Negative results and failure-envelope mapping

A success predicate may require correct output, bounded resource use, and a latency or recovery target simultaneously. Specify each requirement and the measurement window. Distinguish a supported configuration that failed from one that could not be executed because a necessary feature was absent. Record both, with their different meanings. A timeout is not a successful request with conveniently missing latency, and a crash cannot simply vanish from the denominator when the claim concerns completed work under a service contract.

Preserve attempted conditions, run identifiers, logs, stop reasons, and partial observations. A timed-out duration is often right-censored: completion would exceed the timeout if the operation eventually completed under that setup, but its exact completion time is unknown. Report a censor-aware analysis or a declared failure metric instead of assigning a fabricated exact latency. If an experiment is unsafe or unaffordable beyond a limit, mark that untested region explicitly. Missing evidence is different from evidence of either success or failure.

Choose stress axes that follow the explanation: offered load for queueing, hot-key skew for partition contention, state size for recovery, or message loss for a coordination protocol. Use a coarse sweep to find transitions and targeted follow-up points to investigate them. Adaptive exploration is legitimate when reported as such; do not describe its selected outcomes as an unbiased random sample of deployments. Repeat boundary conditions because variability can turn a sharp plotted edge into a probabilistic success region.

A deterministic fluid backlog provides a deliberately simple example. With initial backlog q₀ and constant service faster than arrival, the drain time is q₀ divided by excess service capacity. At equal rates the backlog does not drain, and above service it grows. This is not the waiting-time distribution of a stochastic queue. It isolates a resource balance that can motivate a test, while real arrivals, burstiness, scheduling, and finite buffers can produce different failures even below average capacity.

A failure to establish improvement is not proof that no improvement exists. An imprecise interval can leave useful and harmful effects both plausible. If practical equivalence matters, choose an equivalence margin and an appropriate analysis instead of interpreting a nonsignificant difference as equality. Distinguish an implementation bug, an invalid baseline, and a mechanism that genuinely loses in a valid regime. Publish the unfavorable cell alongside its resource and correctness context so readers can decide which explanation remains credible.

The queue instrument makes two fresh fluid experiments explicit. Reset, set both λ=2 and μ=3, predict the initial four jobs drain after eight Steps, then Step eight times. Separately Reset, set both λ=4 and μ=3, predict backlog eight after eight Steps, then Step eight times. Each Step advances 0.5 seconds, and each run starts with backlog four at time zero. This fixed clipped-linear model provides no arrival randomness, tail-latency samples, or production stability certificate; it demonstrates how a failed load condition can bound a narrow model claim.