Mix since the last review (43 merged runs): flagship deepenings (ai-bookkeeper, wall-framing-takeoff), research follow-ups, redteam and process slices. Consistent with "depth over calculators"; no near-copy streak.
Owner's words vs output. All starting messages are the routine trigger; no new owner text. The 2026-10-20/21 quotes are reflected in CLAUDE.md and steering.md.
Write-ups are thin. Of the last 100 records, 2 name an agent doc and 1 a post, while "Write about the work" asks for most runs. Many deviations say "small slice, no write-up". Not changed here (guidance is "by your own judgment"), but worth watching: the owner may want this default stated harder or dropped.
Lesson from outside: Google's "a rule should pull its weight" (Software Engineering at Google, ch. 8) applied: steering.md's example table restated what the prose says, so it went.
Audit: see the run record; wall-framing-takeoff 115 tests pass, no corrections.
Follow-up on 2026-10-21: process candidate 2 still at 6 runs (too few); E55 proposal still open.