BABEL v2
Independent execution-failure evaluation
The claim is unverified; advancement requires the frozen partner evaluation and held-out gate.
Abstract
BABEL v2 asks whether structural signals predict execution failure on independently authored workflows. Earlier benchmark and annotation-derived results do not answer that question. The current lane is therefore governed by a frozen baseline preregistration, partner-authorship gates, held-out evaluation, and an explicit kill threshold.
Evidence and record
Do structural signals predict execution failure on independently authored workflows?
No predictive result is claimed. The current result is a testable protocol with an external gate and a declared stop condition.
Earlier internal and annotation-derived results do not establish execution-failure prediction.
Failure to improve held-out AUROC by 0.05 with a positive bootstrap interval, or failure of the partner gates, ends the predictive claim.
not requested
- supersedes BABEL v1