type666firewall/resh

Resh is a Python tool for analyzing philosophical and theoretical texts to evaluate how well their arguments actually hold up. It does not judge whether a text is true or false. Instead it looks for structural weaknesses: hidden premises, logical fallacies, circular reasoning, dogmatic assumptions, and rhetorical bias. The result is a reproducible numeric score called epsilon-resh, which measures the structural soundness of an argument on a scale from 0 to 1. The tool operates in two independent layers. The first is fully deterministic: it uses linguistic annotation, natural language inference, and sentence embeddings to compute a score from ten components, covering things like fallacy density, implicit premises, and logical validity. Because it uses no AI language model, the same text always produces the same score. The second layer is optional and uses a language model to apply a "critical arsenal" of deeper philosophical questions, including diagnosing which horn of the Munchausen trilemma the text falls into (infinite regress, circular reasoning, or dogmatic stopping point). The two layers are always kept separate in the report, and if they disagree, the disagreement is shown rather than hidden. The tool can analyze a single passage or a full document. For long documents it splits the text into chunks, processes them in order with a configurable call budget, and aggregates the results at the end. All AI calls are logged explicitly, failures are recorded as discarded contributions with error messages, not silently ignored. Installation requires Python 3.11 or higher and cloning the repository. A minimal install skips the machine learning stack and falls back to simpler methods, which is useful for trying the command-line interface without committing to a full GPU setup. The full stack adds Stanza (a linguistic annotation library), PyTorch, and an embedding model. The inductive layer needs an API key for a language model configured in config.py. The primary language of the README is Italian, but an English translation is included. The project supports both Italian and English input texts. The license is MIT, which means you can use, modify, and distribute it freely.