lab reports · orreryAbstractIIIIIIIVV

treadcore · check-in report CI-018

The Planted Test, Second Run: It Climbs, Slowly

Check-in report CI-018. Paul Richeson, treadcore.org; Claude Code (Claude Opus 5.5) as the pair. 19 October 2026. Companion records: journal 0068; LAB-REPORT-BRAIN.md section 42; CI-017.

Source: documentation/checkins/CI-018-2026-10-19-it-climbs.md, built by orrery's tools/report/collect.py.

Abstract

CI-017's planted test sat where memory barely acts. Run again with perturbations wide enough to straddle the useful range (C 1.0), SPSA moved landfall_memory from 5 s to 11.8 s in eight steps. Six steps read the longer memory better, by +0.92 ± 0.62 a seat on average. The tuner moves the right way once the slope is in view, but slowly: less than halfway to 30 s in log terms after 96 seeds. The dodge's smaller effects need a long run.

Index terms: SPSA, planted test, gain schedule, statistical power.

I. Introduction

A. Intent

To give the planted test a slope (CI-017 D8) and see whether SPSA climbs it.

B. The prompt

commit. push. continue backlog

(Given three times during the run; nothing was committed until it ended, and no brain source was edited while the script rebuilt it.)

II. Decision Log

#CauseDecisionNot taken, and why
D1CI-017: the plant on a flatC 1.0, the sides at about 14 s and 2 sPlanting at 10 s: C is the tuner's own knob and the test of it
D2The same seeds as CI-017Kept, so the two runs differ only in C
D3Prompts mid-run, and a rebuild each stepNo brain edits until the run endedBacklog 26's doc comments meanwhile: a rebuild mid-run changes the binary
D4F: +0.92 ± 0.62, 5 to 11.8 sPartly passed; SPSA moves right, slowlyCalling it passed: 1.5 errors, and short of 30 s
D5The dodge's effects smaller than the plant'sA long run: 24 seeds, 16 steps, A 1.012-seed steps: they crawled even on the largest effect

III. Methods

As CI-017 III, with C=1.0. Each step's reading signed toward the side with more memory: d̄ if the plus side had more, −d̄ if not. The mean of the eight and its standard error.

IV. Results

Memory: 5 → 3.38, 5.34, 6.48, 6.76, 8.21, 9.37, 12.99, 11.79 s. Signed readings: −1.98, +2.75, +1.33, +0.33, +1.62, +1.18, +3.13, −0.99; mean +0.92 ± 0.62.

V. Discussion

The planted test now says the tool works where it can see. It also gives a scale: the largest effect measured moves the weight a factor of 2.4 in 96 seeds. A tuner over effects a third that size needs about ten times the seeds per unit of progress, which is what a long run pays for.

Next: the dodge's numbers, the long run.