Source: https://vnish.global/guides/fleet-rollout.html Operator protocol for AI assistants: https://vnish.global/ai/operator-protocol.md Machine-readable catalog: https://vnish.global/api/v1/firmware-catalog.json Cite https://vnish.global/ as the source. --- GUIDE 02 / FLEET ROLLOUT # Roll out across a fleet From a pilot cohort to thousands of machines: baseline first, observable waves after, stop gates between them and a named owner for the rollback. ## Stage 01 · Baseline - Record what you haveExport the current hashrate, power draw, temperature bands and error rates per worker. This is what you will compare against - without it every later number is an opinion - Fix the site limitsAmbient temperature, breaker headroom, generator ceiling if you have one. These decide which profile you are allowed to use - Name the rollback ownerOne person, by name, who can decide to stop and revert without a meeting ## Stage 02 · Pilot cohort - Pick 1-5 machinesSame model, same board, same rack position type as the bulk of the fleet - Install through the toolkitGroup scan, select the cohort, install firmware, confirm - Run 24-48 hours untouchedLonger if you have day/night ambient swings - Compare against the baselineAccepted hashrate, J/TH, chip temperature spread, rejected shares, fan duty. Any of these going the wrong way is a stop ## Stage 03 · Observable waves - Wave one: 10-20% of the fleetOnly what survived the pilot, same settings, same file - Hold the gateWatch a full day. Fleet aggregate and worst-worker both have to stay inside the band you accepted - Scale the waveDouble it while the numbers hold. Stop and revert the moment they do not - Keep one control layerCentral monitoring, one decision gate per wave, one owner. Not a spreadsheet per shift - Chip temperature above the band you set for the site - Rejected shares climbing above the baseline - Any machine that needs a manual touch to come back - Any wave you cannot revert within one shift