01 / fixed rule
Send once
One attempt. Then live with the uncertainty.
Try this first
Let the receiver forget.
Press Show outcome. Keep the seed, then change Key retention to 8 ticks. An old key expires; its delivery does not disappear. A late retry can now produce a second effect. Return retention to Whole run, then reduce Credits to 24. Bookkeeping protects uniqueness, but this model charges for it.
Delivered. Confirmed. Different numbers.
A missing receipt could mean the request vanished, the reply is slow, or the work finished and the reply vanished. Send once accepts the uncertainty. Retry oldest tries fresh jobs first, then the oldest unconfirmed job whose timeout has elapsed. Random retry uses the same eligibility rule with random selection. None learns.
Retry + key uses the same dispatcher as Retry oldest. Its receiver atomically stores a unique job key with the effect, then suppresses repeats while that key survives. This atomicity is an assumption, not a feature that arbitrary external services automatically provide.
No rule gets a free lunch.
| Rule | Delivered | Duplicates | Confirmed |
|---|---|---|---|
| Send once | 14.016 | 0 | 8.375 |
| Retry oldest | 18 | 13.844 | 17.844 |
| Random retry | 18 | 14.172 | 17.844 |
| Retry + key | 17.328 | 0 | 13.797 |
Retained keys eliminate duplicate effects here, but the extra credit per attempt leaves some work undelivered. At a 24-credit ceiling, the keyed rule averages 9.5 delivered; oldest retry averages 15.984. At 100 credits, both deliver all 18 in every sampled default-channel world.
Eight-tick key retention produces 5.281 duplicate effects on average. Even a loss-free channel gives oldest retry 2.484 duplicates: a reply can take 12 ticks to return, longer than the 8-tick timeout. Try a longer timeout before deciding that every silence is a failure.
What you can change
- Seed and credits
- The seed reproduces a world; credits cap all attempts. Keyed requests consume two each, other requests one. Unused credits are not spent automatically.
- Loss and delay
- Request loss happens before delivery. Receipt loss happens after arrival, including a suppressed repeat. Delay is uniformly 1 to the selected maximum on each leg.
- Retry after
- Ticks since the last send before that unconfirmed job becomes eligible. New jobs have priority. Long timeouts can leave work unresolved at the deadline.
- Key retention and forgetting
- Retention starts when the first effect is recorded; suppressing a repeat does not refresh it. Forget keys clears the receiver ledger at tick 40, preserving every effect and all sender memory. Both controls affect only the keyed receiver.
Where this experiment stops
All rules see only local send history and acknowledgements, with matched world draws per job and attempt. They share a dispatch horizon and credit ceiling, not equal attempts or measured compute. Processing at the receiver is instantaneous; there is no congestion, malicious sender, key collision, changing payload or real service integration.
All 640 seeded worlds and 2,560 policy outcomes are retained across ten predeclared settings, including total request loss, total receipt loss, short retention and ledger reset. These are invented rules, not measurements of a provider, organism, LLM or trading system. There is no composite score hiding the tradeoff.
Background: the AWS Builders’ Library on safe retries explains request identifiers and retry semantics. This model is an original teaching simplification.
Method & parameter guide ↓Source package ↓All batch results ↓Read the model
Previously: One Door, Many Hands · Explore the eight small worlds.