// the find
elder-plinius/GLOSSOPETRAE
LINGUISTIC ENGINE FOR AI
GLOSSOPETRAE is two things bolted together: a zero-dependency JS engine that procedurally generates constructed languages (phonology, grammar, script, audio, evolution), and a research project measuring how LLMs acquire and can be tricked by those languages, including built-in modules for steganography and tokenizer/guardrail exploitation. It's aimed at NLP/conlang hobbyists and AI-safety researchers, not general app developers.
The generative engine itself is genuinely solid — no external dependencies, 25 modules covering everything from phoneme inventories to SVG glyph rendering to formant audio synthesis, and it runs offline in a browser or Node. The experiment harnesses have real engineering discipline: echo-fidelity guards that return null instead of a false zero for unmeasured trials, Bonferroni correction, checkpoint/resume for long runs, and all 78 raw result JSONs shipped alongside the paper so claims can be re-derived rather than taken on faith.
The 'offense' modules — SteganographyEngine, TokenExploiter, LanguageAttributes with a literal 'Phantom' guardrail-evasion attribute — are shipped as production-status code in the same package as the language generator, not gated behind anything; anyone importing this gets covert-channel and safety-evasion tooling by default. The README itself admits the paper's first draft contained fabricated numbers caught only by self-audit, which is an honest disclosure but also a reason to distrust the rest of the unaudited headline claims (100% delivery, J=0 detection, etc.) without independently re-running the harnesses. Documentation is almost entirely marketing prose and tables — there's no real API reference beyond a handful of code snippets, so using the engine programmatically means reading source. Reproducing any of the experimental results requires burning OpenRouter API credits across nine-plus frontier models, which isn't cheap or fast to verify.