WMTrace: Open forensic tool for detecting LLM text watermarks

WMTrace addresses a growing practical problem: as major AI providers including Anthropic, OpenAI, Google, and Meta deploy watermarks to mark their outputs, the ability to verify and understand those marks becomes important for transparency, policy compliance, and academic research. The tool explicitly rejects the temptation to offer a single opaque confidence score (e.g., "97% AI") and instead provides evidence bundles with categorical grades and mandatory scope statements that spell out what the result does and does not support. This design reflects a mature understanding of forensic inference: a watermark detection is evidence for compatibility with a declared scheme under a specific configuration, but it cannot prove authorship, provider identity without trusted key provenance, factual accuracy, or intent.

The research specification, preserved in the repository at docs/RESEARCH.md and current as of 2026-08-14, covers the full taxonomy of watermarking techniques (KGW, unigram, unbiased reweighting, semantic and multi-bit schemes, and cryptographic marks), provider adoption status, attack and robustness analysis, and a published roadmap (P0–P6) for future detectors including model-based approaches (SynthID-Text reference, Fast-DetectGPT, Binoculars) and C2PA verification. By separating watermark detection, generic AI-text detection, model attribution, and provenance verification into distinct code namespaces and never averaging them, the tool avoids conflating different forensic questions and encourages users to think precisely about what they are trying to verify.