Preprint / scholarly article

Observable-Only AI Safety from Public Data: Robust Bottleneck Diagnosis with Auditable No-Meta Dynamic Programming, Anytime Confidence Sequences, and Dynamic IQC

K. Takahashi

Published
DOI
10.5281/zenodo.18615875

Full text PDF (Zenodo)

Abstract

The preprint presents an observable-only AI safety framework for robust bottleneck diagnosis from public data, combining no-meta dynamic programming, partial identification, anytime confidence sequences, and dynamic IQC to produce auditable interval diagnostics with fail-closed replay contracts.

Keywords

  • observable-only AI safety
  • public data
  • robust bottleneck diagnosis
  • no-meta governance
  • dynamic programming
  • anytime confidence sequences
  • e-processes
  • partial identification
  • dynamic IQC
  • deterministic replay
  • auditable diagnostics

Identifiers and source records