The architecture
One World Model
One state. One transition function. Six decision domains.
Every other planning vendor sells modules and integrates them. Each module builds its own view of the world, and the seams between them become the integration tax the customer pays forever. Azirella built the shared world model first, and layered the decision modules on top.
The shared state
One representation of the supply chain's reality, read and written by every decision module.
Positions at every node, on-hand, allocated, safety, obsolete.
Production, storage, transport, and labor, states and calendars.
Orders in motion across suppliers, internal network, and customers.
Customer, production, transfer, purchase, their commitments and promise dates.
Product states, in development, launching, mature, phasing out, EOL.
Promised service levels, contracted volumes, and policy guardrails.
A promotion decision, a production decision, and a dispatch decision all modify the same state variables through the same physics.
Five hubs
Some state objects are touched by many planes. These are the hubs. Every hub has explicit ownership rules, so every plane can read the truth without stepping on anyone else.
Forecast
The most contested hub in any enterprise.
Demand owns calibration. Promotion and Portfolio contribute priors. Supply, Production, and Transport consume quantiles. Versioned, with an auditable override log.
Inventory state
Touched by every plane, physically owned at the facility.
Warehouse owns writes at each facility; Transport writes on receipt and putaway; everyone else reads. Event-sourced consistency, no reconciliation.
Capacity state
Three views on one truth.
Production owns actuals. Supply Planning owns the committed plan. MPS owns week-horizon reservations. Three writers, one substrate, with explicit conflict resolution per resource.
Lifecycle state
Portfolio owns; everyone else subscribes.
Pre-launch, launching, mature, phasing out, EOL. Portfolio emits lifecycle events; Demand, Supply, Production, Transport, Warehouse all consume. Fan-out is first-class.
Commitment ledger
A meta-hub: every promise every plane has made.
Service commitments (Supply to customer), volume commitments (Promotion to retailer), build commitments (MPS to S&OP), appointment commitments (Transport to Warehouse). This is what the outcome measurement service uses to close loops on every plane.
Most vendors can't answer "who owns this write?" in under five minutes. The hubs are how we answer it before the question is asked.
Four coordination mechanics
Planes collaborate without integration plumbing because the substrate runs four mechanics in parallel.
Constraint cascading
Upstream commitments become downstream hard constraints where necessary. Portfolio's launch schedule constrains Supply; Supply's deployment plans constrain Transport. Necessary but brittle on its own, so we complement it.
Lagrangian dual prices
Instead of hard constraints, upstream passes prices that shape the downstream cost function. Supply Planning doesn't dictate "build exactly this" — it passes a marginal cost of capacity and a shadow price on inventory, and Production runs its own solve against those prices.
The key property: when reality diverges from expectation, the downstream plane substitutes within its own flexibility envelope — without violating upstream intent. This is why the architecture bends without breaking.
Shared-state observation
A decision at one plane instantly changes observable state for every other plane. A factory breakdown updates capacity; the next Supply run sees the new reality; the next Portfolio cycle sees any long-term implications. The state ledger is the message. No explicit integration needed.
Upward feedback
Downstream realities flow back up as training signal. Chronic plan misses aren't just variance, they're evidence that the upstream model is miscalibrated. Health signals, calibration drift, and override-derived training all route back to the plane that was actually wrong.
Together these four mechanics replace the N² integration surface every other vendor ships, with a single substrate every plane reads and writes against. See the intersection catalogue →
Six decision domains. One state. One stream.
Each domain operates at its native cadence and horizon, with a policy class suited to the decision structure. The modules communicate through the shared state, not point-to-point integrations.
| Domain | Cadence / Horizon | Policy class | Objective | Status |
|---|---|---|---|---|
| Portfolio | 1–3 years | DLA with stochastic lookahead | NPV, market share, strategic fit | Roadmap |
| Demand | weekly | CFA / learned PFA | Forecast accuracy, shaping impact | Live |
| Supply | daily | CFA / VFA | Cost–service balance | Live |
| Production | hours to weeks | PFA / rolling DLA | Throughput, adherence | Roadmap |
| Transport | minutes to hours | PFA / DLA | Carrier cost, appointment windows | Beta |
| Warehouse | seconds | PFA | Labor utilization, pick-face throughput | Roadmap |
Upstream decisions become constraints or targets for downstream ones. Portfolio sets the product set and lifecycle curves that demand plans against. Demand sets the signal that supply and inventory plan against. Production sets loads that transport dispatches.
Why it matters
Adding a module isn't an integration project
Because every new decision module reads and writes the same state, expanding from supply to transport to warehouse to portfolio is additive, not re-architectural. No data duplication. No reconciliation logic. No silos.
CFA becomes a coordination mechanism
Upstream layers don't need to solve the downstream problem. They shape the cost function the next layer optimizes, a far more powerful and scalable pattern than trying to solve everything as one monolithic DLA.
The Decision Stream spans the whole enterprise
Because every decision is generated against the same world model, every decision is ranked against every other decision on the same urgency × likelihood scale. One queue. One mental model for the operator.
Scope grows without breaking the pitch
When a customer asks "can you also do promotion planning, or factory scheduling, or yard management?" the answer isn't "we'll build an integration." It's "we'll add a policy module over the state you already have."
"The unified state and transition layer is the strategic moat. It's what lets you add modules without integration tax."
How we add planes
We sequence by how much new state we have to model, not by perceived customer demand — because the state model is what compounds.
Supply Planning
The wedge. Already carries the substrate that the other planes lean on.
Transport
Close to live. Your substrate already tracks orders, inventory positions, and commitments, dispatch and carrier state are modest additions.
Production Scheduling
Changeover state, machine state, shift/labor roster, modest state additions on top of capacity state already modelled.
Portfolio & Demand Shaping
Significant state additions, lifecycle curves, cannibalization graphs, market response models, elasticity surfaces. Highest commercial leverage, because no other vendor does demand–supply unification well.
Warehouse
The largest state addition, facility-level slotting, equipment, labor pools, congestion. Comes last because the substrate extension is deepest.
This sequencing is an architectural argument, not a product roadmap. Every plane we add deepens the substrate, which makes every subsequent plane easier and every existing plane sharper.
See how the world model drives the Decision Stream
One state, six domains, one ranked queue. Live today for supply; expanding across demand, transport, production, and portfolio.