Orchestration Lifecycle
Every session that enters QUORUM is a controlled execution unit. The orchestration layer manages session state across the nine-stage decision pipeline, enforces sequencing guarantees, handles concurrent session isolation, and ensures that every stage transition is logged before the next stage begins.
Sessions transit nine discrete stages in strict sequence. Stage skipping is not permitted. Each stage records its entry and exit timestamps before handing off to the next stage. The cumulative latency budget from INTAKE to COMMIT is 50ms under normal operating conditions.
Total pipeline budget: 50ms p50. TSA request is asynchronous and does not block session completion.
Each session is assigned a unique session_id at INTAKE. Sessions share no mutable state. The rule engine evaluates each session against a read-only snapshot of the current rule set — no session can modify rule state. Model weights are never mutated during a session.
No session can observe another session's signal vector, score, or evaluation state. Sessions are evaluated in independent execution contexts with no shared mutable memory.
Concurrent sessions cannot interfere with each other's scoring. The model inference layer is stateless per invocation — session A's score cannot influence session B's inference path.
Session state is garbage collected after COMMIT completes. Ephemeral session data does not persist beyond the COMMIT stage. Only the audit ledger entry and verdict record are retained.
session {
id: uuid_v4() // assigned at INTAKE, immutable
rule_snapshot: rule_set_ref // read-only snapshot, never written
model_version: weight_hash // immutable reference, never swapped mid-session
signal_vector: encrypted_blob // isolated to session context
state: RECEIVED → FINALIZED
gc_after: COMMIT.complete // ephemeral state cleared on commit
}
Pipeline stages run in strict sequential order within a session. Across sessions, the concurrency model is differentiated by stage: early pipeline stages parallelize freely; arbitration and consensus stages are serialized per-entity to prevent race conditions on entity state.
Each stage has a defined failure posture. Critical path stages halt on failure and return a safe-block verdict. Non-critical stages degrade gracefully, falling back to a simpler evaluation path. All degradation events are written to the audit ledger and are visible in the governance dashboard.
All degradation transitions are logged to the audit ledger with the triggering condition, affected stage, and degradation tier entered.
Every session progresses through a defined state machine. Error states are terminal — a session that enters an error state does not continue to subsequent stages. The error state, its triggering stage, and the error classification are recorded in the audit ledger before session termination.