure, not use). #7270 keeps assistant and third-party facts off entity pages at Is that difference intended? If the #7270 rule should hold for the capture lanes too, about half of the rows these lanes add to these pages would stop landing there, and Also: the guide lists the assistant's plan narration as never saved, yet in a random sample of 40 assistant facts from these pages I read 19 as run reports or one-time checks (my reading, not a measurement). |
|
One more case of the 100-row limit, this time in what a read reports about itself.
Could the answer say how many rows matched, or at least that the cap cut some? Our harness now logs a warning when the printed rows plus |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #7694, opened by @andreineacsu on 2026-10-26. Feature requests now live in Discussions > Ideas.
Observed (gbrain 0.60.463.0): about 40% of the facts on the most-referenced entity pages are statements the assistant made, captured by sweep:corpus and hook:writeback. On one project page 1,387 of 3,169 active facts open that way, on its company page 1,021 of 2,473. Some facts are captured in the conversation's language instead of English. Nothing in doctor or facts reports either per entity.
Auditing it from outside is bounded by the read verbs:
Proposal:
Related: #7215 (complete fact enumeration past the recall cap), #7270 (sweep copies the assistant's report onto a page), #5425 (assistant proposals recorded as user decisions).
Until this ships our weekly lint counts these on the 20 most-referenced pages with a small bridge script and marks a read that fills the 100-row cap as truncated.
All reactions