HEMO · TEXT OBFUSCATION
Same meaning. Different hand. Paste any model-generated text and get a rewrite that erases the statistical tells detectors look for — then scores itself, so it keeps iterating toward the only bar that matters: a human reading it and not flinching.
Deterministic and transparent. Same text + same seed = same result. New seed = a different fresh variant.
500+ model-favorite words and phrases get swapped — leverage → use, furthermore → plus, delve → dig. The vocabulary fingerprint is the loudest signal detectors use.
Run-on sentences get split, tiny ones get merged, adjacent sentences get safely reordered, sentence starts get varied, and paragraphs occasionally open colloquially. Real writing is bursty; model writing is smooth and metronomic. Smoothness is the tell — so vanish reintroduces the noise.
Common tokens get shifted onto the road less traveled, formal contractions loosen, dash styles randomize, and each paragraph gets its own drifting voice — casual, neutral, or terse. Detectors ride on smooth low-perplexity surface; vanish roughens it.
Every candidate is scored against detector heuristics: tell density, burstiness, lexical diversity, marker rate. Hybrid mode also fires a real paraphrase on the edge model (Cloudflare Workers AI) and the loop picks whichever candidate the proxy likes most. It exists to trek us toward fooling a human. That is the hard bar. Some humans.
| …leverage cutting-edge technology… | …use new tools… |
| It is important to note that… | Worth noting that… |
| Furthermore, the ecosystem… | Plus, the system… |
| …plays a crucial role… | …matters a lot… |
| In conclusion, one might argue… | Bottom line, you could say… |
Honesty clause: no watermark scheme is unbreakable and no heuristic detector is a court. Local mode lowers the statistical profile with deterministic rewriting; hybrid mode gets a genuine rewrite from an edge model (inputs go to Cloudflare Workers AI inference only — not persisted, not used for training) and lets the proxy pick the winner. Neither promises invisibility. It never fakes facts — it rewrites, it never invents.
Your text is processed in the request and discarded. Nothing written to KV, nothing logged, nothing trained on.
{"text":"…","seed":1,"strength":0.7,"mode":"auto"} → {result, stats, swaps, engine}. mode:"local" is deterministic per (text, seed); "auto" adds hybrid edge-AI rewrites and picks the best-scoring candidate.
/api/score?text=… → heuristic human-likeness 0–100 with per-dimension breakdown.
Input cap 150000 chars. CORS open. Works from any script, any n8n node, any agent.