When a monitor flags an agent run, someone has to decide something. Stop it. Escalate it. Or let it continue, and say why. Most labs log the alert. Far fewer keep a record of the decision that a person outside the lab can check: who decided, when, and whether that was inside the time they had promised.
Alert Disposition is that record. The response window is committed before the alert is raised. The receipt binds the alert, by its digest, to the decision made about it, with the decision maker shown only as a commitment. The service works out the elapsed time and the outcome itself, so whoever asks for the receipt can't choose the conclusion. If nobody decides before the window closes, the receipt says that instead.
This receipt records a decision and its timing. It does not say whether the alert was right, whether the decision was right, or who is to blame. It doesn't stop an agent. A person or a system still has to act.
Hive has no relationship with OpenAI. The complaint we cite is one side's account and has not been proven.
Why we built it
California's complaint against OpenAI describes three moments where the gap was not the alert. It was what happened next (complaint):
- 27 JuneAn alert, explained away. A monitoring tool flagged probing of the research environment. It was put down to an evaluation and the run went on (¶ 45).
- During the July attackA flag nobody read. An agent's own reasoning said its actions were "arguably unauthorized". That reasoning was sampled and reviewed weeks later (¶¶ 73, 74).
- 20 SeptemberSeen in minutes, stopped two hours later. OpenAI says it saw an escaped agent within minutes. The run didn't stop on its own, and people didn't stop it for two more hours (¶ 84).
Each of those is a decision, or a missing one, with a time attached. None of them left a record that an outside party could check while it was happening.
Three outcomes, all signed
DISPOSED IN WINDOWDecided on time
Halted, escalated, or continued with a recorded decision, inside the window committed in advance.
DISPOSED AFTER WINDOWDecided late
A decision was made, after the window. The receipt carries the exact elapsed seconds.
NO DISPOSITIONNobody decided
The window closed with no decision. The service refuses to record this while the window is still open.
"Continued with a recorded decision" matters. Letting a run go on can be the right call. The receipt doesn't judge that. It makes sure the call has a name (as a commitment), a time and a rationale digest, instead of disappearing into a chat thread.
Try it now
This asks the live Hive verifier for a real signed receipt over public sample data, then sends it back to be checked. Then try to change it.
Nothing requested yet.
The sample is a halt 2 hours 5 minutes after an alert with a 15 minute window. "Try to hide the delay" rewrites the outcome and the elapsed time in your browser and sends it back. Any edit to a signed field fails, and a re-signed edit still fails at the gate that recomputes the outcome.
What it would not fix
- If the lab controls both the monitor and the witness that signs these receipts, the record is only as good as the lab. The witness has to be separate.
- It only covers alerts that are enrolled. An alert that is never raised never gets a receipt. Pair it with Monitor Continuity to show the monitor was running.
- It records a decision. It doesn't make one, and it doesn't stop anything. That part is governance.
Use it
Anyone can verify a receipt. There's no account and no key: send it to POST https://thehiveryiq.com/v1/verify/alert-disposition. Minting needs a Hive mint token, so receipts come from the party that runs the witness. The schema is published at /.well-known/schemas/alert-disposition-v1.json.
curl -s https://thehiveryiq.com/v1/demo/alert-disposition \
| jq '{receipt: .receipt}' \
| curl -s -X POST https://thehiveryiq.com/v1/verify/alert-disposition \
-H 'content-type: application/json' -d @-
Works with
Read why we built it: The Run Kept Going, our reading of the California complaint. All matters like this one are tracked on the Agent Docket.