Work

BRIDGE: An Acceptance Gate to Block False Improvements in Model Updates, with a Retrospective Clinical Evaluation

Year: 2026

Type: preprint

Abstract: Class-balanced metrics for more than two classes. bridge-gate 1.1.1: --metric bacc and --metric macroF1 now apply to tasks with more than two classes, in addition to weighted F1. A model is fitted once and judged under several metrics through protocol.run_metrics. Tested on the eight multiclass datasets of collections (D) and (E), 480 decisions (MINI-54): every weighted-F1 decision reproduces the stored one, and BRIDGE meets the pass criterion under all three metrics (weighted F1: 216 accepted, 0 false improvements; balanced accuracy: 253 / 1, bound 2.93%; macro-F1: 248 / 0, bound 1.68%). Run script, per-trial outputs, analysis and manifest are in experiments/. All earlier decisions, stored outputs and reported numbers are unchanged. Every run manifest verifies (experiments/verify_manifests.py), and reproduce_paper.sh reproduces the main table.

Source: arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29515

Author: Sunaryono, Dwi

Institution: Institut Teknologi Sepuluh Nopember

Language: English

Topic: Machine Learning and Data Classification

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funders: Institut Teknologi Sepuluh Nopember, Lembaga Pengelola Dana Pendidikan, Direktorat Riset dan Pengabdian Masyarakat

Awards: G3489933243, G3855682619, G4708790182, G8443298323

OpenAlex ID: W7223139003

Cites: 3

Cited by:

Related to: 3