Gold MMICKAI® LABS

Open the box

Sources & acknowledgements.

These are teaching experiments. Public algorithms and openly available tools make their mechanisms inspectable. No lab establishes product security, certification or compliance.

Voice studio

Speech is generated locally by Pocket TTS, Kyutai Labs, using KevinAHM’s browser implementation and ONNX model export. English 2026-04, INT8 models. We adapted the worker for local assets, one inference thread, no preset voices and explicit failure handling.

The model export is licensed CC BY 4.0. The browser source is Apache-2.0. Model card. Model processing uses ONNX Runtime Web 1.20.0 (MIT; third-party notices) and SentencePiece tokenisation (Apache-2.0). The included synthetic reference was generated with Windows speech synthesis; no person was recorded.

Browser-local open models

Transformers.js 4.3.1 (Apache-2.0), ONNX Runtime Web 1.31.0-dev.20260914-8d85527a0 (MIT; third-party notices). All model files are pinned by revision and SHA-256 in the asset manifest. These are pretrained third-party models, not MICKAI-trained foundation models.

Some converted repositories omit licence metadata; the manifest cites the upstream model licence. Whisper also retains its original MIT notice. Apache-2.0 licence · Whisper MIT notice.

OCR uses Tesseract.js 7 (Apache-2.0) and English tessdata (Apache-2.0). Interactive 3D uses Three.js 0.186.1 (MIT). Embedding projections use ml-pca 4.1.1 (MIT). Library notices are included in the distribution.

Immersive architecture and trace sources

Selected adaptations from pinned source snapshots, reviewed 21 October 2026. These are educational visuals, not live product telemetry.

Interface: React 19.3.1 / React DOM 19.3.1, MIT. React notice · React DOM notice · Scheduler notice. The transformer source also retains its third-party notices.

Trust experiments

ML-DSA-65 signatures use noble-post-quantum 0.7.1 (MIT). FIPS 204 specifies ML-DSA. Hash links use the browser’s SHA-256 implementation. The teaching format is separate from Mickai production formats. Keys are fresh and disposable.

Learning experiments

The tiny language model is a small neural character predictor with stochastic gradient descent. Memory retrieval uses explicit profile fields, event order, lexical TF-IDF/cosine similarity and graph traversal. Agent, reward and scheduling labs are labelled simulations. Scores are calculated from their fixtures, not invented product benchmarks.

The four original Trust Agent labs and twelve Agent gate challenges are preserved in the development repository. The coding studio preserves the original worker sandbox and supported DOM subset.

Brand and artwork

The gold M is the original MICKAI asset. The registered wordmark follows Trust Agent’s live-text design: Inter Tight, MICK in the text colour, AI and ® in gold. The hero and experiment covers are AI-generated photorealistic conceptual artwork, not photographs of MICKAI equipment. The supplied portrait sample is a fictional AI-generated adult. Branding is composed from the official asset and live text. Original covers are 1672 × 941 pixels, the native output of the built-in image tool. WebP derivatives are optimised without upscaling. Charts and 3D displays are computed from experiment values.

Typeface licences: Inter, Inter Tight, Playfair Display, JetBrains Mono.

Your data

The core Labs experiments collect no email addresses and send no analytics. Core experiments remain free without login or an email gate. The delivery team can offer optional newsletter signup separately and connect consent-aware usage analytics. Usage events contain only experiment ID, event type and app version, never inputs or outputs. Recordings and generated speech remain in memory until cleared or the experiment is left. Notebook results are stored in this browser, and can be exported or cleared. They do not include voice recordings. Loading local model assets makes requests to the same host as the page. Course and source links open external pages only when selected.

← Return to the Labs