Preprint / scholarly article

Multi-Constraint Certified Bottleneck Estimator for Large-Scale AI Training

K. Takahashi

Published
DOI
10.5281/zenodo.18057956

Full text PDF (Zenodo)

Abstract

The preprint introduces MCCBE, a telemetry-only and fail-closed auditing framework that certifies conservative bottleneck floors across tail latency, integrity gaps, I/O, network, and power constraints for large-scale AI training under incomplete logs.

Keywords

  • bottleneck estimator
  • AI training
  • telemetry-only
  • fail-closed auditing
  • tail latency
  • integrity gaps
  • I/O limits
  • network limits
  • power caps
  • certified floors
  • resource ledgers

Identifiers and source records