Methodology

A clear measurement contract. Every number has a source and a boundary.

Measurement contract

Mainnet · read-only review

Hadal compares tipped Beam HTTP delivery with untipped Solami RPC. Both share a 20,000-unit compute limit and configured priority fees, with no preflight and no client retries. Route and tip change together; this measures the combined treatment.

Assignment schedule: Loading…
Configured cadence: —
Beam tip: —
Priority fee: —
Observer region: Not declared
Protocol fingerprint: —

Probe sending remains gated by explicit wallet and spending authorization.

The JSON download contains public settings and archived protocols for recorded attempts. It excludes credentials, wallet files, and provider URLs.

Timing & sources

Reception ≠ execution

01SendBefore the HTTP broadcast

02AcceptedTransport acknowledgement

03ProcessedFirst live stream reception

04ConfirmedConfirmed stream reception

New probes store monotonic durations tied to the sending collector session. UTC wall timestamps remain display context. A clock adjustment cannot turn a measured duration into a wall-clock subtraction. After a collector restart, missing cross-session durations stay unknown.

Legacy records retain their wall-clock label. RPC recovery is confirmation evidence, not live latency. Historical signatures have unknown send and first-seen times unless this collector observed them. Chain block time has second precision.

Transport acceptance can arrive after a stream observation. Processed reception is not confirmed inclusion, and confirmed inclusion is not consensus finality. Node timing API ↗

Statistics & uncertainty

Inspect the dashboard

Nearest-rank percentiles & tail support

p50 and p95 use successful live gRPC receptions in the complete one-hour stored cohort. RPC recovery, execution failures, future timestamps and unavailable timing are excluded. Below 20 samples, nearest-rank p95 is the sample maximum. The small-tail warning below 100 samples is a display heuristic, not a precision guarantee.

Empirical cumulative distribution

The distribution view groups tied durations and plots the right-continuous fraction F(t) = count(duration ≤ t) / n. Its threshold means the fraction of observed live successes within that duration. It is not the chance that every submitted transaction lands by that deadline.

Wilson reference ranges

Nominal 95% Wilson ranges describe landed counts among settled attempts under a fixed-sample IID Bernoulli model. Empty samples remain unknown. Rolling samples, temporal dependence, and settled-only selection prevent interpreting this range as an anytime-valid winner test.

Pending-outcome bounds

For A attempts, L observed landings and P pending outcomes, final observed inclusion lies between L/A and (L+P)/A. Expired attempts remain in the denominator; execution failures count as landed. These are accounting bounds, not confidence intervals or a censoring model.

Timing provenance & protocol mixtures

Live coverage is valid live durations divided by successful inclusions. The app separately counts monotonic, legacy wall-clock, RPC recovery and invalid timing. Mixed protocols are flagged; changing treatment settings does not create one controlled comparison.

Route assignments

Persisted blocks

The default schedule randomizes the first route in each two-attempt block, then assigns the other route. A persisted transport failure still consumes its assignment. Restarting resumes an unfinished block only when the protocol matches. Optional alternating mode preserves strict Beam/RPC ordering.

The dashboard reports descriptive Beam-minus-RPC latency differences only for complete blocks with two successful live observations from the same monotonic clock session and a send gap no greater than three configured intervals. Incomplete, recovered, failed and long-gap pairs do not become fabricated matches. No confidence or causal winner follows from this summary.

Temporal blocking reduces one scheduling bias. Tip, observation coverage, geography, workload interactions and dependence remain relevant. A future inferential experiment needs an append-only cohort, an estimand and defensible assumptions. NIST randomized block designs ↗

Research applied

Reading library

Why Does My Transaction Fail?Separate execution failure from delivery failure.Brown, Cai & DasGuptaWilson ranges with explicit model assumptions.The Tail at ScaleDistribution, tail support, and inspectable slow observations.AlpenglowKeep client reception, confirmation, and finality distinct.Munzner · Nested ModelMatch trend, distribution and trace views to the task.Heer & Bostock · Graphical PerceptionCommon axes, explicit values and restrained chart encodings.Time-uniform confidence sequencesApplied guardrail: rolling Wilson ranges are not a continuous winner test.

Reading the limits

A missing Beam receipt is unknown, not a failed landing.
An expired attempt does not identify the validator responsible.
The leader of a confirmed slot supplies context, not proof of a drop.
Unavailable streams and a zero sample yield unknown performance.
Randomization does not remove the tip difference or establish independence.
Sequential inference, bandits and survival estimates stay unimplemented until their experiment and censoring assumptions are justified.