Lessons · Lesson 5 of 5
Running forty of them at once
The weekly routine, the two exception queries that replace the status meeting, and the three measures that tell you whether the calendar itself is healthy.
Lesson 5 of 5 · 24 min
Twenty-five minutes, not a meeting
Everything so far has been about one order at a time. Ruwanthi has forty, and about four and a half hours a week that are not already spoken for. Spread evenly, that gives each order under seven minutes. That is enough to look at forty orders and understand none of them.
So the routine is not "review the book". It is two queries, run every Monday morning. Between them they turn forty orders into a list that is usually four to seven rows long.
Query one: which orders lost float this week, and how much. Not which orders have little float. That board belongs in the buying office, it is course 27.3's, and it answers a different question. This one measures the change, not the level: for every live order, float last Monday, float today, and the difference. An order that has sat at three days of float all season is stable and known. An order that went from twelve to four in seven days is the one that has just happened. It is invisible on a board sorted by level, because at four days it is still not the worst row on the page.
Thelawatte's threshold is a fall of more than two days in a week. In the first week the query returned 13 of 40 rows, which is what happens when a backlog surfaces at once. By the fourth week it was 5, and it has run between 4 and 7 since.
Query two: which milestone rows are in the past with nothing recorded. Every open row whose forecast date has passed and whose actual is empty, sorted oldest first, with the owner's name. This is the staleness measure from the first lesson, run as a worklist rather than as a statistic. It started at 168 rows over five days old. After a quarter it was 21.
The whole thing takes about twenty-five minutes and nobody attends it. It is a list, not a meeting. What the list produces is four to seven conversations, each with one person, each about one thing. Each of them is the meeting that was worth having.
What the register carries
The float register is one row per live order and it is deliberately narrow. Six columns, and two of them are the ones people leave out.
| Column | Why it is there |
|---|---|
| Order and on-board date | Identity |
| Float at confirmation | The baseline for slack, not for dates. Without it, nothing below can be read |
| Float today | The level |
| Change this week | The movement — the column the exception query sorts on |
| Largest consumer since confirmation | Which decision, in which screen, took the most days. This is the ledger from the previous lesson, in one cell |
| Row last touched | Whether to believe the rest of the row |
The fifth column is the one that changes behaviour, and it does so for a plain reason: it puts a name next to a number in a place other people read. In the first quarter it was a purchase-order screen thirteen times, a line plan nine times and a plant calendar four times. After two months, sourcing began quoting a lead-time difference in days and in float on its own comparison sheets, without being asked, because the register kept naming them.
Three measures of the calendar, not of the orders
There is a permanent temptation to judge a critical path by whether the orders shipped on time. That is a measure of the factory. These three are measures of the instrument. They are worth separating, because a calendar can be in perfect health on an order book that is going to be late anyway, and it will tell you so earlier.
- Edit latency. The median gap between the event that changed a date and the moment the system was told. Thelawatte went from 19 days to 3. Anything over a week means your calendar is a history book.
- Staleness. Open rows whose date is in the past with no actual, as a share of all open rows. It began at 60.5% and settled near 4%. Anything above about a twentieth means the exception list is not being worked.
- Forecast accuracy at thirty days. Of milestones whose forecast was set at least thirty days ahead, the share that completed within two days of it. Thelawatte: 41.2%, then 63.8% two quarters later.
The third one needs its ceiling said out loud, because somebody will otherwise chase it. A forecast accuracy near 100% would not be good news. It would mean one of two things: either nothing difficult is being forecast, or forecasts are being adjusted to match outcomes shortly before the event. The second is the previous lesson's baseline problem wearing a different hat. What a number on a page does to the person judged on it is course 10.4's territory, and this is a live example of why that course exists.
That ceiling is the same shape as two other findings in this course, and by now it should look familiar: a set of dates that never move (lesson one), a gate approved 98.1% of the time (lesson three), and a forecast that is always right. An instrument that never disagrees with reality has stopped touching it.
What it bought, honestly
Thelawatte ran this for four quarters. The detection lead time is the mean number of days before the on-board date at which an order's float first went negative in the record. It went from 9 days to 31.
That comparison needs a caveat, and the caveat is not small. Before the third date field existed, the record could not go negative at all until somebody typed it. So the earlier figure is partly measuring when a planner updated a spreadsheet, rather than when the order became late. The direction is solid and the size of the gap is not. Say that when you present it, because the person you are presenting to will otherwise repeat the number to somebody else with the caveat removed.
What can be attributed directly is narrower and more useful. In the first quarter of running the exception list, 6 orders were found at negative float. Four were recovered by re-sequencing at no cash cost. Two needed part of a colour air-freighted, at USD 8,900 in total. In the equivalent quarter of the previous year, 5 orders reached the same state and were found at around nine days out, when re-sequencing was no longer available, at USD 61,400 in air freight and late-delivery discounts.
Two different years with two different order books is not a controlled comparison, and it would be dishonest to present it as one. The thing that is genuinely comparable is the point of detection, measured the same way in both years, and everything else follows from it. The exception list did not make the factory faster. It moved the moment of finding out from nine days before the vessel to thirty-one, which is the difference between an option and an invoice.
Check yourselfYour float board is sorted by float, lowest first, and you read the top ten every Monday. What are you systematically missing?Show the answer
Everything that has just changed. A board sorted by level shows you the orders that have been in trouble for weeks, which you already know about. It hides an order that fell from twelve days of float to four this week, because four is still not the worst row on the page. Sort by the change as well as the level. Steady bad news is a known position. A sudden fall is an event, and an event is the only thing you can still act on.
Check yourselfWhich three numbers tell you whether the calendar itself is healthy, as opposed to the order book?Show the answer
How long a date stays wrong before it is corrected; the share of open rows sitting in the past with nothing recorded against them; and how accurate a forecast set thirty days out turns out to be. All three are properties of the instrument rather than of the factory, and all three can be computed from a log the system already keeps. Watch the ceiling on the third one: perfect forecast accuracy means either nothing difficult is being forecast, or the forecasts are being tidied up shortly before the event.
What you own at the end of the course
Three date fields with three owners. A permission table saying who may move which. A completion definition for every milestone where a tick and the world can come apart. A shortlist of gates chosen by their refusal rate. A slack ledger for one order that went wrong. And a Monday list that is four rows long instead of forty.
None of it makes anything arrive sooner. All of it moves the moment you find out, and everything you can still do about a late order is on the other side of that moment.