No experiment closed this week. Several systems did become less willing to confuse producing an answer with proving one.
The same bug, wearing different hats
Across the lab, the failures looked unrelated. One system could produce a useful conclusion without keeping its sources close enough. Another could repeat work without knowing whether the repetition was a retry. Another could act before its authority had been made sufficiently boring.
The common defect was not intelligence. It was the distance between an output and the conditions that made the output trustworthy.
What moved
Evidence and interpretation became easier to tell apart. Interrupted work became easier to resume without accidentally performing it twice. Agent work acquired tighter boundaries around where it may run and what it may carry home.
Elsewhere, paper results stopped borrowing the wardrobe of live ones, and an independent review found a failure that self-review had politely overlooked.
None of those changes makes for a single dramatic demo. Together, they make the systems harder to impress with their own output.
What did not move
No new portability claim was proved. No trading result became a profitability claim. No development milestone became a public product launch. No private machinery became public merely because it had a productive week.
The unresolved work remains unresolved. This is a field report, not an attempt to sneak several side quests through the experiment ledger in a large coat.
The late part
Friday happened. The update did not.
So this one ships on Sunday, visibly late and otherwise intact. A public record is more useful when it includes the missed cadence than when it quietly repairs the calendar.