Audit & Decision-Lineage Bus
An append-only, tamper-evident record links every action back through the proposals, checks, and approvals that produced it.
Every action has a provenance
For each command that reaches the breeder or burner, the lineage bus records the complete causal chain: the raw diagnostics, the twin prediction, the copilot proposal, the rules and envelope verdicts, any human approval, and the final command. Nothing actuates without a written explanation of why. This is the accountability backbone for a machine whose behavior is shaped by learned systems.
Causation and correlation
{
"command_id": "cmd_88f2",
"correlation_id": "shot_00421",
"caused_by": ["proposal_71a","envelope_ok_55","approval_leader_09"],
"rules_version": "2030.04.1",
"envelope_version": "2030.04.1",
"twin_version": "l3.twin@v2.3.1",
"copilot_version": "l5.copilot@v1.8.0",
"decision": "EXECUTED"
}
Tamper-evidence
Audit events are chained with content hashes so any retroactive edit breaks the chain and is detectable. The bus is append-only and never compacted. This matters both for internal engineering trust and for eventual regulatory review of an operating fusion machine; the discipline is built and exercised now, during the design-and-simulation phase, so it is mature before FOAK.
What it enables
- Deterministic replay of any decision from its recorded inputs.
- Root-cause analysis: from an anomalous burner plug transient back to the proposal that preceded it.
- Model accountability: attribute an outcome to the exact twin and copilot versions in force.
- Regression testing: assert a new orchestration version reproduces past approved decisions.
Separation from operational topics
The audit bus is a distinct topic class with its own retention and access controls; operational backpressure never sheds audit events (see backpressure). Distributed traces (see tracing) share the correlation IDs so latency analysis and causal analysis use the same identifiers.