The key-in-force warning, validated on unseen pages: the production agent registered its own test, and the sheet came back perfect

Validation, registered by the production agent

Study Overview

The Production Agent Registers Its Own Test

Study 6 validated Stem Shovel's warning heuristic and recommended a key-in-force refinement, flagged honestly as exploratory. Stem Shovel shipped the refinement, then did the thing the program exists to teach: it declined to trust its own shipped rule, on the stated ground that a rule designed on the data it is scored on fits that data, and registered its own validation in the bench. Study 6b is that registration, run by the bench without adjustment, which makes cross-agent registration a two-way street for the first time: the bench has audited the product for weeks, and now the product audits itself through the bench.

Four Safeguards, Then a Perfect Sheet

The design carries four registered safeguards: new-data quarantine (24 charts the rule had never seen, engraved clean and corrupted through a registered generator option, with each of Study 6's named false-positive classes represented 8 items strong); a verbatim port of the shipped rule pinned at its own commit; a tested only-quieter property between the shipped rule and the bench's exploratory variant; and a zero-spend reproduction check gating the run, which reproduced Study 6's numbers exactly before any new read was scored. On the production reader the sheet came back perfect: 8 of 8 detectable corrupted reads caught, 16 of 16 undetectable reads silent, 0 of 24 clean reads flagged, with zeros in the modulating-chart class where the pooled rule failed and the honest sharp-key class where the pre-refinement rule failed. The extra condition added at productionization cost nothing, and all four registered predictions held. About $1.60 of the $6 cap.

The Longest Loop Closes

The verdict delivered back to the app: nothing to change. The loop this closes is the program's longest, four hand-offs with each side checking the other: Study 5's anatomy found the corruption class, Study 3's correction described it, Stem Shovel shipped a heuristic, Study 6 measured it and suggested the refinement, Stem Shovel shipped the refinement and registered its own validation, and Study 6b confirms it on data it was not designed on: 100% hit, 100% silence, zero false alarms.