AKI, Atomic Kernel Inference, AgDR, cryptographic invariant, PPP triplet, BLAKE3, Ed25519, Merkle commitment, court-admissible AI, accountability by design

AgDR-Phoenix v1.8.0, April 2026

Atomic Kernel Inference

The cryptographic invariant ensuring every AI decision output is atomically bound to its accountable context, making provability a prerequisite, not an afterthought.

← AgDR Protocol Performance →

The Formal Invariant

AKI(i) = { commit(AgDR(PPP, Trace, Delta)) ⇔ output(Result) }

Interpretation: For any inference instance i, the cryptographic commitment to an AgDR package is logically equivalent to the emission of a result. No commitment, no output.

This invariant is enforced at the kernel boundary. It is not a logging policy. It is a type-level guarantee.

Note on provenance: Phoenix v1.8.0 is the first implementation to enforce atomic provenance binding per the AgDR specification. Earlier versions capture context metadata but do not satisfy the PPP Triplet atomicity requirement.

Core Principles

Atomic Commitment

Cryptographic sealing is a gatekeeper, not a side-effect. The kernel cannot emit output without first committing to context.

No Phantom Decisions

If the kernel cannot sign the AgDR package, the inference result is discarded before reaching any API boundary.

Provenance by Construction

Every outcome is mathematically traceable to its inputs, execution environment, and oversight chain, by design.

Contextual Integrity

The PPP triplet binds each decision to accountable authority, ensuring decisions cannot be decoupled from their normative context.

AgDR Package Components

// AgDR Payload Structure
{
  "ppp": {
    "principal": "did:accountability:...",
    "purpose": "credit_assessment_v2",
    "permission": "cbc_s122_delegated"
  },
  "trace": "blake3:...",
  "delta": { "state_before": "...", "state_after": "..." },
  "commit": "ed25519:..."
}

Each field is cryptographically bound. Mutation of any component invalidates the signature.

Implementation Notes

The AKI invariant is enforced at the kernel boundary. Reference implementations prioritize:

  • Memory safety: Rust-based kernel with zero-cost abstractions
  • Deterministic cryptography: Ed25519 signatures, BLAKE3 hashing
  • Verifiable scaling: Merkle tree commitments for O(log n) audit
  • Interoperability: PyO3 bindings for Python integration
  • Tamper-resistant timing: CLOCK_MONOTONIC_RAW for inference timestamps

Performance characteristics are documented separately to preserve the timelessness of this formal definition.

"In standard AI deployments, logging is an asynchronous side-effect. In AKI, the record is the prerequisite."