Abstract
The overnight run played 30 generations and 380 censuses on the new defaults, with no failures. Seven trials read past two errors on 55 seeds. None passed both held-out reads of 400 seeds, and one passed one. There is no candidate, as the run expected. The 0.2.0 draft release waits for day 1, and the new report pages are live.
I. Introduction
A. Intent
The morning read.
B. The prompt
goodmorning. please check
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | LR 100's check-in was left unwritten at midnight | Written and committed first | — |
| D2 | No trial through both held-out reads | No candidate; the run goes on | Read gen25/sample-2 a third time now: one pass in two is the noise's shape at this rate |
| D3 | Day 1 of 0.2.0 opens at 18:34 | Left to Paul (DEPLOY.md) | — |
III. Methods
Over ssh: the run's progress.txt and candidates.txt; gh release view v0.2.0; gh run list; curl of the three report pages.
IV. Results
LAB-REPORT-BRAIN.md section 101's table: seven trials, fourteen held-out reads, one past two errors (+0.58 +/- 0.26).
V. Discussion
Seven readings that looked like gains on 55 seeds fell back into the noise on 400. That is the reason the held-out rule exists.