The recorded cases behind See the Proof, one post per case and the catalog that frames them. Local, authenticated results from September 2026; not a production acceptance.

Six recorded proof-of-concept cases from September 2026, stated at their real size: three carry the main story, and three are narrower checks.
A proof page that states its evidence boundary first (recorded, local, not re-run, not a production acceptance) makes every claim on it easier to inspect.
Recorded 15 Sep 2026, local authenticated PoC. A promoted request waited as a held task. A human authorized it. Only then was it claimed, processed, and completed through Gate Zero. Authorization was a precondition, not a postmortem.
Put the human gate before the model call, not in the incident channel after. A held task that cannot start is cheaper than a completed task you have to unwind.
Agent reliability is not only a model problem. It is also an execution-order problem: establish authority, dispatch within the authorized lane, then produce a reviewable receipt.
Recorded 15 Sep 2026. An operator approved a correction. A later Forge run applied it. The receipt names the same correction id as both active and rendered. The fix was not a chat comment. It was an input to the next run.
Much of what gets called memory is a transcript. A transcript does not know which line was approved; a receipt that cites the correction id does. That is the difference between a note and a governed change.
The expensive failure is often not the first wrong answer but the second run that never saw the approved correction. If the fix is absent from the next receipt, the governance loop remains unproven.
Recorded the same day. A held task accepted exactly one approved delta at its checkpoint and finished by reusing the held response. No second model dispatch. One change. No extra generation.
Dispatch, review, and redispatch start another generation for one approved change. One recorded checkpoint applied one approved delta to a held response without redispatching the model.
Feedback in a chat thread is not proof that a change reached execution. Two recorded cases show approved corrections landing with an inspectable receipt.
Manual gap-to-function: 31 function records discovered in the PoC, canonically registered, bound to their implementation, independently read back, ending at Candidate 31.
A presenter replay passed nine recorded checks and a separate browser sequence passed its own. That is a bounded check suite, worth reporting as exactly that, not production traffic.
Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, partly for inadequate risk controls. This is what those controls look like when they come before the action.
Recorded proof, illustration, live read view and explanation: labeling each one keeps curated, current and proposed work from passing as executed.