Determinism
Identical inputs. Identical outputs. Every time. Everywhere.
Execution A
Bit-perfect parity
Execution B
The Requirement
Deterministic inference is not "close enough" or "mostly the same." It is exact, bit-perfect reproducibility. Given identical inputs, the runtime must produce identical outputs on any hardware, at any time.
The Challenge
Floating-point variance introduces drift. Non-deterministic RNG breaks reproducibility. Undefined evaluation order creates ambiguity. A deterministic runtime eliminates all these sources of variance.
The Mechanism
Fixed-point arithmetic (Q15-style) reduces numeric drift. HKDF-SHA256 provides deterministic seed derivation. Canonical serialization ensures stable inputs. The result: perfect replay capability for debugging and auditing.
The Evidence
After execution, cryptographic hashes of the outputs are compared. If the hashes match, the outputs are identical down to the bit. This is the foundation of verifiable AI.