Research

Things I tried, and what happened.

I build these tools by trying things. Often I expect a small local model to beat ordinary code, and often I am wrong. This is where I show the working, including the results that went against me.

You do not need any of this to use a tool.

So far

I tried a legal AI model at writing complaint letters. It kept my facts straight, and I am not building the tool.

12 October 2026

On four letters a small legal model added nothing to my notes, and a general one invented two dates. I preferred the legal model’s drafts and still dropped the idea, because writing a letter from notes is what people use a chatbot for. What I kept is the check.

Keep in mind: four letters to one housing association, notes written by an AI assistant, and no chatbot was run.

I tried three small models at reading a typed request. Ordinary rules did better.

6 October 2026

Asked which changes a request about a picture wants, ordinary rules got 59 of 60 right and three small models got 47 or 48. The next day a different kind of model, one that fills in a form, got 56. I ended up with a few plain rules and that model behind them, and two days later swapped both for a better form.

Keep in mind: every request was written by an AI model, which flatters the rules. Nothing typed by a person was tested.

I gave a frontier agent a guide to my local tool. It used more model work, not less.

19 September 2026

Claude Code with my guide and Tables did more work than Claude Code reading the picture itself, and got one fewer of seven small tables right. It also gave me a figure for what one table costs an agent.

Keep in mind: seven small tables I made myself, one agent, one pass.

What has been checked about Tables

10 to 14 September 2026

4 of 5 named synthetic tables exact; unruled adds an empty row. Two cropped phone photographs were read correctly. In my own use, mistakes remained in two prepared correction tasks.

Keep in mind: development examples and my own use. Nobody else has tried it, and I report no accuracy percentage.

I have not measured energy, and I do not claim a reduction in it. Tokens describe demand on a model, not energy.

How I keep the evidence straight

Three kinds of evidence, and three separate claims

Software checks show that the code does what it says, such as how a CSV is written. They say nothing about whether a model read your table correctly.
Model observations are results on named inputs with exact model files. They hold for those files and those inputs only.
Human-use findings are what happened when a person used the tool, including the mistakes they did not spot.

Three claims also stay separate, and each needs its own evidence: that a tool is useful; that a whole task uses less of a named resource at the quality required; and that publishing a finding helps someone else. So far: some evidence for the first, from one person; a token budget for one task, which is not a measured reduction; nothing yet for the third.