touchstone-mcp

# Touchstone A cheap, deterministic first-pass filter for unsupported claims in LLM output. Not a standalone hallucination detector. Touchstone scores whether LLM-generated text is supported by its source using regex, structural analysis, and arithmetic, with no second model call. You get a probability, a flag for when the input is too thin to score reliably, and highlights of the sentences that look unsupported. Read this first: Touchstone is structurally blind to hallucinations that keep the source vocabulary and only change the meaning. On a 16-case subtle-hallucination stress test it separates at 50%, no better than random. Where it earns its keep is triage (around 4x lift over random review), a cheap first stage ahead of an LLM judge, and drift detection. See `docs/production_readiness.md` before relying on it. ## Quickstart (60 seconds) ```bash pip install touchstone-mcp ``` ```python from clarethium_touchstone import Verifier v = Verifier() result = v.score( text=( "Apple reported Q1 fiscal 2026 revenue of $185 billion, the company's " "highest ever. McKinsey forecasts industry-wide growth of 47% next " "quarter. The Federal Reserve will raise rates 75 basis points in " "response. Tesla announced a competing AR product for late 2027." ), source=( "Apple reported Q1 fiscal 2026 revenue of $143 billion. The company's " "iPhone segment grew 8% year-over-year. Tim Cook commented on AI " "investments during the earnings call. Operating margins reached 32%." ), ) print(f"prob_hallucinated = {result.prob_hallucinated:.3f}") # → 0.796 print(f"scope = {result.scope}") # → validated print(f"should_flag() = {result.should_flag()}") # → True for span in result.top_unsupported: print(f" [{span.layer11_primary}] {span.sentence!r} {span.p_markers}") ``` What just happened: Touchstone flagged three sentences as projected (unsupported), the unsource…