Addendum v0.9 · Multi-fabric, multi-pipeline SOTA architecture
No single model, optimizer, or pipeline is allowed to run the company.
Multiple independent analytical and decision pipelines run against shared canonical state. Their outputs are reconciled, disagreement is exposed, and the best admissible action is selected or presented. Design / stub — not a live federated runtime.
Canonical Event State → N independent pipelines → parallel solvers → cross-model coherence → Pareto evaluation → policy/human constraints → ranked decision frontier → execution → receipt → settlement → champion/challenger → repeat.
Decision Fabric
The Decision Fabric reconciles candidates instead of letting one pipeline control operations: evidence weighting, cross-pipeline agreement/disagreement, hard-constraint validation, uncertainty, expected-value/risk scoring, ranked action set, then human or authorized autonomous execution. Agreement raises confidence. Strong disagreement is an operational signal — not something to hide inside an average.
Functional fabrics
Real-Time Operations
Seconds/minutes: booking, dispatch, call-ins, emergencies, Van state, rebase, windows.
Planning
Hours/days/weeks: staffing, PTO, maintenance capacity, protected reserve, borrowing.
Economic
Labor × cost × profit, contribution, margin, opportunity cost, pricing, procurement.
Evidence
Receipts, provenance, media, hashes, source conflict, point-in-time truth.
Learning
Training, calibration, historical replay, skill/demand learning, drift.
Simulation
Counterfactual schedules and no-impact synthetic operating runs.
Analytics
Reports, graphs, models, scales, probabilities, human-review queries.
Parallel pipelines
| Pipeline | Primary role |
|---|---|
| Deterministic Rules | Hard constraints, safety, qualification, SLA, policy. |
| Optimization | Dispatch, routing, capacity, scheduling, contribution, survival. |
| Predictive ML | Demand, duration, callback, inventory, conversion, urgency. |
| Simulation / Digital-Twin | Counterfactual schedules and alternative futures. |
| LLM Reasoning | Notes, synthesis, explanation, root-cause hypotheses. |
| Historical Analogue | Comparable past calls/days/branches as empirical evidence. |
| Cross-Branch / Federated | Company-wide Van, labor, inventory, and booking balance. |
| Human-Heuristic | Recurring override patterns — without assuming the human is always correct. |
Pareto frontier
Do not collapse every material choice into one opaque score. A manager may see a fastest-response plan, highest-contribution plan, lowest-reschedule plan, and best-balanced plan. Trades recommends one; alternatives remain visible and auditable.
Champion / challenger
Challengers stay in Shadow until settlement against actual outcomes. Do not delete the prior champion immediately. Preserve rollback and version lineage. Branch-specific ensembles are versioned; one universal mix is not assumed.
Fault tolerance
Any model, fabric, connector, or optimizer may fail independently. A Simulation or Analytics outage must not stop dispatch. An AI outage must not stop booking. A local branch continues if central analytics is unavailable. Cross-fabric contradictions are surfaced — Operations vs Workforce availability, Economic vs Schedule, Inventory vs stale Evidence — before the Decision Fabric emits an action.