Research Scientific Accuracy

Scientific Accuracy

Overview

It is easy to claim "no fabricated numbers," but TokenPrint enforces this mechanically through architecture and CI/CD pipelines.

How TokenPrint proves it

1. The Ban on Math.random()

The build script (frontend/scripts/verify-data.sh) mechanically greps the codebase for Math.random(). The only exceptions allowed (via a strict allowlist) are for purely visual layout jitter, such as scattering the points slightly inside a tensor's bounding box in the Architecture Explorer to make the cloud look volumetric instead of perfectly uniform. All data values, lengths, widths, and opacities must come from the Zustand store.

2. GGUF Parsing Verification

The client-side GGUF parser has been verified against raw binary dumps of Llama 3 and Qwen models. It accurately extracts the byte offsets, decompresses the Q4_K quantizations, and retrieves the exact vocab_size and context_length metadata.

3. PyTorch Backend Integrity

The verify_real_data.py script ensures that the generate_steps loop produces the exact same token sequence and probability distribution as a standard model.generate() call using the HuggingFace transformers library.

4. Trace Diffing

The CI/CD pipeline uses scripts/trace_diff.py. It records a generation trace, makes a code change, records another trace, and diffs them. If refactoring the PyTorch engine alters a floating-point probability by more than a specified tolerance, the build fails. This guarantees that visual updates don't break the underlying math.

Related pages

• Design Philosophy
• Debugging

Further reading

• Verification Docs

Navigation

Previous	Home	Next
Design Philosophy	Home	Roadmap