hinanohart/faithful-echo

faithful-echo is a small, independent post-hoc tool that detects where an LLM softened, paraphrased, hedged, or numerically rounded user wording. It classifies differences along seven deterministic drift types (entity, numeric, quantifier, hedge, abstraction, subject swap, tone smoothing) so humans or automated gates can decide which matter. Who uses it: prompt engineers, reviewers, QA/MLOps, or anyone integrating Claude Code agents who need deterministic drift diagnosis (not rewriting or fact‑checking). Key features: standalone Node (≥20) CLI and Claude Code agent, TypeScript programmatic API, per-rule modules, locale-aware small dictionaries (en/ja), task modes (verbatim_quote, paraphrase, summarize, translate) that control strictness, pretty/json output, exit codes by severity, and an overrideable dictionaries dir. Limitations: heuristic (regex/dictionaries), bilingual coverage, not a rewriter or semantic fact-checker.

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.