TypeSafe AI Playground — markjaquith

Watch out

The README calls this a playground, and the shape of it shows why: the PHI command deliberately refuses to produce a legal conclusion. Its four metrics mirror the independent elements of 45 CFR 160.103, and it says in as many words that real PHI status also depends on facts text cannot establish — who created or received the information, whether the holder is a covered entity or business associate, regulatory exclusions, and whether a Safe Harbor or Expert Determination process was completed. The identifiability rubric follows the conjunctive reading in American Hospital Association v. Becerra, requiring identifiability plus at least one actual health nexus. Operationally, a file is sent whole in one evaluation with no truncation, so request limits are your problem; PDFs and other binaries must be converted to text first; empty input is an error; and a directory scan includes hidden files without recursing or following symlinks. The parts-of-speech tagger leans on conventions the model has to interpret: option definitions live once in shared state while short mnemonic codes repeat with null descriptions, which the README admits is interpreted rather than a native reference feature, and the local shortcuts for articles, conjunctions and pronouns assume ordinary usage, so quoted words and unusual names will confuse them. Sentence boundaries are heuristic, not a parser, and the whole input plus question batch has to fit the model's limits. Copy reviewed against the repository README on 2026-10-04, not from running it.