Watermark Lab — planned, v2

In-browser watermark sandbox: generate text with a small local model (transformers.js + WebGPU) with a green-list watermark applied via an HMAC-SHA256 keyed spec — then watch the detector's z-score degrade live as you edit or paraphrase the output.

Also: verify any text generated with keys you hold. Byte-exact detectors for stock HF / SynthID configs require PRNG-compatible reimplementations — tracked as a later interop task.

Reality check: production watermarks (e.g. Gemini's SynthID-Text) are statistically invisible without the provider's keys. No local tool can prove or disprove them on a single text — anything claiming otherwise is snake oil.

AI Likelihood — planned, v3

Binoculars-lite: cross-perplexity between a base + instruct sibling pair (SmolLM2 / Qwen-0.5B class) running locally via WebGPU. One-time model download of a few hundred MB, cached by the browser.

Ships with a calibration mode — paste samples of your own writing to anchor the false-positive rate before trusting any threshold. Probabilistic signal, never proof.

Probe — planned, v4

Bring-your-own API keys (kept in sessionStorage only). Determinism test (same prompt, T=1, n≈30) and constrained-choice frequency tests across varied contexts to check whether an endpoint carries a sampling watermark at all. Burns a few thousand short completions per provider — cents to a few dollars per run.