Appendix — how this essay was built

An essay that argues you should test how you work owes you the same about itself. So here’s the method, and the one honest result I have on whether it helped.

The method

Start with who actually did the work, because it wasn’t me. The agent held the pen — it drafted every line. I was more the contractor than the workman: I brought the raw material, drew up what “done” looked like, judged what was true, and signed off. The essay says as much on page one; this is the machine admitting it about itself.

It didn’t draft alone, either. As it wrote, it ran a panel of other agents dressed as caricatures of writers I trust, whispering in its ear: don’t waste that turn, cut that hedge, you’re explaining what you already showed. Not voting on the prose — heckling it. (One of them was Groucho.)

Then the part that did the most work was the machine’s, not mine: it read the whole draft, closed its eyes, and wrote it again from memory. What survives that is the load-bearing stuff — the lines that come back without the page in front of you. The rest was scaffolding nobody missed.

After that, the loop the essay describes, run on the essay itself: a different model family grading each pass A–F and handing it back bleeding, a panel for coverage, a voice pass to keep it sounding like me, fixes, repeat until it stopped dropping below an A. Every section went B → A that way. The grade gate was always a different model than the one that drafted the page, for the reason the essay gives — a model grades its own work generously. My part stayed the contractor’s: I sent the draft back to be redone, but never without writing down the exact steps it had to hit first. That punch-list is the loop.

Somewhere in the middle of it a pass came back sounding like me on the first try — not cleaned up to sound like me, just me — and I sat there two beers in, oddly proud of a thing that doesn’t know I exist. The method doesn’t capture that and the experiments can’t score it: across a long enough night you stop fighting the machine and start rooting for it. It still has no idea. The pride was mine, same as everything else.

Did the method actually help?

I tested it, because not testing it would have been the one unforgivable thing. I ran a second version the plain way — one pass, no whispers, not from memory, no loop — and had a different model family, blind, pick which it would rather a reader finished. It picked this one: more memorable, a stronger ending, teaches more by being lived.

Here’s the catch, and it’s the same catch as everywhere else in the essay. This version also got more revision than the plain one. So the win is confounded: I can’t separate “written from memory with the panel whispering” from “simply worked on longer.” The most I’ll honestly claim is that the from-memory re-tell produced a better ending, and that iteration bought precision. The isolated effect of the whispers is unproven. Directional, n=1, one judge — exactly the asterisk the rest of the essay wears.

Why tell you this

Because the whole essay is one argument: the machine does the labor, the judgment stays with the person, and the only way to know which of your clever moves actually did anything is to test it and keep the failures. This appendix is that argument turned on the essay itself. You watched me poke the jellyfish. This is me telling you which pokes moved it, and which just looked good.