A project, played back

Charter and business case, 12 months of pallet despatch data by line, cold end layout, shuttle car cycle times, the vendor's throughput specification.

The 6.2 minute round trip is the vendor's figure and it assumes a clear run. I timed the corridor at shift change with a stopwatch: 8.9 minutes with pedestrian stops. Re-ran the model on the real number and the headroom on one route drops from 30% to about 8%. That changes my answer on vehicle count even for the single route. Better to find that in week 2 than in week 24.

35 minutes, most of it standing in the corridor with a stopwatch. The model took two.

Saying yes to a scope you have already modelled as undeliverable is a decision to fail later, in public, on your own signature.

Right answer, wrong delivery. The sponsor asked for five lines because the board asked her. If I hand back a flat no, she carries it alone, and I need her in week 20.

Reopening an approved budget is expensive in a way that doesn't show up on the budget. And picking the highest-traffic corridor for the first deployment is choosing to learn on the hardest case.

ISO 3691-4, vendor safety spec, cold end layout, pedestrian route survey over four shifts, incident log 2019 to present.

The model treated the 214 transits as evenly distributed and recommended zone speed limiting on that basis. They are two spikes, twenty minutes long, twice a shift. That makes a door interlock and a hard barrier at two pinch points far more effective than slowing the vehicle for the whole shift, and cheaper than the full barrier. The averaged number would have given me the wrong control.

The options table, 15 minutes. The transit survey was four shifts of someone standing there with a counter, and it is the only reason the answer is right.

The vehicle being certified is not the same as the installation being safe. She is telling me my design has one layer of protection and she needs two.

The safest option is not automatically the right one when it cannot physically fit. I'd rather bring Aiko three options that exist than one that doesn't.

This is what the averaged data told me to do, and it is the option I'd have taken if I hadn't gone and counted. Controlling for the average when the risk is a spike costs you throughput permanently and still leaves the spike.

Wi-Fi heat map (4 passes, loaded and empty racking), AGV vendor comms spec, plant network topology, access point inventory.

The fourth row is the one that mattered and the model took the vendor datasheet at face value. I asked Lucas directly: mission buffering holds a mission through a dropout of up to 8 seconds. Our dead stretch takes 40 seconds to traverse at survey speed. So the feature is real, and it does not solve our problem. Datasheet said supported; the number underneath it said no.

20 minutes. The question that saved the project was 'supported for how long', and it isn't in the table.

'Supported' is a marketing word until you make it a number. Eight seconds of buffering against a forty second dead stretch is a coincidence waiting to fail.

The trap here is that it will pass its acceptance test. You'll sign it off in a quiet week and inherit an intermittent fault that takes three months to diagnose because it only happens when you're busy.

This is the answer if we were designing the plant. We're not, we're inside a project with a window, and a solution that arrives after the window is not a solution.

Forklift telemetry 12 months, shift rosters, task observation over 6 shifts, warehouse vacancy list, business case assumptions.

The business case's 70% is a reduction in corridor traffic, not in driver hours, and somewhere between the model and the notice board those became the same sentence. They are not. I checked the telemetry myself rather than the case: 54% of hours, and the residual tasks are the ones that need a person. That distinction is the entire answer to Dan's question, and the business case wording is what created the problem.

40 minutes. The number was never the hard part. Noticing that two different percentages had been written as one was.

A question asked in front of six people gets answered in front of six people or you have answered it anyway, badly.

Never promise something you can't deliver to buy peace in a difficult meeting. The relief lasts a fortnight and the damage lasts the rest of your time there.

I'd do this as well, not instead. It's the best thing you can do for the project and for them. But it is a job offer to some of them, not an answer to all of them, and Dan asked a numbers question.

Fleet manager fault log (6 weeks), lane despatch records, shrink tunnel temperature trend, pallet profile spec from the vendor.

The obvious read is 'the AGVs are too fussy, widen the tolerance'. The night-shift clustering says otherwise: this is a shrink tunnel running cold at night and not shrinking the film properly. The AGV is the first thing in eleven years that has been unable to ignore it. Widening the tolerance would have hidden a real process fault that was already costing us in transit damage.

18 minutes to find the correlation. Rob knew the tunnel ran cold at night. Nobody had ever had a reason to write it down.

Turning off an alarm is not the same as fixing what set it off. The vehicle found a real defect that the plant had been absorbing manually for a decade. That is the automation earning its money.

This is a scope cut dressed as a routing decision. It also parks a driver next to the defect so nobody ever has to deal with it.

Check whether you're inside your own spec before you accuse a supplier of missing theirs. I've been on the wrong end of that conversation and it costs you more than the week.

Signed interface specification, vendor SOW, WMS message schema, integration test logs, two comparable deployments' rate cards.

I read §4.2 myself rather than trusting the summary, because this is the kind of finding that is satisfying and occasionally wrong. It says what the model said it says. But I also checked whether we ever asked for double-handling moves. We did, in week 11, and it is minuted. So the requirement is ours, it is in the spec, and it predates the CO. That sequence is what makes this a defect rather than an argument.

25 minutes. Twenty of them were finding the week 11 minute that proves the requirement wasn't invented after the fact.

Sometimes buying the date is right. Not when the thing you are buying is already in the contract, and not at week 20 when you still have the riskiest fortnight in front of you.

The contractual position is correct and I'd hold it. What I wouldn't do is hold it in a way that leaves Lucas nothing to take back to his own management.

Never solve a supplier's integration problem by weakening your own system of record. The cost lands on people who were never in the project.

Commissioning test logs, safety case sign-off, marshal training records, localisation fault dumps, charger utilisation, six weeks of dry runs.

Failure 1 is the one that decides tomorrow. Both vehicle 3 faults are within 6 metres of each other and both are after week 9, which is when we put the new access points in over the shrink tunnel. That is a stale feature map in a stretch we physically changed. Re-teaching that section is four hours, and it explains the symptom completely. I would not have connected those two facts without the fault dump coordinates.

30 minutes. The pre-mortem was going to happen anyway; the coordinates are what turned a mystery into a four-hour job.

This is the failure I have watched happen, and the technology was fine. The moment you take the manual route away before the automated one is proven, a one-hour fault becomes lost production on every line behind it. Labour savings of a hundred and sixty thousand a year do not survive that, and the loss lands in a completely different budget line from your capex, so nobody sees it coming until the plant number comes out.

Postponing to fix a four-hour map re-teach and six days of training is an inability to sequence. If the safety case were open I'd take March without blinking.

Parallel running is genuine risk control in a data cutover, where the old system costs nothing to leave switched on. Here the old system is people, and running both tells them which one you believe in.