AIKernel.NET
version: 0.0.2 / status: Refactor / edition: Draft / published: 2026-05-16 / updated: 2026-05-16

Replayable Execution Dump Format

Defines execution dump structure for debugging, auditing, and deterministic replay guarantees.

1. Purpose

Model outputs are probabilistic, but an OS-grade AI runtime must guarantee reproducibility or full post-hoc traceability. This schema provides:

  • Deterministic replay: reconstruct failure-time state for reproducible debugging.
  • Audit-grade evidence: record PDP and routing rationale as tamper-resistant hash chains.
  • Automated regression detection: use historical dumps as baseline truth sets against provider/model updates.

2. Required Fields

The dump must be self-contained enough to reconstruct execution context.

Field Role
DumpId Unique session identifier (UUID/ULID).
ExecutionTimestampUtc Execution start time; dynamic-capacity reference point.
KernelVersion Runtime AIKernel version used for execution.
Seed Random seed used by model runtime for reproducibility.
PipelineDefinitionHash Hash of DAG and phase-boundary definition.
PromptArtifactHash Hash of governance-verified prompt artifact.
ProviderManifest ProviderId, Version, and runtime CapabilityVector hash.
MaterialSnapshotHashSet Hash set for quarantined material snapshots.
OrchestrationSnapshotHash Pre-execution orchestration context hash.
ExecutionOutcomeHash Hash of final output and intermediate thought artifacts.

3. Optional Fields

Additional diagnostics and optimization metadata.

  • Latency metrics: phase duration for Structure / Generation / Polish.
  • Token accounting details: input/thought/output token and cost breakdown.
  • Policy evaluation trace: detailed PDP permit/deny decision tree.
  • Routing score details: per-candidate IVectorMatcher score breakdown.

4. Serialization Rules

Rules to guarantee hash stability and binary-level comparability.

  1. Canonical Ordering: JSON keys are serialized in deterministic sorted order.
  2. Binary Content-Addressing: large binary payloads are referenced by content hash instead of inline embedding.
  3. Strict ISO-8601: timestamps use UTC YYYY-MM-DDTHH:mm:ss.sssZ.
  4. Floating Point Precision: scoring floats are normalized to fixed precision.

5. Integrity Checks

Before replay, IExecutionVerifier validates:

  • Resolution Check: every referenced hash resolves from snapshot store or repository.
  • Chain of Trust: PromptArtifact -> OrchestrationContext -> Provider -> Outcome chain matches exactly.
  • Signature Linkage: outcome remains logically linked to signer identity and kernel version.

6. Fail-Closed Rule

Any non-verifiable dump is treated as contaminated evidence.

  • Verification Failure: reject replay and emit security alert on hash-chain mismatch.
  • Mandatory Field Missing: reject load when required fields are absent.
  • Version Incompatibility: abort replay when KernelVersion drift makes logical reproduction invalid.

7. Storage Concept

Dumps are typically packaged as ExecutionDumpBundle (for example .admp) with three layers:

  1. Header: metadata and index.
  2. Context Map: references to runtime buffer snapshots.
  3. Trace Log: per-phase input/output hashes.

Changelog

  • v0.0.0 / v0.0.0.0: Initial draft
  • v0.0.1 (2026-05-06): Version upgrade aligned with documentation guidelines
Source: architecture/15.REPLAYABLE_EXECUTION_DUMP_FORMAT.md