A guarantee is conditional on a model
A failure model specifies how participants and communication may misbehave. Crash-stop means a process stops and does not return. Crash-recovery permits return, with explicit assumptions about retained state. Byzantine behaviour can include arbitrary or conflicting messages. A protocol proved safe for one model is not automatically safe for another. Include persistent storage, correlated failures and communication partitions in the assumptions. A useful model is concrete enough to explain the guarantee being offered and the situations outside that guarantee; merely saying the design is fault tolerant leaves too much unspecified.
A timeout is suspicion, not proof
Suppose A requests work from B and hears nothing before a deadline. B may have stopped, a path may be partitioned, messages may be delayed, or B may still be computing. In an asynchronous model with no known message-delay bound, a finite silence cannot conclusively distinguish a crash from a slow execution. Timeouts are useful triggers for action but require a design that tolerates mistaken suspicion. Reducing the timeout can improve reaction time while increasing false suspicions. Record the observation separately from the operational interpretation: no reply is evidence, while permanently failed is a hypothesis.
Timing assumptions are another dimension
Timing assumptions form a separate dimension. A synchronous model uses known bounds; partial synchrony introduces bounds that hold under stated conditions, such as after an unknown stabilization time. That does not specify whether a participant can lie. A crash-tolerant consensus protocol can require sufficiently stable communication and reachable members for progress. Distinguish safety, which excludes forbidden outcomes, from liveness, which concerns eventual progress. Waiting indefinitely can preserve safety while failing to make progress. Conversely, choosing an answer quickly does not establish safety if the protocol has ignored conditions needed to avoid conflicting decisions.
Test examples do not prove the whole model
Inject process termination, restart, disconnection and delay separately, and state which property each experiment checks. A successful crash test does not establish tolerance of arbitrary Byzantine behaviour or every combination of faults. During leader replacement, consider the old leader resuming after a long pause; fencing can restrict its ability to perform stale writes. Avoid relying on a belief that the suspected process has disappeared forever. Preserve invariants when it returns. Report test results only for conditions actually exercised, and compare operational reality with the failure and timing assumptions used in the protocol argument.