# A timeout is an observation, not a crash certificate

This executable teaching example accompanies the [crash-stop definition](https://hesela.dev/terms/crash-stop/) and [historical fail-stop definition](https://hesela.dev/terms/fail-stop/).
It uses four constructed schedules, not device measurements, fleet data or a new impossibility proof.

## What the observer sees

A receives a heartbeat at logical tick 0, sends one probe at tick 5, and times out at tick 10 unless the reply has arrived. A reply at exactly tick 10 is processed first by convention. The reply payload is identical in every schedule and contains no sender timestamp.

| Constructed schedule | B crashes | B sends reply | A receives reply | A times out |
| --- | --- | --- | --- | --- |
| Crashed | 7 | Never | Never | 10 |
| Delayed | Never | 11 | 12 | 10 |
| In flight | 7 | 6 | 12 | 10 |
| Timely control | Never | 6 | 8 | Never |

A's local event sequence through tick 10 is identical in the first three rows. In the first row B has stopped; in the second it has not. At tick 12, the delayed and in-flight rows still have identical observer histories although only one sender remains running. A late delivery can establish an earlier send without certifying the sender's present state.

The times are arbitrary logical ticks with no mapping to milliseconds. Request delivery to B coincides with reply send in rows containing a reply. No maximum request or reply delay is assumed. The crashed row's probe may reach B after its crash; the example makes no recovery or reply obligation for that process. A never crashes.

## Reproduce

Requires Node.js with its built-in test runner; no third-party package.

```sh
node model.mjs --check
node --test model.test.mjs
```

Run `node model.mjs` to regenerate `results.json`. The tests cover equal observer prefixes, the in-flight reply, deadline ordering, invalid schedules and deterministic output. The full source schedules and observer projections are in the JSON.

## Scope and engineering use

The hidden crash time is used only by the verifier; A cannot read it. The model does not estimate a failure probability, choose a timeout, implement a detector class, prove FLP, or verify a production protocol. There is no random sampling, seed, empirical confidence interval or hardware benchmark.

For an actual storage service, record separately which process can stop, how a replacement is authorized, which state survives and how restart changes identity. A timeout action needs a safety protocol even when the suspicion is wrong. The example's local observations alone do not settle those contracts.

## Source context

The example is Hesela's own construction. It illustrates the separation of actual failures and detector histories in Chandra and Toueg, *Unreliable Failure Detectors for Reliable Distributed Systems*, JACM 43(2), 1996, sections 1 and 2.1-2.4:
https://www.cs.utexas.edu/~lorenzo/corsi/cs380d/papers/p225-chandra.pdf

For the different historical processor abstraction, consult Schlichting and Schneider (TOCS 1983, sections 2 and 3.1) and Schneider (TOCS 1984, section 1). Their source links and interpretation are in the fail-stop record. Third-party papers are not redistributed here.

Created 2026-11-20. Regeneration time is not a scientific observation date.
