FuguSTX

FuguSTX finds the tells of machine-written prose in English technical text. The engine reads a document and reports one finding per sentence: the byte offsets, a verdict, and a tell category. A tool or an agent repairs the text from the findings.

The engine is a Qwen3-0.6B fine-tune under llama.cpp, on the CPU only. The stx harness — Perl 5 over Fugu — computes every byte offset, and the model judges each sentence. The model learns from pairs of a human document and a machine mirror of the same facts.

The project is the pilot of FuguTTX. The build rehearses the FuguTTX production pipeline at small scale, on the same components, at real prices.

Documentation

• The source repository and the specification