The Activation Receipt
Switch between FP32, FP16, INT8, and INT4 precision to see how provider activation fingerprints diverge from the reference — and why DiFR flags model substitution in just two forward steps.
⬢ loading artifact…
Appears in
The Committed Sample: Activation Fingerprints for Verifiable LLM Inference
LLM non-determinism has been the blocker for output-based verification. DiFR commits to a random seed before inference — and activation fingerprints built from random orthogonal projections detect 4-bit quantization in just 2 tokens at AUC > 0.999.
8 min read ⬢ interactive