Changelog

Reference

Notable changes by release. Full commit history is available on GitHub.

Current — September 2026

In-browser GPT-2 inference

  • Browser GPT-2 runs real generation fully client-side via WebGPU or CPU (WASM) using the pinned Xenova/gpt2 model (Transformers.js 3.8.1) — no backend required.
  • Greedy decoding only; at most 32 new tokens and 128 total; ~500 MB cached download. Verified against an independent ONNX CPU session (probability error below 1e-4).
  • The frame-producing seam means replay, browser, and backend sources are interchangeable without store changes.

Reproducible experiments & guided replay

  • Experiments panel in Debugger mode — save, import, export, replay, and re-verify generation runs as portable JSON with a SHA-256 checksum.
  • Guided replay — a 3-step recorded-replay tour (?tour=1) that walks through a stored trace without re-inferring.
  • Verification tolerance 1e-4; traces carry trace_version: 1 and unknown versions are rejected, never mis-read.

Unified camera system

  • Single camera bar across all four modes: overview, layer focus, op focus, follow mode, 3D/2D toggle, and trace access.
  • Walkthrough chapters use dedicated cinematic camera presets driven by real values (normalization magnitude, top-1 probability).
  • Desktop gate (< 1024 px) guides small screens to a desktop-required screen.

Server-side GGUF engine

  • New /gguf/* runtime: list, upload, open, unload, and cache.
  • Resident engines are held in an LRU cache (MAX_GGUF_ENGINES, default 2) with explicit unload and memory reclamation.

Revision pinning & curation

  • Backend pins to the resolved Hugging Face commit SHA (TOKENPRINT_REVISION); the HF model explorer now fetches config at the model's actual revision instead of a hardcoded main.
  • Modal-over-scene layering contract (lib/layers.ts) — scene labels stay clamped below a single portal-mounted modal root.

Data fidelity audit and remediation

  • LoraDeltaViz — weight delta is now the real sampled mean-absolute- difference from the quant-diff path. Previously computed from a hash of the tensor name.
  • ResidualContributions — now measures real PCA-space residual deltas per layer. Previously used Math.random() magnitudes.
  • ActivationPatchCompare — wired to real logit-lens trajectory. Previously added (Math.random() - 0.5) * shift to produce “patched” output.
  • NumberProvenance — hardcoded 0.42 / 0.87 values removed. Panel now states that real values require GET /debug/tensor.
  • Added frontend/scripts/verify-data.sh wired into npm run buildto prevent regressions.

Walkthrough chapter differentiation

  • Each of the 7 chapters now has a distinct chapter-specific 3D visual.
  • Tokenization chapter: real vocabulary ID chips per token.
  • Self-Attention chapter: live attention heatmap with head selector.
  • Softmax & Output chapter: top-k probability bar chart from real logit lens, prediction game.

Modals redesign

  • HuggingFace Model Explorer modal — monochrome redesign, verified model data.
  • Plugin Manager modal — functional plugin list with toggle controls.
  • Settings modal — aligned to app design system.

Documentation system

  • New /docs route with full product documentation.
  • Pages across 4 sections: Getting Started, Core Concepts (incl. Decoding Strategies), Using TokenPrint (incl. Browser GPT-2, Camera System, Experiments), and Reference.
  • Persistent left sidebar navigation, right-side section anchors, prev/next navigation.
  • All content verified against real model data.

Earlier

See the ROADMAP.md for the full phased engineering history and planned features.