A Writing System That Learns What I Would Have Cut

Asking a model to "write like me" produces something that reads like a model imitating a person. The failure is consistent: it over-crafts. It reaches for the constructed punchline and the three-beat close. I do not write like that, and neither does anyone who is just trying to say a thing.

This is the system that fixes it, and it is retrieval rather than fine-tuning, which is the right call for one person's writing on a frontier model. There are five register cards (raw personal, reflective, professional, academic, public) and a corpus of 12 real samples I actually wrote. On a generation it pulls two to five exemplars filtered by register first and topic second, because matching on topic measurably hurts style fidelity: you get a draft about the right subject in the wrong voice.

The part I would defend is the gate. Output goes to a separate judge agent scored against a fixed rubric before I see it, and the judge's agreement with my own rankings is logged every time. That number is currently 75%, across three rounds, against a trust threshold of 80%. It is not trusted, and it does not gate on its own yet.

What those three rounds actually found is more useful than the score. Round one, the judge over-rewarded an aphoristic line I found poster-y. Round two, it picked a punchy construction that the spec and the corpus both endorsed and I rejected anyway, which is how "no constructed punchlines" became an explicit rule rather than a preference. Round three, I rejected both options in two of four pairs, which meant the bottleneck was never the judge. It was the generator over-crafting before the judge ever saw anything.

So the fix went into the generator, not the scorer. That is the whole result: two rounds of tuning the wrong component, and the diagnostic that caught it was logging disagreement rather than assuming the judge was right.