I · deterministic scan
No taste, no model, no network. It flags what can be flagged mechanically and cites the corpus study behind each severity tier. It reports. It never rewrites.
deslop is not a detector-bypass tool.
It will not iterate your text against GPTZero, Turnitin, ZeroGPT, or Originality.ai scores. There is no rewrite-until-the-number-goes-green loop, by design. That arms race produces mangled prose and teaches the writer nothing. Fix the writing. Ignore the score.
Detector evasion is the most common feature in this category. Every tool that shipped one was read in full during the teardown that produced deslop, and every one of those cores was left out on purpose.
The third pass is the one that makes the other two safe. A humanizer is a persuasion amplifier: it makes prose land harder, including when the prose is wrong. Smoothing a false claim is the real risk in this category, and most of the field does not address it.
One "however" is not a confession.
Every finding carries a bar-tag for how much license you have to act on it. Most findings are cluster, and that restraint is the whole design.
$ python3 scripts/deslop_scan.py scan draft.md 1:18 critical always artifact '[Your Name]' -> (remove; verify source) 3:7 critical cluster word 'delve' -> dig [Kobak et al. 2025] 3:22 medium cluster word 'vibrant' [Liang et al. 2024] 3:30 high cluster word 'tapestry' -> mix [Liang et al. 2024] 4:31 high cluster word 'leveraging' -> using [Liang et al. 2024] 5:1 low cluster word 'Moreover' -> also [Rosenfeld 2024] ... 2 more 8 findings: 2 critical, 3 medium, 2 high, 1 low
Twelve repositories, read in full.
Not the READMEs. The files. A README is marketing surface, and in this category it is frequently a work of fiction. One tool presents an AI probability, a burstiness index, and a type-token ratio, and it looks like an instrument. The burstiness index and the type-token ratio are both a random number generator. Paste identical text twice and get two different readings.
Four of the twelve were rigorous, and what made them good was never the word list. Every tool has a word list. It was knowing when to leave your sentence alone, which is the harder half of the problem and the half almost nobody built.
Two are in the current build by name: conorbronsdon/avoid-ai-writing (MIT), for the rhetorical and framing tell families and the asymmetric normalization thresholds, and mikiane/claude-watermark-cleaner, whose protected-region idea is reimplemented here rather than copied.
document.getElementById('meta-burstiness')
.innerText = (Math.random() * 8 + 2).toFixed(2);Hand the rulebook to your agent.
deslop is markdown all the way down, so the whole thing fits in one file: the README, the rulebook the model reads, the tell catalog, and the teardown that verdicts all twelve tools. Download it, drop it into your own assistant, and let it edit with the same discipline.
Clone it into your skills directory and it answers to its name.
Then ask for it: "deslop this draft", "kill the AI tells in this", "clean this copy". Requires Python 3.10 or newer.
Want the whole content system built and run for you, not by you? That is what Elorati does.