held · the principle that replaced it
Steady ground is the root of the request pulled whole, and seen to completion.
The instruction we set ourselves was plain. Confidence should not depend on how fuzzy the material was, or on how hard the base was to find. Fuzziness is the medium; difficulty is the work. Neither is ground. Ground is the root truth of the request, pulled whole, followed to completion, and returned clean.
The kernel cannot judge whether a reading was right. There is no model inside it, and we will not write one in — a similarity score invented in the arithmetic would be a measurement we made up. But there is something the kernel can compute without judging anything at all: whether the reading was accounted for.
held · checkable in the engine
Every look carries pieces, and each piece either lands or keeps riding.
A look does not arrive alone. It carries shards — the parts of what was seen — and each shard either lands under one of six headings, or it rides: kept, unplaced, offered back to whoever comes next with a place for it. A look whose shards have all landed has a complete trace. One with riders has open ends.
So the middle channel now reads how completely the trace is accounted for. It starts full. Every direction a look opened and left open costs it, and so does every place two looks part. It still only drains, and nothing can talk it back up — not repetition, not agreement, not averaging two readings you would like to see agree. Certainty is what you pay for what you left open.
And the disagreement between looks is now measured over the whole set of them at once rather than as each one arrives — so that part of the number reads the same in any order, and the queue at the middle is gone.
The word for this is not certainty. A trace with no open ends does not say the reading was right; it says nothing was left riding. Purity belongs to the trace, never to the answer.
held · the trial, run our way
Three engines, the same looks, and the mark set before the run.
three engines · the mark set first
We wrote down what the channel should read before we ran anything. Three fresh engines received identical looks at one concept. The first landed every shard it was handed. The second placed nothing and let everything ride. The third landed half. The predictions were specific: full, empty, and the point between them the arithmetic named.
It read as written. The engine that placed everything kept its whole saturation; the one that placed nothing spent all of it; the half-and-half engine sat exactly where it was predicted to sit. The light channel was identical across the three — the control, again, that says the difference was in the one channel and not in three different machines.
It could have gone wrong in five distinct ways, and we listed them first. Had the empty engine read as sure as the full one, or the half-and-half read anywhere but between them, the reading would have changed and this note would say so.