Qwen-Image Cyrillic FP8 reproducibility configuration

Configuration and proof bundle for:

• 777Radik/Qwen-Image-2512-FP8-E4M3FN
• 777Radik/Qwen-Image-Cyrillic-ControlNet

Files:

• training-config.json — ControlNet training inputs, hashes, and hyperparameters
• runtime-config.json — component repositories, precision, generation, and retry settings
• benchmark-config.json — immutable heldout100 command and quality gates
• benchmark-results.json — BF16/FP8 metrics and manual audit result
• quantization-manifest.json — per-shard source/output hashes
• controlnet_benchmark.py — exact generation harness
• controlnet_accuracy.py — OCR/glyph diagnostic and contact-sheet generator
• verify_hf_release.py — remote LFS size/SHA-256 and required-file verifier

The accepted production result is 99/100 manual exact Cyrillic on the fixed benchmark. FP8 regression is zero: all 100 quantized outputs are byte-for-byte identical to the source-storage run.