Generate
Configurable watermarking during text generation
AI provenance · research toolkit
XAMMY
Make generated content traceable.
Bring watermark generation, detection experiments and provenance artifacts into one research workflow. Explore the relationship between attribution signals and text quality, and keep the evidence behind each run.
Evaluate
Detection experiments and text-quality comparisons
Trace
Generation records and retained provenance artifacts
Status
Research toolkit · pre-release
From generation to evidence
Follow the provenance workflow
A watermark provides a signal to evaluate. A signed record authenticates a retained payload. Explore how those pieces support a traceable generation workflow.
XAMMY / Generate
A signal inside ordinary language.
XAMMY explores watermarking at generation time. A configured sampler shapes token choices while the reader sees ordinary text.
MLX on Apple Silicon · vLLM on NVIDIA CUDA
Signal illustration
Our engineers found that the proposed design reduces latency without increasing cost.
Illustrative emphasisHigh
Explore the signal
Semantic conditioning and cached synonym candidates guide the choices available to the sampler. Adjust the watermark settings, then examine their effect on generated text and detection results.
Keep the evidence
Retain a generation report, fragment index and verification material. The provenance workflow supports signed records through optional ML-DSA integration; verification depends on the configured signing method and the artifacts retained.
Measure the tradeoffs
Compare baseline and watermarked outputs, evaluate semantic and structural quality, and test detection after rewriting. Results belong to a specific model, configuration and evaluation set.
Two generation backends
From local experiments to GPU evaluation
Choose the backend that fits your hardware and model. Shared configuration and evaluation tools help you compare experiments across environments.
vLLM backend
NVIDIA GPUs
Explore watermarking through the GPU generation pipeline, with configurable sampling and cached-teacher workflows.
MLX backend
Apple Silicon
Run local generation experiments through the same engine interface. Model compatibility and available features depend on the selected backend.
Evaluation
Test what survives a rewrite
Use the evaluation pipeline to explore where the signal holds and where it degrades. Compare detection with text quality and retain the run context needed to interpret the result.
- Baseline and watermarked generation comparisons
- Repeated paraphrase and detection experiments
- Semantic and structural text-quality metrics
- Checkpointed runs and generation traces
Pre-release
Bring a concrete use case
XAMMY is a research toolkit with generation backends, evaluation pipelines and provenance experiments. An evaluation starts with your model, content and attribution requirements, then defines the evidence needed to judge the result.
Contact the team