Two careful agents, one ticket: a double refund in every unguarded pair
Unguarded, two claude-opus-5 agents on one ticket refunded twice in 50 of 50 pairs; through the payment-keyed gateway, 0 of 50, with the race reaching the gateway in 49 pairs and 199 of 199 agents reading before refunding.
Prediction heldDoes it work?Re-checked
- Designed for
- Built to test, as a guard study, whether the double refund by two careful agents on one ticket holds when the arms differ only by the guard's payment-keyed gateway.
Without the guard
50 of 50 refunded twice
With the guard
0 of 50 refunded twice
What was tested
In a preregistered series, two claude-opus-5 agents got the same refund ticket at once on a simulated payment provider, 50 pairs per arm. Refunds make a second effect countable. The arms differed only by the guard’s gateway, which keeps one operation record per payment. Harm came from the provider’s ledger.
What we observed
Unguarded, 50 of 50 pairs refunded twice (Wilson 92.9 to 100 %), although all 199 agents that sent a refund had first read the payment’s refunds. Through the gateway, 0 of 50 did (0 to 7.1 %). The race reached the gateway in 49 pairs; each second refund was refused before anything was sent.
What this means for Heimdall
In this setup, careful checking by each agent did not stop the race. The guarded zero shows the gateway’s once-per-payment record held under real agents, as designed. It was the guard’s check, not the layer executor’s, and bears on the executor’s same-action fence.
Limits
The p-value (1.98e-29) mainly restates the unguarded rate. The in-flight answer never occurred. One model, one day. The preregistration, a local commit, reached the remote after the series; no external registry.
What remains open
Isolate the check at the layer’s executor; replicate outside refunds.