Research and writing
A study, two open model repositories and the founder's opinion pieces on AI evaluation, education and public policy, grouped by year.
Study
Zenodo, author preprint, not peer reviewed. DOI 10.5281/zenodo.21286873 · 10 July 2026
Coding agents are often given test-runner feedback and then credited with improving a repair. The preprint audits three experimental designs from the author's own research programme and finds that the feedback frequently exposed the expected output itself: in the primary QuixBugs arm, 156 of 169 consumed feedback messages (92.3 per cent) rendered the fixed expected-output field. The conclusion is that feedback which exposes a visible test assertion is causally confounded with verification, and that any claim of a verification benefit needs pass-or-fail-only feedback, a same-candidate-pool selection control, task-clustered inference and immutable transcript-to-row hashes. It is how NASCIA thinks about evaluation: an evaluation must test what its feedback reveals.
Read on ZenodoTwo repositories on Hugging Face under the founder's account, which he describes as LoRA adapters for British English prompt optimisation on Mistral 7B. Beyond that, only what the repositories themselves state is stated here.
Hugging Face
A LoRA adapter for mistralai/Mistral-7B-v0.1, trained with supervised fine-tuning (TRL and PEFT) and published in August 2025 with the adapter weights and tokenizer files. The card carries the base model, library and training tags; a full model card has not been published yet.
View on Hugging FaceHugging Face
A repository created in August 2025 under the same account. Its model card and files have not been published yet, so nothing is stated about it beyond its name.
View on Hugging FacePublished in the Brazilian and international press, in the language of each outlet. Titles are kept as published; links open the outlet's own page.