Cimba discrete event simulation library
Copyright (c) Asbjørn M. Bonvik 1994, 1995, 2025 - 26.
Licensed under the Apache License, Version 2.0 (see LICENSE).

Preliminary project policy on use of Large Language Model (LLM) AI

As of mid-2026, LLM-driven tools have become very useful for coding tasks.
However, the state of the art is still not close to where a LLM model can be
trusted to independently write critical code. For that reason, the de facto
AI policy in the Cimba project can be summarized as follows:

* Nothing "agentic" whatsoever. Everything needs to have a human in the loop.
  No exceptions.

* Code belonging to the Cimba library as such is not AI generated. In some
  cases, AI (i.e., Anthropic Claude Opus in its latest version) may have been
  asked to propose code for some specific task. This first draft was then
  reviewed and modified as needed by me before use.

* Non-critical code, such as test scripts and tutorial examples, may be partly
  AI generated. I am not an expert in, say, CUDA programming or the finer
  details of the HDF5 file format, so it was helpful to get initial AI coding
  assistance for Tutorial 5 boilerplate code instead of starting from scratch
  with the manuals and a clean sheet of paper. However, it turned out that
  debugging that code and iterating until the program worked as I had intended
  may have taken as much time as if I had started from scratch instead. AI can
  make very strange mistakes.

* Based on this experience, please do not contribute AI generated code.
  Reviewing and debugging it can easily take more time than it is worth. It
  may also raise copyright issues if it includes snippets from a copyrighted
  training set. This will be very hard to chase down.

* The latest generation of AI, e.g., the "Mythos-class" such as Anthropic
  Claude Fable 5, is ruthlessly effective in pinpointing coding mistakes. This
  has quickly become a very helpful tool for debugging a complex concurrent
  program like a Cimba model, although the models' proposed fixes may still be
  misguided.
  The workflow is to ask the most powerful model available (currently Fable 5)
  for a review, work through the issues that were surfaced with support from
  the next lower-tier model (currently Opus 5), and have this model verify that
  all issues raised by its big brother eventually were resolved. And repeat.
  For full transparency, we collect the outputs of these reviews in the repo
  folder code_reviews, file names starting with the review date in YYYYMMDD
  format.

AI technology develops fast. The above should be understood as a provisional
policy at a specific point in time. There may be changes when new technology
becomes available, but the basic principle will not change: Always a human in
the loop. Only human creative effort can be copyrighted, and this library is
copyrighted. "Agentic" AI-generated code cannot be incorporated.
