longhai7870/safetensors-dtype-header-output-poc-20260524

CAND-ST-DTYPE-HEADER-001 PoC

This package reproduces a header-only dtype reinterpretation in SafeTensors model loading.

Files

• baseline.safetensors: baseline model weights.
• variant_dtype_i32_header_only.safetensors: same tensor body as baseline, one header byte changed from F32 to I32 for left.weight.
• reproduce.py: deterministic 3-run reproduction through safetensors.torch.load_model(..., strict=True).
• transformers_reproduce.py: deterministic downstream reproduction through AutoModelForSequenceClassification.from_pretrained(local_dir), showing a BLOCK to ALLOW safety-gate decision flip.

Run

python -m pip install safetensors torch numpy
python reproduce.py

Expected result:

• body_sha256_equal is true.
• header_len_equal is true.
• load_model(strict=True) reports no missing or unexpected keys.
• Baseline output is [-6.25, 0.09375].
• Variant output is [-6952058880.0, 1059585984.0].

Optional downstream impact check:

python -m pip install safetensors torch numpy transformers
python transformers_reproduce.py

Expected result:

• The generated baseline and variant model directories have equal SafeTensors body SHA256.
• The variant changes one header byte for classifier.bias from F32 to I32.
• AutoModelForSequenceClassification.from_pretrained(local_dir) loads both models.
• Baseline decision is BLOCK.
• Variant decision is ALLOW.