Abstract
CI-044 found the a-raid-on-a-guarded-base family under its floor (70.2% against 75%) and bisected the drop to 5899e8e, which set the front read's three weights to 0 by default. That commit gave the front scenario its old w_front back so the scenario still tested the front read. It missed this family, whose header names w_front_push as its mechanism. By a rule written before the run, the family now sets w_front_push to its old 0.75 and reads 81.5% +/- 1.9% (326 of 400), close to the 81.8% it read before 5899e8e. Every family holds and the brain's 87 tests pass. The defaults are unchanged.
Index terms: scenario families, regression, test fixtures, defaults.
I. Introduction
A. Intent
To commit and push LR 77 once the turtle census had landed, then fix backlog 62.
B. The prompt
commit and push when turtle lands. then fix backlog 62
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | "when turtle lands": turtle's census had finished (pooled.txt, 08:18) and was written up in LR 77 | LR 77 committed and pushed on main (7c18851) | Wait for another run: none was playing |
| D2 | The family's header: the push point does it, 85.5% against 50.5% with w_front_push at 0 | Set w_front_push alone in the family's brain | All three front weights at their old values: the family tests one of them |
| D3 | 5899e8e's precedent: the front scenario sets w_front = 0.5 | The weight in the family's toml, with a comment why | Lower the floor: the family then tests nothing it was written for. Turn the push back on by default: LR 75's census found it costs in play |
| D4 | The rule before the run | Written to build/b62/NOTES.txt: every family at or above its floor; if not, stop, no second weight | — |
| D5 | F: 81.5%, every family holds | Backlog 62 done | — |
III. Methods
scenarios/families/a-raid-on-a-guarded-base.toml: weights = { w_front_push = 0.75 } under its [[brain]], and a comment in the header. make families (400 seeds for this family, its own seeds) and cargo test -p treadcore-brain --release, on the Mac at 7c18851.
IV. Results
In LR 80: the family at 326 of 400, 81.5% +/- 1.9%, floor 75% (281 of 400, 70.2%, without the weight); every other family as before; 87 tests passed.
V. Discussion
A scenario that tests a mechanism has to keep the mechanism on when a default turns it off. 5899e8e did that for the scenario it knew of and missed the family. The rule CI-044 added to CLAUDE.md (make families at each default change) would have caught it on that commit. The fix is in the test, not the brain: the push point is still off for play.