The race window by start offset, and the preregistered prediction that failed
The preregistered prediction failed: harm fell below half of pairs at a 2 s offset (9 of 20), not at 3 to 3.5 s. Harm followed the race in 51 of 51 pairs against 0 of 129, and 4 to 8 s spacing still left 3 of 80 harmed.
Prediction failedHow we measureRe-checked
- Designed for
- Built to measure, in the guard's duplicate-delivery series, how wide the double-refund window is in time between two agents' starts, against a preregistered prediction.
What was tested
A race model fitted to pilot timing predicted, for each start offset between two unguarded claude-opus-5 agents on one refund ticket, the share of pairs that would refund twice. Preregistered test: the share first falls below one half at 3 or 3.5 seconds. Nine unguarded and two guarded offsets of 1 to 8 seconds ran, 20 pairs each, on a simulator.
What we observed
The prediction failed: the share fell below one half at 2 seconds (9 of 20). Both model assumptions held: no trend in the second agent’s read time, and 0 of 440 agents refunded after seeing a refund. Harm occurred in all 51 pairs where both read an empty refund list, none of the other 129. At 4 to 8 seconds, 3 of 80 pairs were harmed; through the gateway, 0 of 40.
What this means for Heimdall
The failed prediction concerns our race model, not the guard; the window roughly tracked the agents’ read-to-write time, about 2.5 seconds that day. Spacing starts, a simpler defence, narrowed the race without closing it. This bears on the executor’s fence at the point of effect.
Limits
The 2-second cell sits below that day’s own timing, unexplained. One model, one day.
What remains open
Another model or day needs a new measurement.