Abstract
The night's SPSA point passed its held-out rule by 0.002 of an error. The rule's second guard, make families, passed: all nine hold, the duel 5 points lower, inside the noise. Because the pass was so narrow and the point lay far outside the range the screen read, one more census read it on 400 fresh seeds, its rule written first: +0.09 +/- 0.27, inside the noise. The point is not promoted; no default changes. The second read caught what the first held-out read let through.
Index terms: brain census, SPSA, held-out read, replication, scenario families.
I. Introduction
A. Intent
Fetch the night's results from the census machine, score them, and take the SPSA point through the rest of its rule.
B. The prompts
good morning. remember our last conversation. what do you say>
yes. conenct and get reults.
yes. start with families
yes. i love the work you are doing. give yourself a tip.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | No session memory; the repo is the record | Read journal 0100, CI-053, LR 88 first | — |
| D2 | The machine's name is in no record; ~/.ssh/known_hosts | Probe the LAN hosts read-only with BatchMode: the machine is saskatchewan, 192.168.8.235 | Ask Paul: two probes cheaper |
| D3 | z 2.002, a pass by 0.002 | Report it as at the bar, not as a clear pass | — |
| D4 | The night's rule: families in the morning with Paul | make families SET= the five on the Mac | — |
| D5 | Families at the point alone show no change | Families at the defaults beside it (30 s) | — |
| D6 | A pass at the bar from a point outside the screen's range; LR 74 | A second read, 400 fresh seeds from 400000, rule and expectation in build/reread/NOTES.txt before it ran | Promote on the rule's letter: LR 74's lesson unheeded |
| D7 | The second read inside the noise | Not promoted; defaults stay | — |
| D8 | No code or default changed | No CHANGELOG entry (LR 87 and 88 the precedent) | — |
III. Methods
On the Mac at 7c97aa9:
make families SET="strikes_max=46.6825 longshot_every=11.0390 \ w_focus=0.4439 mass_for=52.3568 castle_inside=18.6594" make families
On the census machine at 7c97aa9, in tmux:
TRIAL=build/reread/theta.toml SEEDS=400 SEED0=400000 MAP=all \ JOBS=24 make census
The census in build/census/20261026-073751; every log ends in test result: ok. z to four places from the logs' "a seed:" lists, as scripts/pool.sh reads them.
IV. Results
| Read | Every map | z |
|---|---|---|
| held-out, 300000 (the night) | +0.597 +/- 0.298 | 2.002 |
| read again, 400000 | +0.088 +/- 0.267 | 0.33 |
| the two pooled, 800 seeds | +0.34 +/- 0.20 | 1.71 |
Families: all nine hold; the duel 73.5% +/- 2.2% at the point, 78.5% at the defaults; the rest unchanged. Fortressislesx below both reads (-1.33 +/- 0.66 pooled). Per map in LR 88.
Errors: none in the runs. The first draft of LR 88's morning part miscounted the expectations (13 of 22; 12 of 19) and claimed a step scaling for SPSA not checked; both corrected before commit.
V. Discussion
The rule as written would have promoted a point worth nothing. A single held-out read at two errors is right about 97 times in 100 when there is nothing to find, but SPSA points are chosen for reading high, and the bar meets that bias at its weakest. Two proposals for Paul (LR 88): promote an SPSA point only after two held-out reads, or one of 800 seeds; and bound SPSA to the range the screen has read.