88 vs the Lexicon
Can you score a decipherment?
We have been reading crop glyphs by function for a year — not hunting a secret alphabet, but asking what each mark does. Closure is stillness. The opening is where time enters. A saw-tooth ring is a ratchet: it turns and forbids reverse. Tangency is an optional door; penetration alters the core.
It works, in the sense that it produces readings. What it has never had is a grade. There is no answer key for a crop formation. Every reading is a proposal, and a language built entirely from proposals can drift a long way without anyone noticing.
So we borrowed an answer key from somewhere else entirely.
→ Open the working bench — all 88 classes rendered twice each (one struck cell, and a random field), grouped by the collisions. The numbers below are the summary; that page is the evidence, and the whole argument turns on being able to look at it.
The anchor
Wolfram's elementary cellular automata are the simplest interesting rule you can write: a line of cells, each black or white, each one looking only at itself and its two neighbours. Three cells means eight possible neighbourhoods, which means 2⁸ = 256 possible rules. That is the entire universe of them. Collapse the ones that are mirror images or colour-swaps of each other and 88 distinct classes remain.
Those 88 have properties the crop corpus does not: they are complete, they are enumerable, anyone can regenerate them from scratch, and none of them are anybody's opinion.
So: describe all 88 using only the glyph vocabulary. Two classes that end up wearing the same word are a gap — a distinction the language cannot make, with two pictures attached.
This is a law we learned the expensive way on a different project: a soft reader with no hard anchor is confidently wrong about 40% of the time, and a better-worded prompt made it worse. The fix is not more care. The fix is to point something checkable at it. The 88 are the checkable thing.
The projection — where all the judgment lives
The lexicon reads a mark stamped into a field. A rule is not a mark. So something had to decide what the mark is, and that decision is the only place judgment enters the whole experiment.
The choice: the glyph is the space-time raster the rule stamps. One cell struck at the top, and whatever grows below it.
| Lexicon mark | What it does | Measured as |
|---|---|---|
| Filled disc / thin ring | mass vs path | long-run ink on the page |
| Dots on a ring | quantised steps, a clock | cycle length on a 24-cell ring |
| Saw-tooth ring | the ratchet, directed time | reversibility, checked exhaustively |
| Open arc / non-closure | handedness from a closure defect | does the rule equal its own mirror |
| Concentric rings | depth / iteration levels | self-similarity at half scale |
| Radial spokes | broadcast from a source | how far one struck cell travels |
| Tangency / penetration | optional door vs alters the core | damage from touching one cell |
Argue with any single row and the whole result moves. That is the honest position, and it turned out to be the thesis of the entire exercise: the catalogue is cheap. The lens is the job.
Result one — the word
42 of 88 classes resolved. 48%. Sixty-seven classes fell into twenty-one collision groups.
But the number worth acting on was not the headline. It was this:
AXIS POWER — classes only this mark can separate
Tangency / penetration 13
Dots on a ring 8
Open arc 7
Filled / thin 6
Concentric rings 6
Saw-tooth 4
Radial spokes 3
The single most load-bearing mark in the language carries more than the next two combined — and its destructive half has never been read from an actual glyph. Tangency was read at Battlesbury in 2017. Penetration is still a placeholder, a meaning we guessed because the parser could detect the shape. The mark doing the most work is half-attested.
That is not a criticism. That is a work order.
A mark can be true and non-informative
The saw-tooth — the ratchet, the icon of directed time — came out validated as a law and nearly useless as a word.
Of the 256 rules, exactly six are reversible: 15, 51, 85, 170, 204, 240. Look at which six. One is the identity. One is the colour-swap. Two are shift-left and shift-right. Two are shifts combined with a swap. The only reversible rules are the ones that do no work — they reflect, or they translate, and that is all. The moment three inputs genuinely combine into one output, information dies and the arrow appears.
Which is the lexicon's first law, stated as a census: closure equals timelessness; the opening is where time enters. Six out of 256, checked exhaustively on every ring from four cells to twelve.
And as a distinguishing word it separates only four classes out of 88, because 84 of them are ratchets. It is right about the world and almost silent about which thing you are looking at. Those are different scores and they should be tracked separately.
Where the language objects to the maths
Standard practice collapses 256 rules to 88 by treating a rule and its mirror as the same thing, and black and white as interchangeable. The lexicon refuses both: the open arc is the source of handedness, and filled and thin are two different currencies — mass and path.
Measured: 52 of 88 classes are chiral, and 29 of 88 contain both a filled and a thin reading. The standard collapse destroys two distinctions the language treats as load-bearing.
So for our purposes the deck is 256, not 88. That is the language correcting the mathematics' bookkeeping, which is the direction nobody expects it to run.
Result two — the sentence
Then the obvious objection: 48% is a score for a character. Language strings marks into a sentence.
Fair. The first test flattened each rule's entire life into one word. But the corpus does not read that way. In July 2026 a formation at Foxhill, near Swindon put four distinct closed glyphs on one straight baseline — the first true sentence in the corpus, read left to right like writing rather than outward from a centre like a seal. Its four words: an opening target, a loom, a stripped seed, a closing target with an antenna and a scatter of progeny trailing behind it. Hold, weave, strip, emit.
So: stop flattening. Cut the raster into bands and string the clauses.
CONTROL — one clause, whole window, same three marks: 25 of 88 28%
clauses 1 2 3 4 5 6 7 8
mean 25.3 28.3 31.7 40.3 40.7 58.3 55.3 54.0
28% to 66% on identical data. Same window, same marks, same rules. The only thing that changed is that the line got cut into six. Replicated across three independent seed sets.
Why six: at that partition the first band is about one generation wide. The resolving information is concentrated in the early transient — the first few ticks, before the rules settle into agreeing with each other. Past seven clauses the bands fall below a single generation and it degrades.
That is the transferable pillar: when you string a system over time, spend your resolution at the start of the line, not the end.
Result three — a prediction, and its failure
Foxhill's own headline was not the four words. It was that the sentence does not return to where it started — it returns bigger. Closing target over opening target: 1.44 on the outer ring, 1.49 on the disc. And a formation at Avebury five days earlier had measured a voice-to-body clock running about ×1.5 per step.
Two independent formations, one ratio. So: prediction — a ruler the corpus measured will beat the arbitrary one we picked. Test it by holding the observation window fixed and changing only how the line is partitioned. Same information available; only the ruler moves.
ratio x1.00 x1.25 x1.40 x1.50 x1.60 x1.75 x2.00 x2.50 x3.00 x4.00
mean 29.3 34.7 37.3 34.7 38.7 34.3 34.0 35.0 40.3 36.0
The prediction failed. ×1.50 sits mid-pack, beaten by both ratios bracketing it and clearly by ×3.00, which won all three runs. Run-to-run spread is about ±3, so everything from ×1.25 to ×2.50 is one undifferentiated cluster. The corpus ratio does not transfer.
Recorded plainly, because a shape-library only stays honest if the misses go in the book beside the hits. One correction transferred (chirality). One did not (the ruler).
What did hold: geometry beats equality. Equal quarters scored worst of everything, barely above the unstrung control. How you cut the line matters enormously. What ratio you cut it at barely matters at all.
The encoding result
At every single ratio tested, first clause + the transitions resolved exactly as many classes as listing every clause. 28/28, 35/35, 39/39, every row.
Reporting one word and then how it grew costs nothing versus reporting all the words. Which is precisely what the Foxhill reading did — word one in absolute units, everything after it as a ratio. It is a re-encoding, not a summary.
And transitions without an anchor collapse to 10–22%. Growth needs a starting size. Anchor plus deltas turns out to be the minimum sufficient form, and the corpus arrived at it first.
Inflection versus vocabulary, with a number
Avebury strung three circles but repeated one word across time. Foxhill strung different words. The corpus drew that distinction in July. Here is what it is worth:
| classes | resolve to | rate | |
|---|---|---|---|
| flat — same clause all the way (inflection) | 53 | 18 distinct | 34% |
| moving — changes across the line (vocabulary) | 35 | 21 distinct | 60% |
A sentence that changes carries about 1.8× the resolving power of one that repeats.
What we banked
- Cut the line, and cut it fine at the start. 28% to 66%, free, on data you already have.
- Anchor plus deltas is the minimum sufficient encoding — absolute first term, ratios thereafter. Lossless and more compact.
- Repetition is cheap, change is expensive. If a strung reading repeats itself, it is inflection and carries little. The information lives at the transitions.
- The specific ruler does not matter. Having one does.
- A mark can be true and non-informative. Score correctness and discriminating power separately or you will over-weight your most confident word.
What this does not show
48% is not a score for "is the decipherment real." A wholly invented language could score 48% under a projection we chose ourselves. Nothing here is evidence that crop formations encode automaton behaviour, and the makers remain human as the working assumption throughout — the readings are of topology, which reads the same whoever laid it.
What the test yields is narrower and more useful than a verdict: a specific, pictured, prioritised list of where a vocabulary cannot resolve things. A to-do list instead of a feeling.
An instrument bug found along the way makes the point better than any argument. The first run scored 47% and filed a Sierpinski triangle in the same group as three plain diagonal threads, because a single straight line trivially contains a half-scale copy of itself and passed the self-similarity test. Nothing in the numbers flagged it. It was caught by looking at the pictures.
The catalogue was always fine. The projection was wrong until somebody looked.
Still open
- The two-dimensional run. Five marks in the lexicon could not be exercised at all — convergence on a hub, isolation at the rim, containment, the crescent, the graduated dial. A one-dimensional ring has no centre and no rim. Over a third of the vocabulary describes spatial relations a time-series has no room for, which says something about what kind of thing the language was built to read — and names the only test that could score the rest of it.
- Go and read a glyph for penetration. It is carrying the most weight in the language and it has never once been earned from a formation.