AI Science Tracker
EN

Better classifier discrimination did not improve a defensive gateway’s recall

Iskenderun Technical University · 23 Oct 2026 · computer science

Discovery

Better classifier discrimination did not improve a defensive gateway’s recall

Source check: Source checked. Not independent scientific confirmation.

Finding: In a replay of 25,000 labeled prompts, repairing a learned classifier increased its stand-alone AUC from 0.392 to 0.855, while recall for the complete defensive gateway fell from 80.67% to 80.34%. The repair and substitution experiments measured divergence between component and system performance.

Technology: DeBERTa-v3-small; BARBICAN defensive classification

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
A quantized DeBERTa-v3-small classifier supplied learned verdicts in a layered defensive gateway.
Test and result
The same benchmark corpus was evaluated under original and repaired conditions, with substitutions for additional comparisons.
Limitation / still unknown
The labels and operating conditions are benchmark specific and do not estimate field incidents. The recall drop occurred entirely in the synthetic subset, with no recall change across the four public corpora. A stronger substituted classifier elsewhere in the paper slightly improved recall, so this repair result does not establish a general inverse relationship between component quality and system performance.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.33198v1 · https://arxiv.org/html/2610.33198v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 defensive comparison with classifier and system outcomes.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.