Twelve parallel coding agents on a single laptop, with about $450 in API costs, ran the framework's submission to the British Council's BEA 2026 vocabulary difficulty shared task.
The submission that won the closed track of the BEA 2026 vocabulary-difficulty shared task almost contained its own answers. A $450 open-source framework called Glite ARF ran twelve parallel AI coding agents on a single laptop, caught four feature sets quietly leaking target signals into its own training data, stripped them out, and submitted a less flattering 0.802 RMSE. The honest score still won.
Glite ARF was posted to arXiv on 25 June 2026 (Glite ARF preprint). It is the authors' attempt to keep empirical AI research reproducible when the "researcher" is partly a fleet of large language models. The framing the authors propose is "verifier-driven research": the rules of the research process live in Python code that fails loudly when broken, rather than in prose instructions that the agents are only expected to follow.
The setup has three roles. A human researcher decides which hypotheses to test. Coding agents, named explicitly in the paper as Claude Code and Codex CLI, implement individual experiments inside a fixed structure. Deterministic Python verifier scripts enforce four process invariants: task isolation, immutability of completed work, a corrections overlay, and a materialized project overview (Glite ARF HTML version). When something drifts, the verifier does not write a comment. It throws an error, and the run is no longer reproducible.
That audit machinery is what the authors stress-tested against themselves. To benchmark the workflow, they used it to build their submission to the BEA 2026 vocabulary-difficulty shared task, a competition run by the British Council that asks systems to predict how hard a given word is for language learners (BEA 2026 shared task results).
Mid-run, the corrections overlay caught four engineered feature sets that had drifted into leaking target-difficulty labels into the training pipeline. The team stripped them out and resubmitted. The headline metric, root mean squared error (RMSE, where lower is better), moved from 0.609 to 0.802 (Glite ARF preprint). Without the verifier firing, the authors argue, they would have published a falsely good score and the field would have learned the lesson at peer review, or later.
That corrected entry took first place in the closed track of the shared task, the entry category limited to systems that use only the provided training data, with no external resources allowed. The paper frames the win as evidence that audit infrastructure can survive contact with the research it is supposed to be auditing, not as a benchmark result. A separate Sakura paper from the same period, titled "Sakura at BEA 2026 Shared Task 1: What Makes Vocabulary Difficult?", also competed at the same shared task (Sakura paper). The available excerpts frame it as either built with this framework or as a closely related entry; readers who need that lineage should pull the HTML versions side by side.
The texture of the run matters as much as the result. Twelve parallel agents on a single laptop. About $450 in API spend. Verifier scripts treated as load-bearing infrastructure rather than CI decoration. The authors do not claim Glite ARF replaces human supervision. They argue that when lapses happen, the framework surfaces them in the same pipeline that produced the work, instead of letting them spread silently across a long run of unattended experiments.
The cost is roughly 1% in extra compute for the verifier layer, a number small enough to be boring and large enough to be the difference between an artifact that survives review and one that does not. The paper is a preprint, not peer-reviewed, so the framework's value depends on whether other groups reproduce the audit behavior on workloads they did not design themselves.
Watch next: whether the corrections-overlay pattern gets picked up by adjacent multi-agent frameworks, and whether BEA 2026 organizers publish a public audit trail of feature-set rejections across all submissions, not just the winners. If Sakura does share authors with Glite ARF, it will be the natural place to compare verifier-driven runs against more conventional single-agent baselines on the same task.