As a layperson skimming the paper, this strikes me as equivalent to a dashed-off letter to the editor coming from someone in Knuth's position. It's an incomplete, second-hand reporting of somebody else's results that doesn't really investigate any of the interesting features of the system at hand. The implicit claim (here and elsewhere) is that we have a runtime for natural-language programming in English, and the main method reported for demonstrating this is the partial prompt:

** After EVERY exploreXX.py run, IMMEDIATELY update this file [plan.md] before doing anything else. ** No exceptions. Do not start the next exploration until the previous one is documented here.

and later on, a slightly longer prompt from a correspondent using GPT-5.2 Pro, that also loads a PDF of Knuth's article into the context window. No discussion of debugging how these systems arrive at their output, or programmatically constraining them for more targeted output in their broader vector space. Just more of the braindead prompting-and-hoping approach, which eventually, unsurprisingly diverges from outputting any viable code whatsoever. This all strikes me as being an exercise similar to

You are a cute little puppy dog. Do not shit on the floor. Do not deposit bodily waste or fecal matter onto hardwood, linoleum, tile, and especially not carpet. Do not defecate indoors. Do not consume your own fecal matter.

The cargo-cult system prompt approach is like banging two rocks together compared to what a computational system should be capable of, and I would be much more impressed and much more interested if someone like Knuth was investigating such capabilities, instead of blogging somebody else pretending to have the Star Trek computer.