Gemba

Gemba is a decision twin for operations. Slotting software and digital twins tell a warehouse where things should go and how the floor behaves; they treat the human decisions as fixed inputs, so they never show what a decision triggers next, and when the engineer leaves, nobody knows why the plan was shaped the way it was. Gemba simulates the warehouse and the people who run it. A deterministic hourly floor twin (28 aisles, a six-zone pick module, about 2,000 orders a day, synthetic data shaped on a real site) is run by an org of LLM agents defined in config: a program manager who chairs, ops, a safety lead with a veto, a quality lead who can lock a customer's SKUs, a shift lead who owns the labor truth, a director for escalations. Every simulated morning they hold a standup: options carry simulator what-ifs on forecast demand, stakeholders take positions citing that evidence, code scores the options against doing nothing and checks authority, the chair decides, and typed levers change the floor. A consequence engine turns outcomes into new events with causal links, so a night of aggressive moves can cause a picker learning dip, mis-picks, a key-customer escalation, Quality in the room, SKUs locked and the reslot slowed, and you can walk that chain back to the decision. An engineer uses it to rehearse a change before living it, fork any day with a different call and compare the two futures, run the same month under different company priorities, inject disruptions in plain words or describe whole scenarios in a sentence, and keep the org's full decision record. The org also reviews its own worst days: a coach traces the regret to its decision, proposes a lesson, proves it by forking that day, and the playbook is promoted only if it beats the previous one on held-out scenarios. Done-checks are written by the org within code-set bounds and verified every simulated day; drift detectors catch churn, stalls, prediction error and overspend.