Observed
Confirmed by a scan, a weight or a controller acknowledgement, with its source and time.
A twin that admits what it doesn’t know. A Brain that knows what it may decide.
The twin is built from WarehouseAI’s own records and observations, and it shows how sure it is. The Brain uses those facts, your procedures and an explicit authority model to decide what happens next. It records why, and it never moves equipment itself.
Decision receipt DR-4471, illustrativeAct
Lane D2 is blocked. Route R42 departs in 30 minutes and its packages are staged for D2.
AuthorityBounded automatic
RecordedInputs, options, policy, outcome
Not done by the BrainMoving any equipment
Five kinds of truth, never blended.
“P123 confirmed in B08 at 10:02” is a fact about the last observation. If the evidence suggests an unrecorded move, the twin shows an unresolved location instead of quietly guessing.
Inferred
Derived from evidence, such as a move nobody scanned. Always shown as inferred, never as confirmed.
Planned
Work that has been issued but has no evidence yet. A planned move is not a completed one.
Disputed
Contradictory observations are kept side by side and resolved by rule or by an authorised person.
Unknown
Shown as unknown, because not knowing is useful operational information.
A good-looking 3D view must never show an unconfirmed movement as done.
A twin you can rebuild and replay.
Current state is a projection of durable records, so it can be rebuilt, replayed and explained.
- Corrections point to the record they correct. History is never erased.
- The last message to arrive is not assumed to be the latest physical event.
- Each decision names the twin snapshot it used, and its preconditions are checked again before anything is committed.
- Scoped live reads, change subscriptions and historical reconstruction explain how a location or status was derived and why it changed.
- A language model never invents missing inventory or rewrites physical state.
Policy first. Reasoning only where policy runs out.
Routine cases use deterministic policies, calculators and optimisers. Novel situations, conflicting objectives and uncertainty can go to bounded reasoning or to a person. Continuous robot control never waits on a language model.
ObserveRead the twin snapshot and the procedures that apply
ClassifyDecide which kind of situation this is
EvaluateCheck the feasible options against hard constraints
DecideChoose within authority, or hold and ask
VerifyIssue normal work and evaluate the outcome
Authority is set per action, not per system.
A facility can let the Brain re-slot stock automatically while still requiring approval for disposal, stock adjustments or a disruptive reorganisation. Installing an adapter or a new model never grants broader authority on its own.
Observe onlyThe Brain watches and records. Nothing is proposed.
RecommendIt proposes an action with evidence. A person decides.
Approval requiredIt prepares the action. A named role approves it before anything moves.
Bounded automaticIt acts inside limits and reports what it did.
Procedures are records, not model memory.
Standard operating procedures, merchant rules and approved exceptions sit in a versioned library alongside the twin. Each has a source, an owner, a scope and effective dates.
- The procedures retrieved are the ones that apply to this customer, location, operation and time.
- Rules inferred from history stay proposals until someone reviews them and gives them a scope.
- A one-time override is not silently turned into a lasting rule.
- Every decision and message records the knowledge version it relied on.
When the reasoning service is down, the warehouse isn’t.
Known deterministic workflows continue within policy. Unknown consequential cases are held or escalated with an explanation. A failure is never treated as an approval.
- Model providers can be replaced without changing inventory or execution semantics.
- Inference cost and latency are measured per workflow, with budgets and fallbacks.
- A cheaper model is used only if it passes that workflow’s evaluations.
- Owning model weights does not make them the source of inventory truth or safe robot motion. The twin and the controllers are.
Test a plan before the floor sees it.
Scenarios start from a real layout and twin snapshot, with hypothetical demand, equipment and policies. Scenario events never touch production inventory or real equipment.
- Compare one robot out of service, a larger demand wave, a different slotting plan or a blocked lane.
- See the assumptions, the predicted throughput and delay, and how sensitive the result is to uncertain inputs.
- Promoting a simulated plan re-checks current facts and creates normal, reviewed work.
- Equipment emulators reproduce acknowledgements, timing and faults for virtual commissioning.
Bring us the warehouse you have.
Design partners get the first operating profiles built and tested against their own floor, SKUs and order history. You start with directed manual work and climb from there.