Benchmark Atlas v0.6.0

Figure-verified systems expansion with four new grammars: Hunyuan-Large capacity-aware recycle routing, Seed1.5-VL loss-to-benchmark transfer calibration, and DeepSeek-V3 mixed-precision lifecycle plus fine-grained quantization accumulation.\n\nThe catalog now contains 83 unique SVG components across 42 source lineages. All entries pass provenance, uniqueness, rendering, accessibility, desktop interaction, and 390px mobile QA.

Benchmark Atlas v0.5.0

Adds eight PDF-verified research grammars from InternLM2, ERNIE 5.0, Step-3, and Yi: orthogonal MFU stress, unbiased replay-buffer scheduling, entropy-collapse diagnostics, cross-modal expert collaboration, attention hardware rooflines, training/decoding objective reversal, continuous-vs-discrete emergence, and layer-token similarity bands. The atlas now contains 79 unique components and renderers across 39 source lineages, with 19 figure-level evidence records.

Benchmark Atlas v0.4.0

First public release of Benchmark Atlas. Includes 71 source-grounded SVG components, 71 unique information grammars and renderers, 35 source lineages, figure-level PDF evidence for Kimi, MiniMax, and GLM research series, a generated JSON catalog, JSON Schema, TypeScript declarations, responsive catalog UI, SVG export, and validation tooling.