The ordering was not the actor

When an organ is removed from a stem and the arrangement never repairs, exactly one lattice hop stays rigid: the angle from an organ to the one p places above it is unchanged from the control, organ by organ, while every other lag moves. Which p is a question this thread has been chasing for four rounds.

Two answers have been ruled out and one holds at twenty-five rows of thirty. Ruled out, most stubbornly, is that the survivor is the shorter of the two contact steps — the reading the mechanism suggests, since a rule minimising a sum of inverse powers of distance ought to hold its nearest neighbours hardest. It scores twelve of thirty on the census, and the census turns out to be nearly one-sided in exactly the quantity the reading is stated over, so twelve of thirty is a weaker refutation than it looks.

This essay runs the test the census could not.

The design in one paragraph #

A band is a run of rises around the point where the two contact steps change places. Across it the counted pair is held, the settled divergence is held to a twentieth of a degree, the branch is held — and the ordering of the two steps reverses. Two of them were built, one on each branch: golden 5/8 across eighteen rises, Lucas 4/7 across nineteen.

Cut at every offset of every rise on both, compare each cut run against a control sharing its history to the last digit, and read the rigid lag out of the comparison. That is the same measurement the census makes, made on stems chosen to differ in one thing.

The result #

Fifty-five wrecked cuts. On the golden band, twenty-four of them, and the family left standing is the five at every one. On the Lucas band, thirty-one, and it is the four at every one.

Ten of the golden band’s rises have the five-step shorter and eight have the eight-step shorter. The family that survives is the five on both sides. On the Lucas band, seven rises have the four-step shorter and twelve have the seven-step shorter, and the family that survives is the four on both sides.

So the surviving family is the shorter step on part of each band and the longer step on the rest, and it is the same family throughout. The ordering reverses and the answer does not move.

How a survivor is read, and why it cannot be gamed #

The measurement in each cell deserves a paragraph, because a null result is only as good as the thing that was measured on both sides of it.

A cut run and its control share a history exactly: the same seed angles, the same grid, the same rise, the same organs, up to the moment one organ is removed from one of them. Both are then continued by the same rule for three hundred more organs. Every lag from one to twenty-four is compared between the two — how steady the hop is in the cut run, and how far its mean has moved from the control’s — and a lag counts as rigid when its hop is steady to within half a degree and unmoved to within three.

Nothing in that reads the counted pair, the step ordering, or the rise. It reads two node lists. So a cell of a band’s grid cannot be nudged toward one answer by the quantity the band varies, which is the property a matched design needs and does not always have.

The threshold is not delicate either, and that matters more here than usual. The rigid hops measure between zero and about a tenth of a degree of movement, and the next steadiest lag in any wrecked stem measures tens of degrees. There is no setting of the threshold between those two that changes any cell in either grid.

What that refutes, precisely #

Not “the shortest hop survives” as a claim about the census — that was already refused. What it refutes is the possibility that the ordering is the mechanism’s variable at all, on these lattices.

The distinction is worth spelling out. A reading can fail on a table for two reasons: because the quantity it names is not what acts, or because the table could not tell it apart from a rival. The census could not separate those two, because on twenty-five of its thirty rows the shortest-hop reading and the larger-family reading make the same prediction. A band separates them by construction, since the ordering flips while the family does not.

Scored on the bands, the shortest-hop reading is right at fourteen of the golden band’s twenty-four cuts and six of the Lucas band’s thirty-one: twenty of fifty-five, for an outcome that never changed. That number is not evidence about the mechanism at all. It is the fraction of rises on the side of the handover where the surviving family happens to have the shorter step, which is arithmetic about where the handover sits.

It is worth pressing that one step further, because the twenty of fifty-five is not merely uninformative — it is adjustable, and by the person who drew the band. Each band is a run of rises chosen around a crossing, and the choice of where to stop is a choice about how many rises fall on each side of it. Take three more rises off the coarse end of the golden band and the score falls; add three and it rises. Nothing about the stems, the rule or the outcome changes; only the endpoints do.

A quantity that moves when the design’s boundaries move and not when the stems change is not a measurement of the stems. Quoting it as a score against the shortest-hop reading would be quoting the band’s own construction back as evidence, which is the same failure the census made in the other direction — there, a number that looked like a refutation was a statement about which rises had been sampled. The band fixes that failure by making the answer independent of the sampling, and it would reintroduce it exactly by reporting a hit rate.

So the twenty is written down here to be dismissed rather than to be weighed, and the thing to carry away from the bands is the invariance and not the fraction.

What it does not refute #

The rise, which the band cannot hold. Each band spans a factor of about a fifth in the rise, so a quantity depending smoothly on the rise over that range is not controlled here. What can be said is bounded and worth saying: over a fifth in the rise, at a held pair and a held divergence, the answer does not change.

The offset, which is the reading that scores best and which the band leaves entirely alone. Every cell in both grids is at a fixed offset down a column, and the rule stated over the offset predicts every one of them correctly. The band is silent about it by design.

And the front, which changes by an organ or two across a band as the rise falls. That is visible in the grids as a ragged edge: at the coarse end of the golden band one offset wrecks, in the middle two or three, at the fine end two. Whether the extra offsets that appear are the ones that would have kept the other family is a question this design cannot answer, because they do not exist at the other end to compare with.

The thinness of it, stated #

Twenty-four cuts on one band and thirty-one on the other is fifty-five, and fifty-five is a decent number. But they are not fifty-five independent stems: the golden band’s twenty-four are eighteen rises times one or two offsets each, and offset four accounts for eighteen of them on its own.

So the strongest single statement available is about that one row. Offset four, eighteen rises, the ordering reversing in the middle, the five surviving at every one. The other rows agree and are shorter.

That is a narrower claim than fifty-five cuts implies and it is the honest one. It is still much stronger than anything a census could produce, because a census row is one stem and this is eighteen stems that differ in one quantity.

The predictions it was possible to make in advance #

Before the ablation was run, the band made three predictions and it is worth recording how they came out, because a design’s value is partly in what it committed to.

That the counted pair would hold at every rise. It does, at all thirty-seven, which is asserted in the machinery rather than checked by eye. Had it failed, the band would have straddled a transition and not a handover.

That the front would change by an organ or two rather than by several. It does. The wrecking offsets across the golden band run one, two, three, two, and across the Lucas band one or two. A front that changed by five would have made the grids incomparable end to end.

That the surviving family would either hold or flip at the handover. It holds, which is one of the two outcomes the design was built to distinguish. A flip exactly at the handover would have been the other, and would have made the ordering the actor rather than a passenger.

Three predictions, three confirmations, and the third is the result. Stating them first is what separates a test from a look.

How small an effect fifty-five cuts could have hidden #

A null is worth a bound, and the bound is short to compute.

Fifty-five cuts, no flip. Treating them as independent trials, the largest flip rate consistent with seeing none at the usual level is about three in fifty-five — under one cut in eighteen. So if the ordering does act, it acts at most that often, which is far too rarely to be the mechanism deciding a binary outcome that changes reliably with the offset.

The cuts are not independent, and the honest version of the bound has to say so. Eighteen of the golden band’s twenty-four are one offset repeated down a column of neighbouring rises, and stems at neighbouring rises share nearly everything. Read conservatively — one effective trial per offset per band, which is five or six — the bound loosens to about one in two, and says almost nothing.

The truth is between them and closer to the first, because what the column varies is precisely the quantity under test: eighteen rises spanning a reversal of the ordering are eighteen genuinely different values of the thing the reading names, even though they are one offset. A design that repeated one rise eighteen times would deserve the conservative reading. This one repeats one offset across the whole range of the variable, which is the opposite arrangement.

Where the question goes now #

Two quantities move along a rung. One of them has now been reversed under a controlled comparison and the answer did not follow it. That does not promote the other — the settled divergence — to the actor, because the divergence is exactly what a band holds.

What it leaves is a shorter list. The candidates for what varies along a rung and decides the surviving family are now the settled divergence, the front depth, and the rise itself. And the three are not independent of one another: the front deepens because the rise falls, and the divergence slides because the rise falls.

Separating those needs a different design, and the obvious one is a band around a divergence rather than around an ordering: two rises on different rungs whose settled divergences agree, with different pairs. Whether such pairs exist is arithmetic on the ladder, and the ladder sweep has the numbers to answer it without growing anything.

What the census still has that the band does not #

It is worth being even-handed, because the band is a better design and a much narrower one.

The census spans five counted pairs and both branches; the bands span two pairs. The census reaches offsets from three to nine; the bands reach three to five, because a band sits at one rung and a rung’s front has one depth. And the census contains the two lattices that never wreck at any offset, which are in it precisely because a table assembled from the lattices that produce an answer is a table of its own selection — a band has no equivalent, since every rise in it is chosen for its geometry rather than for its outcome.

So the two are complementary and the order they were done in is the wrong way round only in hindsight. A census establishes what the phenomenon is: that a wrecked stem keeps exactly one hop, that the survivor is one of the two contact families, that the offset accounts for most of it. A band tests a reading once there is a reading to test. Neither would have been much use first.

The negative that is worth having #

It is worth saying why a null result is the outcome this thread wanted rather than a disappointment.

The census’s decisive negative was a pair of runs sharing a pair and an offset and keeping different families, which closed a whole class of accounts: no function of the pair and the offset can produce both rows. That was strong precisely because it was a refutation with two rows in it rather than a score.

This is the same shape, one level down. It does not say what decides the survivor. It says that one of the two remaining candidates does not, on fifty-five cuts across two branches, with the quantity reversed rather than sampled. A shorter list of live accounts is what progress looks like when the mechanism is not going to hand anybody an answer.

A word about how this reads against the mechanism #

The reading that failed here was not arbitrary. It came from the rule: a sum of inverse powers of distance weights the nearest organ hardest, so the nearest family ought to be the one a disturbance cannot dislodge. That is the sort of argument this collection generally trusts, and it is worth asking what its failure means.

One reading is that the weighting is not as lopsided as the phrase suggests. At the exponents used here, the terms of the sum at the growing tip belong to lags across the whole contact neighbourhood, not to one organ — and the five largest terms belong to the pair, the number below it and the two above, at every exponent tried. A rule reading five families roughly equally has no reason to privilege whichever of two nearly equal steps is a per cent shorter.

The other reading is that the surviving family is not about the tip’s neighbours at all but about what the removal did to the arrangement above it, which is where the account that scores best lives. On that account the ordering was never a candidate, and its failure here is a confirmation rather than a surprise.

Both readings predict what the band found. Neither is tested by it, which is the usual position after a null: the list is shorter and the survivors are unranked.

What is left #

The four rungs with a computed handover and no band grown on them. Each would cost a few minutes of geometry and a few of ablation, and each would add a different counted pair to a result that currently rests on two.

And the direct test of the front, which deepens down a rung whether or not anything else does. The ragged edges of both grids say the front changes by an organ or two across a band; a design that held the front exactly — by choosing rises with the same number of wrecking offsets — would remove one more candidate. Whether such a set of rises exists inside a band is a question the existing sweep already answers, and it has not been asked.