ANALYSIS REPRODUCTION - 0.2-A2

Do not trust W.
Rebuild it.

A leaderboard row cannot rest on a self-reported score. Profile 0.2-A2 starts from one audited row per environment, then reproduces the primary RMST, deterministic bootstrap interval, censoring envelope, influence, tail support, and exact-horizon accounting under one frozen algorithm.

Reproduce a claim ->Download templateOpen contract
1CANONICAL ESTIMATOR
10KBOOTSTRAP DRAWS
7REPRODUCED OUTPUTS
7HARD GATES
01 / CLAIM VERIFICATION

Input rows first.
Leaderboard second.

Every displayed aggregate must reconcile to the same environment-level endpoint table and upstream evidence commitments.

A1FROZEN ESTIMAND

One estimator, horizon, event, and tie rule.

REQUIRED
A2ENDPOINT MAPPING

Every exit becomes one declared disposition.

REQUIRED
A3UNIQUE UNITS

One hashed environment per analysis row.

REQUIRED
A4WILD SUPPORT

N≥20 and at least 10,000 resident hours.

REQUIRED
A5PRIMARY RESULT

Rebuild W and S(10K) exactly.

REQUIRED
A6UNCERTAINTY

Rebuild CI, bounds, influence, and support.

REQUIRED
A7INDEPENDENT AUDIT

Bind inputs, code, log, and assessor.

REQUIRED
02 / DETERMINISTIC PIPELINE

Same rows.
Same answer.

The bootstrap is stochastic in method but deterministic in verification: a frozen seed, PCG32 stream, high-word index mapping, and percentile interpolation define every resample.

PRIMARY + UNCERTAINTYrows → KM → RMST → W
rows × PCG32 → 95% CI

Any implementation may be used, but the official submission must reproduce the canonical outputs within 10⁻⁶.

03 / REQUIRED TRACE

Publish the result.
Publish its fragility.

W alone is insufficient. The same manifest carries censoring sensitivity, leave-one-environment-out influence, and observed support at the 10,000-hour tail.

PRIMARY

W and S(10K), with refusal when the horizon is unsupported.

UNCERTAINTY

10,000 environment-level percentile bootstrap draws and valid fraction.

ROBUSTNESS

Pessimistic and optimistic censoring bounds plus maximum leave-one-out shift.

SUPPORT

Pre-event risk set at 10K, exact-horizon rejections, retained completions, early exits, and endpoint totals.

04 / LOCAL REPRODUCER

Paste the claim.
Compute the truth.

The checker reruns every official calculation locally and compares the claimed values to the reproduced values without uploading the endpoint table.

LOCAL ANALYSIS REPRODUCERPROFILE / 0.2-A2
ANALYSIS REPRODUCTIONPASSEDAll seven analysis gates pass.
REPRODUCED W87.500
DETERMINISTIC 95% CI75.00-96.88
S(10K)83.3%
VALID BOOTSTRAPS100.0%
A1
FROZEN ESTIMAND

Estimator, horizon, event, horizon-boundary rule, bootstrap unit, PRNG, and support threshold are canonical.

PASS
A2
ENDPOINT MAPPING

Every disposition maps to a valid WILD event or censor with no terminal competing cause.

PASS
A3
UNIQUE ANALYSIS UNITS

Every row has one unique hashed environment identifier.

PASS
A4
WILD SUPPORT

The endpoint table contains at least 20 environments and 10,000 aggregate resident hours.

PASS
A5
PRIMARY + HORIZON REPRODUCTION

W, post-event S(10K), the pre-event 10K risk set, boundary rejections, and retained completions reproduce without extrapolation.

PASS
A6
UNCERTAINTY + ROBUSTNESS

The deterministic 95% CI, censoring bounds, influence, and tail support reproduce exactly.

PASS
A7
BOUND INDEPENDENT AUDIT

Endpoint table, upstream profiles, code, reproduction log, and independent assessor are bound.

PASS
REPRODUCTION != CERTIFICATION

A matching statistic still requires cohort integrity, exposure integrity, safety, preregistration, and independent source verification before registry admission.

OPEN PROTOCOL KIT ->