Only an authorized decision maker's permanent, uncoerced, confirmed removal request becomes the primary rejection event.
Before survival analysis,
decide what happened.
W depends on one difficult distinction: did the authorized human permanently reject the robot, or did observation end for another reason? Profile 0.2-J1 makes every terminal decision complete, blinded, independently reviewed, and reproducible before it can affect the score.
The label controls
the survival curve.
A rejection is an event. Completion and unrelated exit are censoring. Safety, developer, and consent terminations are competing causes that can block rankability. Those are mathematical choices, not editorial ones.
Every environment receives distinct independent reviews without the aggregate score, manufacturer identity, or other outcomes.
A third blinded review resolves initial disagreement by majority; unresolved outcomes never enter W.
Safety removal, developer withdrawal, and consent or privacy withdrawal cannot be relabeled as harmless censoring.
Evidence first.
Outcome second.
The decision clock starts at signed activation. A rejection occurs at the first unambiguous permanent request. Endpoint events win timestamp ties with censoring, and every decision remains bound to its exposure record and signed evidence bundle.
Authorized + uncoerced + confirmed + independently adjudicated
NO IMPUTATION · NO POST-HOC RELABELRebuild every label.
Then count.
The browser verifier checks the trigger-to-disposition map, request authority, review independence and blinding, tie-break logic, target exposure, evidence bindings, and all aggregate counts.
Paste aggregate decision evidence only. The verifier never uploads participant statements or household identifiers.
An unresolved or strategically relabeled endpoint is not missing paperwork. It changes the estimand and blocks ranked WANTED WILD admission.
OPEN ANALYSIS REPRODUCER →