A donkey is a deterministic check: it reads your AI's output, finds the exact thing that is wrong, and never writes anything itself. Donkeys that work on the same kind of thing share one surface. Four surfaces, by what you are making.
The surface for anything your AI writes. Two donkeys:
Deworm catches slop, stoolworms, and insider jargon.
Pebble catches dead prose: flat cadence, power words, hedging.
The surface for anything your AI says back in a chat. One donkey:
Phaedrus catches a bot that agrees with everything, loops, drifts from its
character, or crosses a safety line.
The surface for pulling data out of documents. One donkey:
WYSIWYD catches the numbers an AI invents when it reads a PDF, by tying every
cell back to the page. Live today.
The surface for anything your AI puts on a slide. Donkey:
Inkwell catches charts that mislead, hide their source, or bury the finding.
In preview.
Each surface is the one doloop machine pointed at that kind of work: connect your own AI and the donkeys run inside a loop that remembers what it learns. The shared source for the prose donkeys is doloop-mcps.
Because it's everything the LLM isn't. The model is the clever one, and clever is exactly what makes it cut corners, improvise, and bullshit when it gets bored. You'd never want your adversary to be clever: clever can be reasoned with, flattered, gamed. A donkey can't.
A donkey hauls the load the smart one won't stoop to: the tedious, exact checking, every line, every time. No part of the job is beneath it.
Sure-footed and plodding. That is determinism with a face on it: it doesn't get bored, doesn't improvise, doesn't have an off day.
Stubborn as a donkey. The one counterparty your AI can't charm, bluff, or talk into agreeing. It holds the standard because it is too stubborn to be talked out of it.
You bring the genius; we bring the mule that won't let the genius get away with anything.