SPEX: Verifiable AI for the Real World

LLMs and AI agents are making more decisions than ever—but how do we verify that they’re doing what they say?

Why it matters:

Cheaper & faster than ZKML, FHE, or MPC

Handles LLMs, float arrays, embeddings, and even agentic plans

Resilient against lazy or adversarial solvers

Open-source Python lib (github.com/warden-protocol/warden-spex)

Key concept:

Hash computational states into Bloom filters.
Sample + verify at runtime.
Tune confidence level (X%) vs cost.

Read the paper: arxiv.org/abs/2503.18899

Deep dives:

SPEX as an AI firewall https://wardenprotocol.org/blog/spex-the-firewall-for-ai

Verifiability in the age of black-box models https://wardenprotocol.org/blog/spex-verifiability-in-the-age-of-black-box

Defending against manipulation
 https://wardenprotocol.org/blog/spex-combating-ai-system-manipulation

Trust but verify.

Then do it statistically.

#ZK #LLM #AIInfra #WardenProtocol #Web3Dev #SPEX

Enter SPEX (Statistical Proof of Execution) by Warden Protocol.

It’s a lightweight, sampling-based framework to prove that a computation (even non-deterministic ones like AI inference) was actually executed, without re-running the entire process.