0. Repairfailures, integritywarnings outLadder bookkeepingL3: fundamental tiersTrails, yield holdsand pair historyOperations research & recovery1. Look1,239 numbersper fleet84 base + 5 × 231field diamondTwo-speed memoryhalf-life 32 stepsPhysics & perception2. Decideone move per fleetfrom one snapshotConv + attentionheads 6 : 4 : 1, LearnerExploration schedulelow, peak, lowMulti-agent policy3. Movepositions, claimsand waitsDecide, guard, movehalf a hop per stepGoal claimsand open pickupsFleet operations4. Judgerewards per fleet3 heads + rung costsPay by consequencefixed value scalesNothing:judged fresh each stepReward design5. Rememberand learnthe whole stepas one transitionSARSA on themove madeBuffer of 20,000,target every 1,000MLOpsFLOWRRAone step, six phasesnew weights