On the Foundations of Trustworthy Artificial Intelligence

arXiv cs.AI / 3/27/2026

💬 OpinionDeveloper Stack & InfrastructureIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that platform-deterministic inference is both necessary and sufficient to achieve trustworthy AI, formalizing this claim as the “Determinism Thesis.”
It introduces “trust entropy” to measure the cost of non-determinism and derives an exact relationship between verification failure probability and entropy (verification failure probability = 1 − 2^{-H_T}).
The authors claim a “Determinism-Verification Collapse,” stating that when determinism holds, verification can be reduced to O(1) hash comparisons, while non-determinism leads to an intractable membership problem for verifiers.
They show that standard IEEE 754 floating-point arithmetic violates determinism, and propose a pure integer inference engine designed to produce bitwise-identical outputs across ARM and x86.
In reported cross-architecture and geographically distributed tests (including on-chain attestations), the integer engine produced zero hash mismatches on models up to 6.7B parameters, concluding that AI trust ultimately depends on arithmetic-level determinism.

Abstract

We prove that platform-deterministic inference is necessary and sufficient for trustworthy AI. We formalize this as the Determinism Thesis and introduce trust entropy to quantify the cost of non-determinism, proving that verification failure probability equals 1 - 2^{-H_T} exactly. We prove a Determinism-Verification Collapse: verification under determinism requires O(1) hash comparison; without it, the verifier faces an intractable membership problem. IEEE 754 floating-point arithmetic fundamentally violates the determinism requirement. We resolve this by constructing a pure integer inference engine that achieves bitwise identical output across ARM and x86. In 82 cross-architecture tests on models up to 6.7B parameters, we observe zero hash mismatches. Four geographically distributed nodes produce identical outputs, verified by 356 on-chain attestation transactions. Every major trust property of AI systems (fairness, robustness, privacy, safety, alignment) presupposes platform determinism. Our system, 99,000 lines of Rust deployed across three continents, establishes that AI trust is a question of arithmetic.