The Transition
AI inference is moving from experimental systems to regulated infrastructure. This transition demands a shift from “trust me” narratives to cryptographic proof.
Then: Non-Deterministic Systems
Traditional inference runtimes treat AI as experimental. Inputs produce outputs, but the internal state remains opaque:
- Unreproducible: Same input, different output across runs
- Unauditable: No evidence trail for model decisions
- Ungovernable: Cannot verify compliance claims
- Unsafe: Debugging requires guesswork, not replay
This approach works for research demonstrations. It fails in production environments where reliability and accountability are mandated.
Now: Deterministic Runtimes
Deterministic inference replaces probabilistic behavior with strict guarantees:
- Reproducible: Bit-perfect output parity across hardware
- Auditable: Token-level trails for every decision
- Verifiable: Cryptographic receipts bind inputs to outputs
- Debuggable: Exact replay for root-cause analysis
The Engineering Shift
Determinism requires runtime-level enforcement, not just model-level claims:
Canonical Serialization
Inputs must serialize to identical byte sequences. No whitespace variance. No key-order ambiguity.
Fixed-Point Arithmetic
Floating-point drift breaks reproducibility. Q15-style fixed-point reduces numeric variance to acceptable bounds.
Seed Derivation
HKDF-SHA256 derives execution seeds from input digests. Same input → same seed → same output.
Receipt Generation
BLAKE3 hashes bind inputs, outputs, and routing decisions into tamper-evident execution receipts.
The Evidence Requirement
Regulated deployments demand proof, not promises:
- Government procurement requires verifiable audit trails
- Security clearances demand offline, reproducible systems
- Enterprise compliance needs cryptographic evidence chains
Deterministic runtimes provide this evidence. Non-deterministic systems do not.
AdapterOS
AdapterOS implements these principles as a production runtime:
- Canonical JSON serialization for stable hashing
- Fixed-point constraints to reduce numeric drift
- HKDF-based seed derivation for reproducibility
- BLAKE3 execution receipts for hostile verification
Patent application filed. Under review. Not an issued patent.
References
- Canonical JSON: RFC 8785
- BLAKE3: https://github.com/BLAKE3-team/BLAKE3
- HKDF: RFC 5869
This is a canonical research note. For an interactive visualization, see ai.jkca.me/then-now.