Determinism & forensic traceability.

Why determinism matters

Three institutional consumers, one underlying reason: a non-deterministic decision cannot be reviewed; a deterministic one can.

• Auditors. With determinism, the auditor receives inputs and policy version and re-derives the same output. Audit is a property of the architecture, not a process performed on top of it.
• Regulators. Frameworks for high-stakes AI — the EU AI Act most cited — require explainability, traceability and human oversight. Determinism subsumes traceability; explainability becomes “show the inputs, show the policy, derive the output”.
• Insurers. Pricing depends on reproducible loss estimation. A reproducible layer makes the loss model tractable; a probabilistic one makes it speculative.

The hash-chained audit trail

Each record carries inputs, policy version, output and provenance, signed with ECDSA P-256, with the hash of the previous record in its header.

Tampering is detectable in O(1) per record: modifying any field changes that record’s hash, invalidating the chain from that point. And the chain is append-only — no operation overwrites a prior record, including authorized correction. A correction is a new record referencing the corrected one by hash. The chain accumulates rather than replaces.

ECDSA P-256 signatures

Each record is signed with a P-256 key bound to the issuing layer. P-256 is the primitive used by TLS, major government PKI and the audit-grade signature ecosystem. Deliberate posture: novel cryptography in an audit context introduces review friction. P-256 introduces none.

Replay capability

Any prior decision can be reconstructed. The procedure is mechanical:

The same procedure whether replayed by operations during incident response, by an auditor during review, or by a regulator during inquiry — an hour after the decision or a decade.

OpenTimestamps proof

The chain is bound to wall-clock time via OpenTimestamps; current sealed-core proof SOV-2026-02-26. The proof anchors in the Bitcoin blockchain — not for any cryptocurrency reason, but because it is a public, durable, third-party-independent timeline that cannot be unilaterally rewritten by the platform’s operators.

Probabilistic AI safety vs. deterministic governance

Output filters, moderation classifiers, alignment training are valuable — and statistical. Good for the median case, weak in the tail, not reproducible from a regulator’s desk.

Deterministic governance is a different kind of guarantee: “given these inputs and this policy, this is the output, every time, and here is the chain that proves it.” The two compose — a probabilistic safety layer can be one input the kernel evaluates. But they are not substitutes.