Data Flow

End-to-end data flows through Jope.SMB — from Raman scan to operator display to audit trail — captured as sequence diagrams. Every flow respects the compliance rules (append-only, hash-chained, signed where required).

Online Inference (Raman → Prediction)

The hot path runs every 3 seconds during a batch: Raman captures a spectrum, Console forwards to Inference, Inference returns predicted concentrations, Console displays + writes to Historian.

Batch Lifecycle (FSM)

Batch state transitions are driven by Stateless; every transition is audited and signed when required (21 CFR 11.100).

Each transition writes an audit event with {actor, timestamp, signature_id, batch_id, from_state, to_state}.

Electronic Signature Flow

All critical actions (batch start / abort, recipe change, model swap) require an electronic signature under 21 CFR 11.100(b).

Signature records are tamper-evident (hash chain over {prev_hash, actor, meaning, reason, timestamp, context_ref}).

Audit Trail Write

Every mutation emits an audit event. Events are append-only and hash-chained to make tampering detectable.

Audit events cover: login / logout / user / role / permission changes, batch transitions, recipe changes, model load, backup / restore, audit export (meta-audit).

Model Upload & Hot Swap

AI research team delivers trained .joblib model file. Operator uploads via Console; Inference Server loads and verifies.

Training Trigger (Async)

Re-training happens when new ground-truth (HPLC) samples are added to the training set.

Training runs on a separate thread inside Inference Server — predict traffic remains live during training.

Alarm Flow

Threshold violations (pressure high, flow deviation, inference timeout) raise alarms.