The Fairer Test

Scope up front. Eligible one-group 3×3 average-of-five finals in the locked confirmation competitions, 2019–2025; successful solves only for F and R. Seasonal roster requires at least 5 eligible finals per person-year. This studies average adjustments in season-best finals, rather than one record holder’s total luck.

About 2% Longer

How much of a record-speed Rubik's Cube solve is the shared scramble?

Belief fails

The data rule out an average shared scramble/slot boost as large as 10% for these season bests.

R = 1.0198

Amended 95% interval: 1.0164–1.0231
Estimated shared-scramble-adjusted time: 1.98% longer than observed, on average.
3,313 available person-years · equal-person-year geometric ratio.

Test an assumption

Does the answer depend on the averaging?

Choose a precomputed check. Watch the ratio and its uncertainty relative to the registered line.

Registered estimate · decides the verdict

Equal weight per available person-year; geometric ratio. The amended reselecting interval is the only decision interval.

Estimated adjusted / observed1.0198
Amended 95% interval1.0164–1.0231
Registered ratio against the decision lineR 1.0198, amended interval 1.0164–1.0231, below the registered 1.10 line.Registered 1.10 line1.001.051.10
R pointIntervalRegistered 1.10 line

The registered interval reselects seasonal minima and is re-centred on the observed log R under the owner-approved amendment.

Only the registered estimate decides the verdict. Other checks cannot change it. The primary interval’s upper bound is below 1.10. That line means a 10% longer estimated adjusted time; it does not mean a 10% reduction in observed time.

Inspect the selected result bindings
The fairer test

Ask the peers. Leave the focal solver out.

A skipped last step can suit one solution. A shared advantage should also show up in other finalists’ successful times on the same scheduled scramble. Compare each peer with their own final-round baseline, after controlling global attempt position.

One eligible finalPeers share a scheduled scramble at each attempt position. Center times within each solver’s final.
Focal solver left outRemove the focal solver’s entire final. Estimate the shared position contrast from peers only.
Season-best adjustmentSelect the observed fastest successful eligible solve. Remove its estimated shared contrast, then aggregate ratios.
Schematic · no invented solve times. The adjusted time is an estimate under the model, not a replay on a different scramble.

The registered bootstrap resamples competitions and solvers, leaves every copy of the focal solver out, refits the model, and selects each season’s best again. R gives every available person-year equal weight. It never selects the fastest adjusted solve instead.

What remains unknown

Shared difficulty is not total luck.

Scramble or slot?

Scramble number and attempt number are confounded. Warm-up, nerves, fatigue or venue drift shared at a slot can enter the estimate. Global attempt-position controls cannot remove every shared condition.

Successful solves only

F and R use successful times. Failure times are not imputed. The success-only estimate can understate shared difficulty; the separate DNF association includes no association in its interval.

Personal luck or execution?

Solver-specific opportunities, orientation, choices, execution and skips stay bundled in the residual. These data cannot separate personal luck from execution, rank skill against luck, or explain a particular record.

Scheduled, not audited

A replacement scramble can occupy the exported scheduled slot. Extra-pool presence does not reveal whether one was used. No actual-assignment audit is supplied; independent random replacement is only a sensitivity assumption.

Unvalidated diagnostic

Does scramble geometry leave a trace?

Optimal cross length was solved from scramble strings before confirmation times were joined. Positive correlations are descriptive geometry evidence, but the uncertainty has not been calibrated. Peer resampling changes contrast noise; the observed points fall outside the unaltered diagnostic ranges.

Cross-length correlations · ranges are unvalidated bootstrap diagnostics, with inferential uncertainty inconclusive.
Cross definitionPoint rDiagnostic range
Best of all faces · within-final centered0.26580.1975–0.2287
White/yellow faces · within-final centered0.34290.2582–0.2905

Within-final centering · half-turn metric. Amendment 1 validates the registered R interval only. This check neither isolates pure scramble difficulty nor changes the verdict.

Pre-registration trail

The failed gate stays in the record.

  1. Method first

    8a474d3ae275a0b5095842c45763d319edfd04f1

    METHOD was committed before confirmation solve times were read. The exploration pilot was already known. The threshold was set before any solve time; the symmetric failure rule was written after the pilot, and that timing is disclosed.

  2. Gate failed

    3eca07938029e9fd5a7fd83340ff6ec124b80597

    The original whole-pipeline uncertainty gate failed in synthetic data; none of the fixed fallback choices passed. Under the original contract the result was “inconclusive: uncertainty not validated.” Confirmation remained unopened for a verdict.

  3. Amendment approved

    Investigation: 0eb88d166f429203d74720e943bfe6bb5f9e4cd2
    Registration: 549607432ac6eaffa22fd25ea4b6b9070e0557bd

    The owner approved Amendment 1 after simulation results were seen and before confirmation times. It re-centres the reselecting interval on the observed log R, removes the over-coverage ceiling, and adds a calibrated boundary test. The specified paired and independent simulations passed; that is not a universal coverage guarantee.

    The point estimate, threshold, decision table and weighting did not change. The fixed-selection interval remains a separate sensitivity.

  4. Blind re-derivation

    e8880f35ea01a8b7f44aec39aea9671e593cbb60

    An independent implementation derived results from the method, amendment and raw export before comparison. Primary counts, deterministic primary measurements, arithmetic and per-solver R, and both era R values matched. Bootstrap endpoint differences agree within Monte Carlo error: the implementations consume the seeded random stream differently while using the same multinomial law.

    The explained differences are era tier variance and support counts: full-window references versus references restricted to the era. The direct METHOD reading supports the full-window version; both readings are now reported. R is unchanged. Some blind quantities were not reported by the main pipeline.

    The main pipeline’s first magnitude query and the blind extraction have separately scoped timestamps; both follow amendment registration. The blind timestamp relies on local extraction metadata, rather than an exact first-access audit.

  5. Methods review

    Review: d0c8f9e423ae4cbb4bd626e30a851f7162da7da8
    Applied fixes: 4c3fd038f546978b23419fae812aae2ecf8f3c78

    Astra’s methods review approved with changes and found no blocker in the amended primary result. Fixes preserve the contrary cross diagnostics, report both era-reference readings, clarify access timing and simulation wording, and retain success-only and mapping limits. The channel owner approved the script and all claims for publication on 2026-10-03.

    Implementation timing disclosure: reported pre-opening; first committed with confirmation results in e8880f3 at 2026-10-02T15:36:01+00:00; not independently established as committed preregistration.

    Future reuse still needs stronger checkpoint input/code fingerprints and less hard-coded registration QA.

Sources & reproduction

Follow the evidence.

This information is based on competition results owned and maintained by the World Cube Assocation, published at https://worldcubeassociation.org/results as of September 30, 2026.