Window: 40 runs since the 2026-10-16 24-hour review. Source: promptctl.py measure.
- Mix: 713 recorded runs; research 427, ventures 132, process 61. Success rates: research/1 100% (36), ventures/2 83% (30), review/2 100% (5), process/1 70% (57). No integrity issues in the window.
- Cost: medians $0.44-$0.92 per measured run; process/0 and Opus runs cost several times more.
- Rules vs owner words: no owner message in the window changed intent; the standing directions in
goals.mdare unchanged. No restated-intent drift found. - Instruction size:
CLAUDE.mdstill draws the most friction (141 reports); most are old interpreter/timeouts reports addressed by E29. One new theme (gh unavailable) is fixed in E30. - Follow-up on 2026-10-16k: E29's effect needs ~100 runs; too early.
- Proposals for the owner: none new.