Benchmark methodology

The suite compares operations that have a close native equivalent in UseCase Forge, Bloc, Riverpod, MobX, and Redux. It does not assign an overall winner and does not treat architecture, guarantees, ergonomics, adoption, and raw operation cost as one score.

Locked environment

The reviewed full runs used Dart 3.12.2 on macOS arm64 and Apple M1, with ten samples after a 750 ms warm-up per case. JIT and compiled AOT ran in dedicated processes. Package versions were locked: Bloc 9.2.1, Riverpod 3.4.2, MobX 2.6.0, Redux 5.0.0, and benchmark_harness 2.4.0.

Power metadata states AC power, Low Power Mode off, and battery reported discharging. That apparent combination is retained exactly as captured rather than normalized after the run.

Measurement boundary

• every sample creates a fresh runtime and performs preparation outside the timed region;
• Stopwatch measures the workload; disposal and result serialization are excluded;
• medians, interpolated p95, min, max, mean, population deviation, and coefficient of variation are retained;
• concurrent rows report batch latency normalized per operation, while raw batch ticks remain in JSON;
• N/A means the selected native primitive has no sufficiently close operation;
• allocation counts and GC pauses were not measured, so explanations based on them remain hypotheses.

Important limitation

The frozen metadata identifies the core as 0.1.0-dev.1 (path: ../..) and records the Git commit as unavailable. The runs describe the reviewed local candidates that led to the published preview line, but they are not a cryptographic measurement of the later 0.1.0-dev.2 pub.dev archive. A new release claim should be backed by a fresh hosted-dependency run with a recorded commit.

Download the evidence

• Complete methodology
• Scenario matrix
• Locked dependency graph
• Artifact checksum manifest

Raw JSON retains every sample; CSV provides reviewable summaries. The performance results page links the current reviewed pair directly.