Work

Certified Benchmarks as Honest Evaluation Instruments: Measured Failures of Small Models on Unique-Solution Puzzles

Year: 2026

Type: preprint

Abstract: The Behnamian-collaboration preprint: certified unique-solution puzzles as the evaluation instrument against which honest failure is measurable. Measured confabulation across a small-model fleet (0.3%-2.7% grounding-gate acceptance; gate-assisted 1.3%), a three-condition ablation with irreducible variability (4/6, 3/6, 5/6), a contamination audit of the estate's own corpus (8/16 rows, fixed and measured), and the continuous-certification contribution (stratified re-proof every commit, full 242-certificate bank weekly, fail-closed).

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.24043622

Author: Lutar, Stephen P.

Language: English

Topic: Logic, Reasoning, and Knowledge

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7225463473

Cites:

Cited by:

Related to: 10