The panel with no corner

Sweeping the rise was supposed to separate two readings of the corner — the contact scale, or any memory at all — and it did, by refuting the second. What it also produced was three different shapes where a single shape sliding along an axis was expected, and the middle one is the least discussed and the most informative.

At the coarsest rise the deeper rule wins at both ends and loses across the middle. At the finest it loses at white noise and wins from about seven tenths upward. At the rise in between it wins essentially everywhere, and there is no crossing anywhere in the range.

A missing feature in the middle of a range is not a gap in a trend. It is either a fact about that lattice or a fact about the comparison being made on it, and this essay is about which.

The distinction is not academic, because the two readings say opposite things about whether the corner exists as an object. If the flatness is a fact about the lattice, then the corner is something a lattice either has or does not, and the thread has found a lattice without one — which is a property to explain. If it is a fact about the exponent pair, the corner is a feature of a comparison rather than of an arrangement, and the thread has been naming a relation as though it were a thing.

What “no corner” means here #

The comparison is between two rules with different falloff exponents, run on the same seeds with the same disturbance, and what is counted is how many seeds agree that the deeper one passed more drift. A corner is a correlation length at which that count crosses a majority.

At the middle rise the count is a majority at every correlation tested, including white noise. There is nothing for a crossing to be between, and the margin is not marginal either: five of six seeds at the low end and six of six across most of the range.

That is a different situation from a corner at the edge of the range. A corner at zero would mean the deeper rule always wins in the region tested and might lose below it; here the deeper rule wins at zero, which is the region where the disturbance has no memory at all and where the original prediction said the shallow rule should do best.

It is worth restating that prediction, because the middle panel is its most complete failure. The reading under test was that a rule reading a deep neighbourhood should average away a fast disturbance and be defeated by a slow one, so the deep rule should win at white noise and lose at long correlations. The middle panel has the deep rule winning at both ends and everywhere in between: not the prediction reversed at a crossing, but the prediction failing at every point of the axis it was made over.

Two readings, and the measurement that separates them #

The lattice reading. Something about the 5/8 arrangement makes the deeper rule better at every correlation, so the comparison has no crossing on it. If that is right the middle panel is a fact about arrangements and belongs beside the other two as a data point.

The exponent reading. The panels compare one pair of exponents. If at the middle rise both of those rules happen to sit on the same side of whatever threshold matters, the comparison flattens without anything about the corner changing. On this reading the middle panel is a fact about the pair being compared, and a different pair would show a crossing on the same lattice.

The measurement that separates them is a third exponent at the middle rise. If the flatness survives — if every pair of exponents at that rise shows the deeper rule winning throughout — it is the lattice. If a different pair produces a crossing, it is the pair.

The exponents available are not unconstrained, which makes the test sharper than it sounds. The comparison requires the two neighbourhoods to differ by at least a factor of three, because pairs closer than that are a coin toss at every colour — a fact this thread established separately and which is asserted by the generator rather than assumed. So the third pair is drawn from a small set, and exhausting it is a bounded piece of work rather than an open-ended sweep.

That is one sweep of six correlations at six seeds. It is the cheapest outstanding measurement in this thread and it has not been run.

Why it was not run #

Because the thread’s question was about the corner’s position, and a panel with no corner reads as a panel with no data. It took reducing all three panels to one line each before the absence became the interesting cell rather than the empty one.

This is a recurring shape in this collection and it is worth naming. A sweep looking for where something happens treats the places it does not happen as background. The asymmetry is built into how a sweep is usually written: the code finds the crossing, returns its position, and returns null when there is none — and a null is easy to render as a blank cell and hard to render as a claim. The generator behind these panels prints the shape of each row rather than only the crossing, which is a small change that turned an empty cell into the subject of this essay. The census that looked for which family survives reports the offsets that heal for exactly this reason, and reports them because a table of only the wrecked offsets would be a picture of the selection.

What a flat panel would mean if it held #

Suppose the third exponent comes back flat too, and the middle rise genuinely has no crossing at any pair. What follows is more interesting than the corner thread’s original question.

It would mean the deeper rule’s advantage is not a monotone function of anything — present at 3/5, absent at 5/8, present again at 8/13 — and a quantity that switches on and off along a ladder is not a scale. It would put the corner closer to a resonance than to a threshold: something that appears when the arrangement’s own numbers stand in some relation to the disturbance’s, and vanishes when they do not.

That is a much stronger claim than anything the thread currently supports, and it would need the sweep run at every rung rather than at three rises. It is recorded here as what the flat panel would license if it survived, not as what it shows.

There is a third reading of the middle panel that neither of the two above covers, and it is the one that most weakens the case for treating its flatness as a finding.

Look at what the panel contains rather than at what it lacks. The deeper rule takes five of six seeds at white noise, then six of six at every correlation from a half to nine tenths, then four of six at the longest. It is not flat because the two rules are hard to tell apart there. It is flat because one of them is winning nearly every run, and a comparison that has run out of room at the top has nowhere to put a crossing.

That is a ceiling, and a ceiling is a property of the comparison rather than of the lattice. Six of six is the largest number the panel can print, so if the deeper rule’s advantage grew or shrank across those four correlations the panel would report the same six each time and no feature would appear. The middle rise is therefore not a lattice on which a corner was looked for and not found. It is a lattice on which the instrument was saturated, and the honest reading of a saturated cell is that the measurement was not made.

The ceiling has a number on it, and the number decides which escape is worth taking. A cell prints six of six whenever all six seeds agree, so it saturates as soon as the deeper rule’s true win rate is high enough that six agreements are the ordinary outcome. Six of six happens more often than not once that rate passes 0.5 to the power of one sixth, which is 0.89. So the panel cannot distinguish a rule winning nine times in ten from one winning ninety-nine times in a hundred, and four of its six cells sit in exactly that blind region.

Which says that adding seeds is the wrong repair, and says it with arithmetic rather than with a preference. Twelve seeds move the ceiling from a win rate of 0.89 to 0.94; twenty-four move it to 0.97. Each doubling of the run cost buys a few per cent of headroom, because the ceiling is a sixth root and doubling the exponent barely moves it. A sweep four times as expensive would still be saturated against a rule that wins nineteen runs in twenty.

This therefore has a cheaper test than the third exponent, and a better one than more seeds. Saturation is escaped by making the comparison harder rather than by measuring it longer — raise the amplitude until the deeper rule stops taking every run, or narrow the gap between the two exponents so its advantage is smaller to begin with. Either brings the middle panel’s cells off the ceiling, and only then does asking whether it has a corner mean anything at all.

Until one of them is run, the middle panel supports a single sentence: at this rise and this amplitude, the deeper rule wins across the whole correlation range.

The saturation reading also explains why the middle panel is the middle one, which the other two readings leave as a coincidence. If the deeper rule’s advantage varies smoothly along the ladder and happens to be largest somewhere in the middle of the range swept, then the panel where it saturates is the panel where the advantage peaks — and the two panels either side, where the advantage is smaller, are the ones with room to show a crossing. That would make the flat panel the strongest result of the three rather than the weakest, and it is a fourth reading nobody has tested. Everything else anybody would like to draw from its flatness is waiting on the saturation being lifted.

What the middle panel already rules out #

Whatever the explanation turns out to be, one thing is already excluded: the corner is not a property of the disturbance.

The colours, the amplitudes, the run lengths and the seeds are identical in all three panels — the same six seed values driving the same two rules against the same six disturbances. If the corner were something the disturbance carried — a property of an autoregressive process rather than of the arrangement it is applied to — it would appear identically in all three, and it does not appear in one of them at all.

That is a real narrowing. It leaves the corner as something produced by the interaction of a rule with an arrangement, which is where a mechanism would have to live, and rules out the tidiest alternative.

It also rules out the reading that was hardest to dislodge by argument. “Two organs is simply the shortest memory the run length can resolve” is a statement about the measurement apparatus, and no amount of reasoning about lattices touches it — only moving a parameter the apparatus does not depend on can. The rise is that parameter, which is why this sweep was the one that had to be run rather than another statistic computed over the sweeps already in hand.

Three cautions about the panel #

Six seeds is few. A majority of six is four, and several cells sit at four or five. The claim rests on the pattern across a row rather than on any cell, and the row is monotone in the finest panel and flat in the middle one — but a seventh seed could move a cell. Six is not arbitrary: it is the number at which the seed-to-seed spread stops dominating the difference between two rules, measured rather than chosen. It is enough to make a row’s shape readable and not enough to make a single cell a result, which is exactly how the panels are read here.

The correlation axis is coarse. Six values from zero to 0.99, unevenly spaced. A crossing between two of them would be located only to within the gap, and a crossing narrower than a gap could be missed entirely. The spacing is not uniform because the quantity that matters is the correlation length in organs rather than the coefficient, and length grows without bound as the coefficient approaches one — so equal steps in the coefficient would be wildly unequal steps in what the disturbance actually does. The axis is chosen to be roughly even in organs, which is the right choice for reading a crossing and the wrong one for claiming a crossing is absent.

And the middle rise is one rise. It is a single point in a rung, chosen because it carries the 5/8 pair — which is precisely the sampling this collection has learned to distrust. A second rise carrying the same pair would say whether the flatness belongs to the pair or to the rise, and it is the same cost as the third-exponent sweep. Running both would give a two-by-two — two exponent pairs at two rises on one rung — which is four sweeps and would settle the question in either direction rather than only in one.

Four sweeps is a number worth attaching a cost to, since the whole argument for running them is that they are cheap. One sweep is six correlations at six seeds for two rules, which is seventy-two settled stems; the two-by-two is two hundred and eighty-eight. That is the same order as a single rung sweep of the ablation census and rather less than one band, so it sits comfortably inside what this collection spends on a question it has already decided is the outstanding one. What it would cost more of is attention: four panels rather than three, each with its own saturation to check before its shape is read.

That third caution is the one nobody here would have thought to make a month ago, and it may be the whole answer. If the corner depends on which contact step is shorter, then a rung containing the reversal contains both behaviours, and sampling one rise from it gives whichever the rise happened to have.

Where this leaves it #

The corner thread now has one refutation, one narrowing, and one cell whose meaning is undetermined between two readings that a single sweep would separate.

Naming an undetermined cell and the measurement that would determine it is worth more than another statistic computed over the cells that are determined. The sweep is six correlations at six seeds at one rise with one new exponent, and it is the next thing this thread should do.

There is a temptation, with three panels and one of them flat, to report the two that agree and treat the third as noise. That would be the same error as reporting the offsets that wreck and not the ones that heal, and it is worth resisting for the same reason: the cells where nothing happens are cells, and a sweep that discards them is a sweep whose shape was decided before it ran.