— Gustavo Jasso, September 2026.
Experimental software
The software below was generated largely by AI with varying amounts of intervation on my part. The software is provided as-is, without warranty, support, or commitment to maintenance. Unless stated otherwise, the implementations have not been independently validated, and results should be independently verified before use in research.
AI disclosures in math.{RT,RA}
A tracker of AI-use and non-use declarations in arXiv preprints with primary category math.RT or math.RA, inspired by Louis-Simon Cyr’s disclosure tracker. It presents monthly statistics, source evidence, and information about disclosed AI models and companies, documenting disclosure practices rather than measuring actual AI adoption or judging individual papers. The website and its automated data-processing tools were generated and tested using AI, but have not been independently validated.
Date: Sep 21, 2026.
Model used: ChatGPT 6 Astra (medium) via Codex
gentle
An invariant calculator for gentle algebras, based on work by Chaparro, Schroll, Solotar and Suárez-Álvarez, and by Amiot, Plamondon and Schroll. The implementation was tested by the AI system used to generate it, but has not been independently validated.
Date: Sep 5-6, 2026.
Model used: ChatGPT 6 Astra (low) via Codex
Running time: Around 90 minutes.
Time I spent on this: Around 30 minutes.
Process:
I prompted the model to create a responsive web application implementing the computations obtained in the CSSS paper, describing the desired functionality of the application in plain words but in some detail (including that the computations should be "certified for correctness"). In particular, I asked that the application be "trivial to deploy". After 36 minutes of running time, a fully functioning web application had been generated.
In the second prompt, I instructed the model to add functionality implementing the results in the APS paper. In the third prompt, I instructed the model to include figures of the corresponding surfaces in the application, suggesting to generate these figures beforehand. For each of these prompts, the model ran for around 30 minutes.
In the fourth and final prompt, I instructed the model to test the implementation and verify that it is correct and matches the papers it is supposed to implement. Afterward, I did minor cosmetic edits to the website, adding disclaimers and removing unnecessary "branding", etc. This is where I spent the most time.
All prompts were entered in plain language, withouth any "engineering". Each of the four prompts more or less exhausted my 5h-period usage with a ChatGPT Plus subscription.
Date: Sep 8-9, 2026.
Models used: ChatGPT 5.6 Sol (high) in Chat mode. ChatGPT 5.6 Sol (middle), ChatGPT 6 Astra (low, extra high) via Codex
Running time: Around 30 minutes.
Time I spent on this: Around 30 minutes.
Process:
I prompted the model to follow the instructions generated through a ChatGPT chat, in which I simply asked the model to "Suggest an approach to 'certify [this software] for correctness using AI' so that it can be used in research." Generating the instructions took only a couple of minutes of minimal interaction, e.g. with me asking which model to use to save resources. The runtime for the execution of the instructions was around 20 minutes. Afterwards, I prompted to suggest how to fix the bugs that were identified and then prompted. This took around 8 minutes. The whole process used 75% of my 5h-period usage.