Wilson score intervals
A reference range for the settled landing rate. Perfect small samples keep a wide range; no attempts means unknown.
Nominal 95% · fixed-sample IID model · no live winner testResearch & visual lab
The papers, methods and visual thinking behind a more useful landing observatory.
Design thesis
See the result. Understand the limits. Inspect the evidence.
Eight closely related readings, explicit assumptions, and algorithms tested on edge cases and connected to real collector data.
- Research readings
- 08
- Visual directions
- 06
01 / FROM RESEARCH TO PRODUCT
Make uncertainty visible.
These methods are implemented locally. The actual result still depends on collecting valid mainnet evidence.
Route cards ↗
Empirical latency distributions
Right-continuous step curves preserve repeated durations. A threshold control asks how much of the observed-success cohort confirmed within a selected time.
Successful live gRPC only · full one-hour cohort · no smoothingPending-cohort bounds
Landings divided by all attempts form the lower bound. Adding pending attempts to possible landings forms the upper bound.
Partial outcome accounting · not a confidence interval · no validator blameProvenance and tail inspection
Show how much successful inclusion evidence has valid live latency. Keep RPC recovery and bad timing outside that cohort. Trace the slowest recorded observations.
Explicit coverage · sparse-tail warning · source-linked records02 / A FOCUSED READING LIBRARY
The work this builds on.
Expand a reading for the product decision and review depth. A source’s result is not a current measurement from Hadal.
Solana Alpenglow Consensus
Votor and Rotor address consensus and block dissemination. Their latency model begins at a different stage from a client transaction send.
Measurement boundaryWhat changes in Hadal
Keep processed reception, confirmed inclusion and consensus finality separate. Hadal does not call its send-to-gRPC clock an Alpenglow finality measurement.
Reviewed v1.1 abstract, overview and latency/simulation sections. Formal safety proofs were not independently verified.Why Does My Transaction Fail?
The empirical study analyzes recorded execution failures, with error categories and temporal patterns. Those records already reached the chain.
AppliedWhat changes in Hadal
Show execution failure as a landed outcome. Separate successful execution, failed execution, expiry and pending evidence; avoid turning on-chain failures into delivery failures.
Reviewed abstract, background/definitions, study findings and threats to validity. Its 2023–2024 historical failure rates are not presented as current network rates.Transaction Latency on Solana
The study compares swQoS, fee and tip groups using a recent-blockhash time proxy. It acknowledges geographical and provider effects outside its controls.
Experimental designWhat changes in Hadal
Prefer our own observed send clock. Label Beam HTTP + tip versus untipped Solami RPC explicitly; the difference does not isolate routing alone. Do not reuse their latency figures as ours.
Reviewed methodology, inclusion distribution and routing comparisons. This is operator research, not a randomized trial or an independent Beam benchmark.Interval Estimation for a Binomial Proportion
The authors compare interval coverage and recommend alternatives to the standard Wald interval, including Wilson for small samples.
Applied · WilsonWhat changes in Hadal
Calculate a nominal 95% Wilson reference range for settled landing rates. Empty cohorts remain unknown; perfect tiny samples retain uncertainty. Mark the fixed-sample IID assumptions.
Reviewed indexed abstract and scanned primary-report recommendation pages. The linked 1999 report underlies the 2001 Statistical Science article; scanned pages were inspected visually.The Tail at Scale
Distributed services can feel unreliable because of rare slow responses, even when typical latency is low. The paper discusses mitigation for replicated service requests.
Applied · tail inspectionWhat changes in Hadal
Add an ECDF and inspectable slowest observations. Do not hedge or duplicate the probe broadcasts: that changes the measured treatment and makes attribution ambiguous.
Reviewed latency-variability discussion and hedged/tied-request sections of the primary PDF. Applying the lesson to transaction measurement is our inference, not the paper’s Solana claim.Time-uniform Confidence Sequences
Confidence sequences support inference across repeated looks under stated probabilistic assumptions. Ordinary fixed-sample intervals do not provide that guarantee.
Deferred · experiment modeWhat changes in Hadal
No live “Beam wins” detector from rolling Wilson ranges. A future experiment needs an append-only cohort, stated estimand and assumptions before anytime-valid comparison is added.
Reviewed introduction, coverage definition and sequential-inference discussion. No stitched or mixture confidence-sequence implementation is claimed.A Nested Model for Visualization Design and Validation
Visualization choices should follow the domain task, data abstraction and encoding before optimizing the rendering algorithm.
Applied · visual architectureWhat changes in Hadal
Match view to question: a trend for when conditions changed, an ECDF for how much of the observed cohort was fast, outcome bars for accounting, and trace for source inspection.
Reviewed model, threats-to-validity framework and validation guidance. Our view choices are design inferences from the model.Crowdsourcing Graphical Perception
The study replicates spatial-encoding comparisons and examines area judgment, chart size and grid spacing.
Applied · precise comparisonWhat changes in Hadal
Use shared axes, explicit counts and restrained gridlines for quantitative comparisons. Flowing ribbons tell the marketing story; numerical decisions remain in precise views.
Reviewed primary PDF abstract, spatial-encoding experiment and rectangular-area results. Its user study is not a usability evaluation of Hadal.03 / VISUAL THINKING, NOT JUST CHART TYPES
Who to learn from.
Original concept sketches below illustrate the principles we would adopt. They are not screenshots of these creators’ work or plots of mainnet data.
ORIGINAL CONCEPT SKETCH
Executable visual thinking
Mike Bostock / Observable
Step curves, common scales and direct inspection. Use the official Plot and D3 examples to develop the interaction, with our own data semantics.
Visit the reference ↗ORIGINAL CONCEPT SKETCH
Choose the right question first
Tamara Munzner
Map each screen to a domain question before picking a chart. Precision and task coverage come before visual novelty.
Visit the reference ↗ORIGINAL CONCEPT SKETCH
A memorable visual signature
Nadieh Bremer / Visual Cinnamon
Study composition, annotation, colour restraint and purposeful movement. Use that care in the delivery illustration; preserve measurable axes in the instrument.
Visit the reference ↗ORIGINAL CONCEPT SKETCH
Context is part of the data
Giorgia Lupi
Keep missingness, collection method and evidence limits in the story. A signature is a record with context, not just a coloured dot.
Visit the reference ↗ORIGINAL CONCEPT SKETCH
Calm, consistent software
Linear design team
Reduce navigation weight, make actions predictable, and refine density. The workspace should let the evidence earn the attention.
Visit the reference ↗ORIGINAL CONCEPT SKETCH
Shipped UI references
Mobbin + Refero
Study real tables, navigation, empty states and toolbars. Use the two supplied dashboards as visual anchors; a pretty gallery is not evidence that a workflow works.
Visit the reference ↗A serious experiment needs proof.
A reproducible instrument should answer a builder’s question in one glance, then let them inspect the transaction and its sources. Research makes the methods inspectable; valid live observations supply the result.
- Establish the observation path.Verify processed and confirmed streams before collecting latency evidence. Keep source outages visible.
- Inspect the declared protocol.Balanced route blocks and archived public protocols are implemented locally. Fees, tip and workload are recorded; observer region remains undeclared. Decide what treatment is being compared before interpreting a result.
- Inspect timing provenance.New probes capture monotonic elapsed times within their sending session. Cross-session receptions remain excluded; legacy wall-clock records are labelled.
- Make inference match the question.Use an append-only experiment cohort and justified dependence assumptions before adding confidence sequences or a route-winning claim. Rolling summaries remain descriptive.
This lab is read-only. Live outcomes come from the observatory; concept sketches illustrate visual principles.