One donkey per kind of output, all of them here. Run any of them and you get the specific problems, where each one is and the evidence for it, with the same verdict on every run. Writing starts on our own homepage copy, because we check our writing with the thing we sell.
Dead prose, hedging, jargon, self-management tics, flat cadence. Paste any AI text.
Sycophancy (the mirror wall), a dialogue stuck in one register, flat task-talk, narrowing. One turn per line, or "Speaker: text".
Upload a chart or slide. Reviewed substance first (does it carry a real finding?), then style (chartjunk, direct labels, color, title-as-finding). This one reads the image with a vision model, so it is not byte-deterministic, and it is rate-limited here.
Upload a PDF. The extraction donkey detects the table structure and pulls the cells deterministically, the same numbers every run. Byte-identical across 90 runs, 2,332 cells audited with 0 errors. (For the full visual overlay on the source page, the extraction donkey also runs standalone at wysiwyd.doloop.io.)
Paste a live URL. The design donkey reads the page and its stylesheets and flags system clutter: too many font sizes, weights, and content widths (the stranded-gutter class), plus color sprawl. Same page in, same verdict out, no model.
The architecture is one donkey per kind of output, so code is the next donkey, not a different product. Same contract: located, evidenced findings, the same verdict every run.