Changelog

0.4.0 — 2026-10-08

Retired the eight references explicitly rejected by the user and replaced them with visually inspected published constructions. The other twelve references and their image hashes are unchanged; the active set remains twenty references and twenty-nine assets. Added a shared user-global installer, documented host adapters, conflict detection, version backups, interruption recovery, and a read-only doctor. Reorganized entry points for install/use/maintenance, added a portable plugin manifest and local catalog, and documented the boundary between local scope, activation and cloud-account installation. No model runner or account setting is invoked.

0.3.0 — 2026-10-07

Added an authored 60-second walkthrough on the gallery and a shareable motion.html: question, reference inspection, visible mechanism, critique, delivery, and reference-quality possibilities. Uses only unchanged approved scientific images; no generated showcase or simulated agent result. Includes playback/seek/replay/chapter controls, reduced-motion mode, text transcript, and visibility-aware suspension. The 20-reference / 29-image canon and scientific review thresholds are unchanged.

0.2.0 — 2026-10-07

Restore every approved real published figure: 20 references, 29 image assets, no text-only replacements. Present the original six anchors first, including full AlphaFold and CellRank figures. Remove self-generated demonstrations from the current public tree. Require two actual reference-image inspections with registered hashes, concrete composition/encoding/style observations, and output-to-reference comparisons. Preserve source-specific attribution and rights rather than applying the repository MIT license to images.

0.1.0 — 2026-10-07

Initial public skill: mechanism-to-figure workflow; twenty calibration case guides; rights-aware image bundle; standard-library contract/preflight/review helpers; offline atomic installer; ten-dimensional anchored rubric; independent-review distinction; analytical examples; diagnostic task packets and matched-evaluation protocol.

The public calibration rebuild repairs a transparency rendering defect found in the earlier Gaussian-splatting figure raster. Original paper content is preserved; source and renderer treatment are documented.

No claim of cross-agent empirical efficacy or universal reference-level output is made in this release.