A streak's price belongs to the marker, not the bolt
Worth reading first: The width of a bolt is worth what its faults leave it · A grade charges by the length and a cutter pays by the panel · A missing end is a fault the length of the piece.
The width of a bolt is worth what its faults leave it found that a warp streak, the fault a missing end leaves running the length of a piece, is the one fault a cutter dodges sideways rather than along. It spoils a whole lane of panels or none, depending on whether the panels can be placed across the bolt so that it falls between them. The four-point grade charges the same streak its full penalty in every metre, wherever it lies.
That essay proposed a specification that charged a warp fault by its position across the bolt: nothing within the spare cloth of a selvedge, a whole lane elsewhere. It left open whether such a charge would order bolts the way a cutting room values them. That question has two halves. Whether a mill’s real cutting records agree needs those records. Whether the charge agrees with the cutting arithmetic it was derived from is a computation, and it turns out to have a sharper answer than the proposal expected — partly yes, partly no, and the no is the more useful finding.
A population of bolts to grade
A grade is a ranking. It exists to say which of two bolts is worse, and a specification is a line drawn through that ranking. So the test of a grade is not whether its number is right for any one bolt but whether it puts a population of bolts in the order of what they are worth, and the instrument for that is a population.
The one used here is five populations of 240 bolts each, every bolt fifty metres long and a metre and a half wide. Each has point faults scattered at a rate drawn evenly between none and one a square metre; weft bars across its whole width at a rate drawn evenly up to one every four metres, the shape a mispick or a stop mark leaves; and none, one, two or three warp streaks, at odds of 45, 30, 17 and 8 per cent, each streak at a position across the width drawn evenly. Those rates and odds are chosen, not taken from a mill’s records. They make a population in which all three kinds of fault matter, and the five draws say how much the answers depend on which bolts happened to be drawn.
What each bolt is worth is what a fault map is worth most where the grade is worst and its successor measured it by: the panels a cutter loses who knows where every fault is. The panels are 450 by 900 millimetres, three across the bolt with 150 millimetres to spare and fifty-five along it, 165 to a bolt. The marker places each lane across the width, cuts each lane around its own faults along the piece, and the panels it cannot cut sound are the bolt’s loss.
The scheme stops at the first streak
The four-point grade scores each fault by its length, one point up to three inches and four beyond nine, and caps any linear metre at four points. A warp streak runs every metre of the bolt and is far longer than nine inches in each of them, so a single streak scores four points in every metre on its own and fills the cap everywhere. Nothing else on that bolt can then be charged. Every bolt with a streak grades at 200 points, which is 267 to the hundred square metres, whatever else it carries and however many streaks it has.
The panels lost do not stop. A bolt with no streak loses 21.8 panels on average, all to points and bars. One streak takes that to 53.1, two to 75.8, three to 91.3. The scheme cannot see the second and third streaks at all, and it also cannot see the faults that share a bolt with a streak. On the scheme’s account a bolt with one streak and a bolt with three streaks and twice the scatter of points are the same bolt.
Across a population that costs a measurable amount of ordering. Spearman’s coefficient between the four-point grade and the panels lost is 0.758 averaged over the five populations, between 0.744 and 0.776, and of all the pairs of bolts that differ in what they lose, the scheme puts 22 per cent the wrong way round. That is the number a better specification has to improve.
One streak is free in four places, not two
A cutter with the streak’s position can place the lanes so that the streak falls in the spare cloth, and the earlier essay placed that spare at the selvedges. It is not confined there.
Three lanes of 450 millimetres use 1,350 of a bolt’s 1,500. The spare 150 can be left at one selvedge, at the other, or opened between the first and second lane or between the second and third, and the lanes still fit. So a streak anywhere in a 150-millimetre band at either edge, or in one of the two interior bands at 450 to 600 and 900 to 1,050 millimetres, costs nothing. That is 40 per cent of the width, twice what the selvedges alone account for. Elsewhere it costs one lane, fifty-five panels, a third of the bolt.
A second streak needs a second gap, and there is only one spare to give. Two streaks can both be dodged only if they fall in the same band or the lanes can be shifted so one gap covers each — which with a single 150-millimetre spare happens rarely. The population shows it: 36 per cent of bolts with one streak lose no lane to it, 11 per cent with two, and 5 per cent with three. On average a streak costs 0.64 lanes when it is alone and less than that each when it has company, 1.13 lanes for two and 1.46 for three, because two streaks close together can fall in one lane and cost it once.
Five grades of the same bolts
With the cutting arithmetic in hand, five grades can be put against the same population. The four-point scheme as practised is the first. The others each change one thing.
No cap: the same length bands, every fault charged in full. Spare at the selvedges free: the earlier essay’s proposal as it was stated — capped as usual, but a streak within the bolt’s spare width of either selvedge is not charged. Each lane lost charged: points and bars scored and capped as usual, and the streaks taken out of the scheme and charged separately, by the lanes of panels they cost a cutter who knows where they are. Its price per lane is the scheme’s own price for a weft bar: a bar spoils one row of three panels and the scheme charges four points for it, so a lane of fifty-five panels is charged 73 points. The last grade, streaks left out, ignores them entirely and is the floor any of the others must beat.
The lane charge orders the bolts at 0.987, between 0.985 and 0.991 across the five populations, and gets 4 per cent of unequal pairs wrong where the scheme gets 22. Almost all of the scheme’s misranking is how it treats a streak. The points and bars it scores well enough: on the 114 bolts of the first population that carry no streak, the scheme orders them at 0.95, which is why the band of unstreaked bolts in the first figure is narrow.
The proposal as first stated does much less. Freeing the streaks near the selvedges raises the coefficient from 0.76 to 0.85 and still misorders 17 per cent of pairs. Most of what it leaves on the table is the interior gaps, the half of a streak’s free positions that are not at an edge, and the rest is that it keeps the cap: a charged streak still fills every metre and still hides everything else on its bolt.
Removing the cap alone is worth almost as much, 0.83, and costs nothing in the way of positions. The cap was designed so that one bad metre could not condemn a piece, and against a fault that runs every metre it does the opposite of what it was for: it lets one streak conceal everything else.
The price of a lane depends on the panel
The lane charge works because it asks exactly the question the cutter answers. That is also its weakness. A lane is not a property of the bolt. It is a property of the bolt and the marker together: how wide the panels are, how many fit across, and how much cloth is left over to move them in.
Compute the lane charge for 450-millimetre panels and give the bolts to a cutter using 500. At 500 millimetres three panels fill the bolt exactly, there is no spare at all, and every streak costs a lane wherever it lies. The 450-millimetre charge has forgiven 40 per cent of the streaks for being in places a 450-millimetre marker could dodge, and the 500-millimetre cutter cannot dodge any of them. The position-aware grade orders those bolts at 0.82, and the crude scheme it was meant to replace orders them at 0.91, because the scheme’s blunt rule — a streak is bad — happens to be exactly true when nothing can be dodged.
The same thing happens in the other direction and at other widths. At 400 millimetres the spare is 300 and four fifths of streak positions are free; the 450 charge orders the bolts at 0.78 and the scheme, which charges every streak, at 0.57. At 550 millimetres only two panels fit, with 400 spare, and the 450 charge scores 0.65. At 350 four panels fit with 100 to spare, and the 450 charge and the scheme are equally poor, 0.75 and 0.76. Charged at each cutter’s own width, the lanes order the bolts at 0.96 or better at all seven widths.
So the essay’s proposal has a precise answer. A specification that charges a streak by its position ranks bolts almost exactly as a cutting room values them — for the cutting room whose marker it was computed for. For any other it can be worse than the scheme. There is no single number that is right for every buyer, because the thing it would have to measure differs from buyer to buyer.
What should travel with a bolt
The grading essay named three repairs to the four-point scheme and judged the cheapest to be keeping the fault positions rather than summarising them. This is that judgement with a number on it. The positions let each buyer compute the lane charge for their own marker, and that orders bolts at 0.96 to 0.99; any summary computed before the buyer is known orders them between 0.57 and 0.99 depending on how well the summary guessed the buyer’s panel.
A streak’s across-coordinate is one number per streak, it is already measured by any inspection frame that records where a fault is, and on a population with a streak in every second bolt it is worth more to the ordering than everything else the grade records. It is the cheapest datum in the whole inspection report, and the one the grade throws away first.
There is a middle course for a mill that must still print a grade. Removing the cap, so that a second streak costs something and a streak no longer hides the points beside it, gets to 0.83 and needs no assumption about anyone’s panels. That is a third of the gap between the scheme and the best, bought without asking the question the mill cannot answer.
The bolts, the grades and the marker, named
The bolts are drawn from a fixed pseudo-random sequence so that every figure here is reproducible: each of 240 bolts takes a point-fault rate, a bar rate and a number of streaks, and then positions for each fault. Five populations differ only in the seed. The scheme is the four-point system’s length bands and per-metre cap as quoted by the trade, with a point fault taken as 25 millimetres long, a bar as the bolt’s width and a streak as its length.
The cutting is the marker of the width essay: three lanes, each placed across the width on a five-millimetre grid by a short dynamic programme and each cut along the piece around its own faults, with the better of that and a banded placement taken. It is a real marker and so a lower bound on the best possible one; on bolts carrying streaks alone the width essay found it meets the ceiling no marker can beat, so the lanes a streak costs are exact. The lanes a streak costs are computed the same way on the bolt with its points and bars removed.
The ordering is Spearman’s rank coefficient between each grade and the panels lost, with ties sharing their ranks, and the share of misordered pairs counts every pair of bolts with different losses and scores a tie in grade as half wrong.
Required of it, and not shown on the page: that a streak at 60 or 1,440 millimetres costs a 450-millimetre cutter nothing and one at 750 costs a whole lane; that every bolt with a streak scores the cap in every metre; that on bolts with no streak every capped grade is the scheme itself; that the lane charge orders bolts better than each of the other four grades in every population drawn; and that the population, the coefficient and the grade refuse two bolts, odds that do not sum to one, a negative fault rate, two lists of different lengths and a panel wider than the bolt.
What a computed ranking cannot say
The population is invented. Real mills have fault rates that differ by an order of magnitude between weaves and machines — a loom’s break log is one way of reading them, and a mill’s own inspection records are another — and real streaks are not placed evenly across the width: a broken end near the selvedge is more common on some looms, where the edge ends work hardest, and a tight end in the middle on others. The coefficients here would move with all of that. What would survive is the structure: that a capped scheme cannot see past one streak, that the free positions include the interior gaps, and that a lane is a fact about a marker.
One panel size is cut from each bolt. Real markers mix panel widths, nest pieces of different shapes and turn some of them, and a nap or a one-way print restricts which way each piece may lie. A mixed marker has spare cloth in more places and of more sizes, which should make the streak’s position matter more rather than less.
A streak is drawn as perfectly straight and perfectly visible, which is the simplest of the shapes a weaving mistake leaves. A tight end may show only after finishing, a streak may wander across a few millimetres, and a busy weave hides some faults that a plain weave shows. And the value a cutting room puts on a bolt includes things other than sound panels — the time to cut around faults, the risk of missing one, a buyer’s tolerance for a fault that falls in a seam allowance — that no count of panels includes.
Who measured what
The four-point system’s bands and cap are a long-standing trade convention, and the observation that a point grade and a cutting yield measure different things is as old as both. Recording fault positions on an inspection frame and passing them to a cutting room is an established practice in mills that have the equipment. What is added here is the measurement of how much each of several grades recovers on one population: that the capped scheme misorders about a fifth of pairs; that nearly all of that is its treatment of warp streaks; that the free positions for a streak include the gaps between lanes and so cover 40 per cent of a 450-millimetre marker’s width rather than 20; that removing the cap recovers a third of what a full position charge would; and that a position charge fixed in advance for one panel width can rank another buyer’s bolts worse than the scheme does.
Still open: whether one streak in a pair is worth breaking a lane for
Everything here charges a streak by the lanes it costs a cutter who keeps three whole lanes. A cutter could also break a lane: cut it short where a streak would have run into it, and fill the rest of the width differently for that stretch. A streak that starts part way down a bolt, as a broken end does when it is mended late, spoils only the length after it, and a marker that changed its lane layout at that point would save the length before.
Whether a streak’s charge should depend on where along the bolt it begins — whether a late-mended end costs a fraction of a lane in proportion to its length or a whole lane regardless — needs a marker that can change its layout part way down a bolt, and the dynamic programme here places lanes once for the whole length. The prediction is that a streak shorter than about half the bolt is charged nearly in proportion to its length, and that the proportion fails sharply once two streaks overlap along the bolt, because a layout that dodges both has to dodge them over the stretch they share.
Named objects
A flat tag is an object no other essay names yet.