A Week of Retractions

This was the least glamorous and probably the most important week of the summer.

Lost in the Middle

My manuscript said every output file was committed to the repository. It wasn't, so I fixed that, and every result the paper cites now has a run directory behind it. I added unit tests for the heuristic model and the metrics (19 in total) and did a full citation audit that corrected several attributions and venues.

I also made three retractions. The paper had been described as 'submitted'; it hadn't been, and the docs now say so. The 'scales linearly with context length' line from May is gone. And I restated the context-length claim to say what the experiment can't show: whether the penalty grows with length. Writing 'this experiment does not test X' felt uncomfortable at first, but it makes everything else in the paper more credible.

Agentic DGG conversion

The Pure Math agent now draws a Graphviz before/after diagram of each rewrite rule, and I render everything into a self-contained HTML page with typeset LaTeX. LLM-generated diagram code broke on mathematical notation: a primed variable like x' contains an apostrophe that closes a string in DOT. I added a balance check, stricter prompt rules for labels, and a fix for a Windows console encoding crash on Greek letters.

I also documented a known issue: the local Llama 3.2 model kept adding prose after its JSON routing decision, so parsing failed and every request fell back to keyword routing. When requests did land on the right agent, it was because of how they happened to be worded.

RAG saturation

LaTeX is now the single source of truth for the manuscript, and I deleted the old Markdown version so the two can't drift apart.