Define sites before outcomes exist.
REQUIREDOne score.
More than one site.
A pooled survival estimate can look stable while one field operation supplies nearly every household. WANTED WILD now requires at least three substantive sites, caps any site at half the cohort, reproduces W within each site, and removes each site in turn.
Spread the evidence.
Expose the dependence.
Topology is a certification gate. Site-specific outcomes are public diagnostics and never become a second score or tiebreaker.
At least three substantive sites.
REQUIREDOne opaque site per environment.
REQUIREDNever replace or reweight W.
REQUIREDReport W or null by site.
REQUIREDRemove each site in turn.
REQUIREDBind assignments and reproduction.
REQUIREDPartition the rows.
Do not move the goalposts.
Every site uses the same A2 event, censoring, tie, horizon, and refusal rules. Unsupported site tails remain null. The official pooled W stays the unweighted environment-level estimate.
Δsite-out = maxs|W-s − W|
Passing limits concentration. It does not impose a performance threshold on site-specific W or its range.
Show where
the score moves.
Opaque site IDs protect participants while preserving every site count, share, identifiable estimate, unsupported tail, and site-exclusion result.
Site count, smallest N, and maximum cohort share.
W and S(10K), or null when the tail is unsupported.
Complete leave-one-site-out table and maximum absolute shift.
No geographic, dwelling, cultural, or population generalization claim.
Bind every site.
Recompute every view.
The verifier runs locally against opaque site assignments and endpoint rows. It checks prospective design, topology, pooled identity, site results, site-exclusion sensitivity, digests, and independent attestation.
Site definition and dominance limits were frozen before benchmark activity.
At least three substantive sites contribute without one site exceeding half the cohort.
Every unique environment maps to one opaque site and a rankable A2 endpoint.
The pooled W and S(10K) reproduce without reweighting or score substitution.
Every site-specific W is reported when identifiable and null otherwise.
Every site exclusion is recomputed and maximum pooled-score sensitivity is disclosed.
Controlled site assignments and the reproduction log are digest-bound and independently attested.
W alone orders eligible WILD entries. Site range, dominance, and leave-one-site-out shift are mandatory public diagnostics and forbidden tiebreakers.
OPEN LEADERBOARD ->