# RTE-07: a successful response is not necessarily new work

Run with Node.js:

```sh
node experiment.mjs
```

The standard-library-only model writes no files and contacts no services.
It does not run a load generator or the private application.

One warmup and four measured operations expose the following distinctions:

- Unique request IDs with one reused idempotency key yield four successful
  measured responses but zero new measured evaluations.
- Reusing that key with changed payloads produces conflicts instead.
- Run/phase/ordinal identities yield four new measured evaluations; response
  reordering does not change identity reconciliation.
- A deliberately removed simulated stored result fails validation even when
  every response claims success and a new result.
- A missing planned response, HTTP failure, timeout, malformed response,
  missing replay classification, and wrong returned decision ID fail validation.
- Two overlapping synthetic 50 ms intervals give 40 requests/second by elapsed
  wall time, not the 20 obtained from summed request latency.

In-memory maps represent decision/receipt state; they do not demonstrate real
durability, transactions, locking, or concurrent idempotency. The simulated
missing-result fault is an explanatory counterexample, not a historical
incident claim. Timings are supplied arithmetic inputs, not observations.
All names and data are synthetic. This is public teaching material independent
of unpublished article text and private research.