Determinism

Identical inputs. Identical outputs. Every time. Everywhere.

Execution A

Input: 0x1a2b3c
Processing...
Output: 0x9f8e7d
Hash: blake3(0x9f8e7d)
✓ MATCH

Bit-perfect parity

Execution B

Input: 0x1a2b3c
Processing...
Output: 0x9f8e7d
Hash: blake3(0x9f8e7d)

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.