Abstract
The first night of backlog 69: about ten hours on the census machine, queued at 18:40. Block 1 reads eight profiles against the new defaults. Block 2 screens sixteen weights at half and double. Block 3 runs SPSA over the five weights the screen ranks first, chosen by a script written before the screen, then reads its point on 400 held-out seeds. Each block's rule and our expectations were written before it played, so the morning can score the predictions as well as the brain. Nothing changes a default overnight.
Index terms: brain census, slope screen, SPSA, preregistration, league.
I. Introduction
A. Intent
A night's census, planned to about ten hours, with the expectations written first.
B. The prompts
so good. can you plan about 10 hours a census.
yes. please. and write about our expectations. the longest run.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | LR 87's plan, step 1 the slope screen and step 2 a league | Both tonight with existing code; the retreat fix (code) waits | Write the retreat fix first: a night lost to review |
| D2 | A 110-seed census about 8 min, 400 about 25 | 8 + 32 censuses of 110, SPSA 24 x 48, one of 400: about 8 h 40 min with slack | More weights: past ten hours |
| D3 | SPSA's weights must not be picked by eye after seeing the screen | pick.py, written before: the five largest gaps | Pick in the morning: a night idle |
| D4 | The playtest held the machine's screen | Closed before the run | Run beside it: both slow |
| D5 | "write about our expectations" | Predictions per run, in NOTES and LR 88, before | — |
| D6 | One step at a time; Paul's word for defaults | No default changes overnight; candidates read again later | — |
III. Methods
build/night/run.sh on the machine at 7c97aa9: make census with TRIAL each profile (110 from 160000), each screen file (110 from 170000), pick.py, make spsa (TUNE the five, STEPS=24 SEEDS=48 SEED0=200000), the held-out census of theta.toml (400 from 300000). Each census's directory in build/night/NAME.dir; a line a census in build/night/progress.txt.
IV. Results
The night ran 18:40 to 00:42, six hours, not ten. Read in full in LR 88's morning part; the morning's steps are CI-054.
- Block 1: six of nine expectations held (read at 19:33).
- Block 2: no candidate; seven rows outside two errors, every one below the defaults; nine weights flat; w_exposure and w_mine read exactly zero.
- Block 3: pick.py chose strikes_max, longshot_every, w_focus, mass_for, castle_inside. SPSA's point read +0.597 +/- 0.298 (z 2.002) on 400 held-out seeds, no map past two errors below.
- Expectations: 12 of 19 held across the night.
Errors: the plan's timing (a census of 110 seeds 6.5 min, not 8).
V. Discussion
Writing expectations down turns the night into two tests: one of the brain, and one of what we think we know about it. LR 74's flat ridge and LR 86's cowboy are the two priors in tension tonight.