Heimdall ResearchSearch
Menu

One effect per intent, and an honest answer

Duplicated delivery is normal in queues and agent teams. When every effect passes one executor that records before it acts, does a duplicate still become a second effect, and what does the requester learn when the answer is lost?

Where this stands

What the evidence shows 3

  • Two agents given one refund ticket at once refunded twice in 50 of 50 unguarded pairs, though each first read the payment's refunds; through the guard's payment-keyed gateway, 0 of 50. A second vendor's agent gave 25 of 30 against 0 of 30. DUP-02, DUP-06 re-checked.
  • Keyed on an operator's mandate, the gateway let both lawful refunds through in 20 of 20 pairs and still stopped a duplicate under one mandate (0 of 20 refunded twice). DUP-08 re-checked.
  • When the answer to an effectful call was lost in integration tests, the provider held 1 payment and the retry was refused as undetermined; when the request was cut before the server, 0 payments were made and an operator act could resolve the reservation. ARC-19 working note.

What did not hold, or was not unique 2

  • Within one call, existing clients already kept one effect: in 70 client-and-fault combinations of three Stripe SDKs on a simulator, no retry made a second effect. The default Python client, though, reported an error for a refund that existed. ARC-21 re-checked.
  • A prediction about our race model failed: harm fell below half of pairs at a 2 s start offset (9 of 20), not at 3 to 3.5 s as preregistered. It concerns the model of the race, not the guard. DUP-03 re-checked.

What remains open 3

  • Two judgements stay with people. With mandate keying the guard enforces the issuer's count, so a duplicate authorized twice passed (18 of 20 pairs); with an answer dropped, the gateway made no lawful refund and left the outcome unknown for an operator. DUP-08, ARC-23 re-checked.
  • The two-agent result has been measured only on refunds, on a simulator. The once-only check has not been isolated at the layer's executor, and whether it is worth more beside the guard's other checks at one boundary is untested. DUP-02, DUP-06 re-checked.
  • Replies are no record of effects: in a pilot, 19 of 19 agents informed that their answer was a replay still told the customer they had made the refund. No rate exists yet for replies after a refusal; a coding scheme was never run. DUP-01 re-checked; OWN-06 working note.

Being studied now 2

13 of the 24 studies in this direction are not yet published: most rest on records not yet in the repository or on a stand whose source material cannot be published, and others are still running or were never run.

Every published study on this question

Other questions