Mix since 2026-10-20 (48 merged runs): venture deepenings (ai-bookkeeper, wall-framing-takeoff, open-house-plans: flagship-style depth, not new calculators), research follow-ups, redteam, process slices, creative. This matches "depth over calculators" and "deepen existing projects" (goals.md). No near-copy streak seen.
Owner's words vs output. Every starting message in the window is the routine trigger ("read ./prompt.md ..."), so there is no new owner text. The 2026-10-20/21 quotes ("cut the deadweight", "guidelines not rules") are reflected in CLAUDE.md and steering.md.
Process runs drifted to slivers. 19 of 70 process runs since 2026-10-20 ended partial, mostly one-deletion slices of P036/P040/P043. That is the opposite of "a person could not finish in twenty hours". Fix: process candidate 2 (take up to three rows per run), E61.
Rules.steering.md 171 lines (budget 120): cut the "improvement loop" section (restated process/prompt-1.md), -10 lines. prompt.md is 103 (budget 100) and review/prompt-1.md 138 (budget 130); unchanged.
Lesson from outside: Netflix's "context, not control" (Hastings and Meyer, No Rules Rules, 2020): give people the facts and let them decide, and keep process only where an error is irreversible. Fits the hard limits vs guidelines split already in CLAUDE.md.
Audit (3 non-doc PRs, re-run): free-modelcost (4 tests pass), infra-netting (7 tests pass as README says, via pytest-style functions, pytest absent so run by hand), open-house-plans (11 tests pass, no-warranty text on the page). No corrections.
Follow-up: E55 proposal (retire rule needs 2+ integrity issues) is still open; research candidate 1 (99%, 128 runs) vs champion (98%, 712) is a tie in practice.