Ask the Vendor What Happens When the Plate Reader Dies
A biofoundry is only as reliable as its weakest instrument, and the scheduler decides whether a failure costs one plate or the whole night's run. Before you sign, ask the questions that make a vendor describe how the system behaves when parts of it fail, not its throughput on a good day. Each question below points to a specific risk that you will otherwise find during commissioning, when it is expensive to fix.
Does the scheduler replan, or does it halt?
Many integrated cells run a static schedule that is computed before the first plate moves. In that design, one instrument timeout either freezes every downstream arm or leaves plates sitting on deck positions until an operator arrives. Ask the vendor to show you, on their own demo cell, what the scheduler does when one device stops responding mid-run: does it send work to a redundant instrument, park plates in a controlled hotel, or only raise an alarm? The risk here is about time, because every step in a DBTL cycle has a hold limit, and a scheduler that ignores hold limits will resume a run whose plates are already compromised. Require a configurable per-step constraint, for example a 30 min ceiling between reagent addition and readout, and ask whether the system automatically marks the plate as failed in its data record when that limit is broken. If the answer is that an operator checks the timestamps afterwards, the system will feed bad data into the Learn phase without anyone noticing.
Who owns the device drivers?
Ask whether instrument control goes through lab automation interoperability standards such as SiLA and the standard microplate footprint conventions, or through proprietary wrappers that only the integrator maintains. This decides whether you can replace a failed reader or incubator with a different make, or whether you have to wait for the original vendor's service queue. Ask a second question too: who tests driver compatibility when an instrument manufacturer pushes a firmware update? A closed driver stack turns a routine update into an unplanned outage, and from a shortlist you cannot see that risk unless the vendor gives you a written support boundary.
Where does run state live?
If run state exists only in the scheduler's memory, and not in a LIMS transaction that survives a controller reboot, one power interruption turns every plate on deck into an unidentified orphan, so ask to watch a cold restart in the middle of a run before you believe any uptime figure.
A decision rule for the shortlist
The most common failure in practice is not a dead instrument. It is a gripper that misplaces or drops a plate, after which the scheduler's model of the deck no longer matches reality and every later move becomes a collision risk. Ask how the system detects that mismatch, through barcode confirmation at each transfer, plate presence sensing, or nothing at all, and how long recovery takes. My estimate is that in the first year, recovery labour will account for more lost capacity than hardware faults, so judge vendors on mean time to recovery, not on headline availability. The rule: if a vendor cannot show recovery from a single instrument failure within one shift, without anyone moving plates by hand, rank them below a slower system that can. Redundancy is still not free, because it costs deck space and capital, so your next decision is which device to duplicate. Pick the one whose failure blocks the most downstream steps, which is usually the liquid handler, not the most expensive instrument on the deck.