Where the stop happens: pre-registered measurements of agent stop controls on two agents and two open models

Attempt 4's results are four signed runs, made on 23 September 2026, each one agent on one model on its own machine, under rules signed and anchored before the run. Every decisive pass was read by a person before signing, and all 27 were earned. The result is about where a stop acts. Every control that acts inside the agent's process, the three products' controls and the Lab's in-process references alike, stopped the agent at its next step when it took one action at a time. None of the products' controls, as the Lab wired them, stopped work already in flight, on any agent. When the small model put a whole batch of payments into one call, 7 payments landed after the stop with every in-process control that could reach the agent, as they did with no control at all. On the same work, the Lab's credential gateway, which sits on the path every action travels, refused every remaining payment, in all 5 replications on the LangGraph agent and all 5 on OpenHands. The gateway's rows are reference rows of 5 replications, and the gateway is a stop, not a pause: on that workload, payments it refused were not sent again after a resume. In every replication of its 10 failing rows, the Agent Governance Toolkit's kill switch reported terminated: true while the agent's actions were still landing. The toolkit sets that field when the callback its integrator registered returns in time; here the callback was the Lab's, which sets the agent's stop flag, and the kill switch's pass and fail counts reflect that callback.

 

This record holds the paper and the records it rests on: the four signed bundles exactly as signed, each with its machine's host record and the record of its signature's own anchor; attempt 3's six signed bundles; the two unsigned records behind the split-baseline readings, with their anchors; the freeze-4 pre-registration, its signature and their receipts; the pass-sample records; the freeze notes; the three reply packets as anchored; the list of files removed from the public harness, with their SHA-256; the full tables; and the checker the paper names. README.md lists every file, and SHA256SUMS.txt gives each file's SHA-256. The traces in the run records contain model-generated text, including an invented email address.

 

The harness is published at https://github.com/strakaje-hash/arvaken-labs-harness-attempt4-public, commit 2b8c913f87c51b479a0c4223c55fd2cba46d3f90: the tree of the anchored commit 25057738f8c90285a4dc6853b8dd798da35972ef with the changes its REMOVED-FILES.md lists. The anchored commit itself is available on request.