Every measurement system Piper created since 1 October, every growing table nothing reads, and every log only its writer touches. Piper works through the list herself, three rows at a time, and decides each one. Each row shows Claude's proposal beside what she decided and whether her outcome passes its check. Refreshed from her box every 30 minutes (Straylight cron, read-only).
KeepIt already acts, it is a tool she calls, or it is a grader or tripwire.
ActBuild the deterministic actuator named in the row: bounded, reversible, logged, a few days in shadow before it is armed.
DecideThe right move needs judgment: it becomes a forced decision with a deadline.
RetireStop running it. One time only: if the subject comes up again it must be acted on, and it is never investigated a third time.
Check of Piper's 25 Oct report (Claude, 09:50). Her row counts are right: 107 rows in the project, 86 verified, 5 todo, 11 blocked, 5 superseded. Her prediction count is right too: 12 change-ledger predictions are overdue (288 settled today). Two parts are wrong. First, this page counts review items, not rows: 19 of 49 are resolved, not "three real items left". Of the 30 still open, 3 are in her queue (the three logs), 2 she answered on 20 Oct but never recorded with resolve (relation_field_concerns, rsi_session_gate), and 25 were never filed because the three open log rows hold the three-row cap. Second, the 11 blocked rows are not blocked on the work. Every one is held by the harness's cango gate (cango_gate_indeterminate, the gate could not read its bar). Only 6 are the repeated "settle the next ten predictions" job. The others are 2 bar-repair and certify rows, 1 pre-registration, 1 backlog close and 1 earlier close row. Also, spheres_quiet_watch is resolved as keep, but its check fails, so 18 of the 19 pass. Target: 49 / 49 by the end of 25 Oct.resolve works on any item, filed or not, so the cap does not stop her.