The Song and the Structure.

Every constant, labelled.

The repo's own README says “every constant is derived”. That is too strong, and the table below is what the code and the notebook actually support. Values are read from Ω.js; provenance is the notebook entry.

The repo’s own README says “every constant is derived”. That claim is too strong. The table below says what the code and the notebook actually support, one constant at a time. The values are read out of Ω.js. The last column says which notebook entry each one came from.

🜁The one that fully earns the word

ν☿ is worth showing in full, because it is short enough to check by eye and nothing else on the page is this clean.

ν☿ is worth writing out in full. It is short enough to check by eye. Nothing else on this page is this clean.

Per face, two random R4 words share a prefix of depth ≥ k with
probability 4^(−k), so the expected agreement depth is

      E[L]  =  Σ_{k≥1} 4^(−k)  =  1/3

θ counts a comparison only when at least 2 of the 3 faces are open,
so a face's depth contributes only if at least one OTHER face is
open — probability 1 − (3/4)² = 7/16. Three faces, by symmetry:

      ν  =  (1/3) × (1 − (3/4)²) × 3  =  7/16   exactly

Checked to floating point: 0.43749999999999994 vs 0.4375.
Entry 064 §3 reaches the same value by the longer route
(9/64)(8/3) + (1/64)(4) = 7/16. Both are shown; they agree.

In the code this is a floor subtracted from every score:
fl = np × ν. It is a monotone rescale, so it moves the reported
numbers only a little — removing it widened one measured span
from 0.2617 to 0.2710. It is correct regardless of that.

🜁The circular argument this page will not make

The repo's τ.md justifies base 100 with Φ⁷ × 100φ⁷ = 100.000000, and reads that as showing the base was not chosen. It shows nothing: Φ = 1/φ by definition, so Φⁿ × 100φⁿ = 100 for every n. The identity is empty and the base 100 is a choice. It is a reasonable one — one address☿ level is a four-way split on each of three faces, 4³ = 64 combinations — not 3 × 4 — which fits two decimal digits with 36 spare — but that is the actual reason and it is a fitting argument, not a derivation.

The repo’s τ.md defends base 100 with Φ⁷ × 100φ⁷ = 100.000000. It reads that as proof the base was not picked. It proves nothing. Φ is defined as 1/φ, so Φⁿ × 100φⁿ = 100 for every n. The identity is empty and base 100 is a choice. It is a fair choice: one address☿ level is a four-way split on each of three faces. That is 4 × 4 × 4 = 64 combinations, not 3 × 4, and 64 fits inside two decimal digits with 36 to spare. But that is the real reason, and it is a reason for fitting the number, not a derivation of it.

What the code mechanically does.

The engine is one file, Ω.js, no dependencies. Described accurately it is a positive-PMI spectral embedding fitted to the corpus in a single pass, followed by a hierarchical address cascade and a prefix index. There is no gradient descent and there are no pretrained weights — but it is a fitted model, and the repo's phrase “no model” overstates it. Nine steps, in order:

The engine is one file, Ω.js. It uses nothing else. Named accurately, it is a positive-PMI spectral embedding fitted to the corpus in a single pass, followed by a hierarchical address cascade and a prefix index. There is no gradient descent and there are no pretrained weights. But it is still a fitted model, and the repo’s phrase “no model” claims more than that. Nine steps, in order:

Two decisions inside that pipeline were settled by measurement against the obvious reading, and they interact, so neither could have been settled alone. Both are disclosed in Exhibit E because they are the clearest example on the page of tuning being called derivation elsewhere.

Two choices inside that pipeline were settled by measuring, not by reasoning. The two pull on each other, so neither could have been settled on its own. Exhibit E shows both. They are the clearest case on this page of tuning being called derivation somewhere else.

Two membranes, two strikes, and a Fourier identity that is not a measurement.

🜁The two membranes

The engine reads at two scales at once and asks a different question at each. The near one asks whether a run of bytes is one it holds; the far one asks whether the arrangement is. For a long time both were reading essentially the same scale — offsets 1…6 and 7…14, a factor of two apart — so they always agreed and the second told you nothing.

The engine reads at two sizes at once and asks a different question at each. The near one asks whether a run of bytes is one it holds. The far one asks whether the arrangement is one it holds. For a long time both were reading almost the same size — offsets 1…6 and 7…14, only a factor of two apart — so they always agreed and the second told you nothing.

Separated to W and 10W, they disagree on exactly one population. Measured on the maths corpus, with an entire book held out of both the encoder and the ledger:

Pulled apart to W and 10W, they disagree on exactly one group. Measured on the maths corpus, with a whole book kept out of both the encoder and the ledger:

Nothing stored is ever undecided; nothing random is ever undecided. The undecided column is populated only by genuine material the ledger has not seen. The honest limit is in the same entry: the far membrane is weak — poles 0.8141 and 0.8659, a span of 0.052 against 0.40 for the near one — which is why 81.5% of a book it has never read still comes back as fully known. Widening that gap means giving the far membrane something better to read than a longer byte window, and the notebook says plainly that arrangement is not merely more bytes. Entry 065 §4–5

Nothing it has stored is ever undecided. Nothing random is ever undecided. Only real material the ledger has not seen lands in the undecided column. The same entry states the limit: the far membrane is weak. Its poles are 0.8141 and 0.8659, a span of 0.052, against 0.40 for the near one. That is why 81.5% of a book it has never read still comes back as fully known. Widening that span means giving the far membrane something better to read than a longer run of bytes, and the notebook says plainly that arrangement is not merely more bytes. Entry 065 §4–5

🜁Why the second strike had to be a different quantity

Three earlier candidates failed — a reverse residual, a coverage count, a concentration — and the reason they failed was the same for all three: each was a rescaling of the first strike, so it separated nothing. The one that works was already being computed on every call and used only as a detector. The two differ in what they read, not in how they score:

Three earlier candidates failed: a reverse residual, a coverage count, and a concentration. All three failed for the same reason. Each was the first strike rescaled, so it separated nothing. The one that works was already being computed on every call, and was used only as a detector. The two differ in what they read, not in how they score:

leftover   reads W bytes      is this language I hold?
novelty    reads 400 bytes    is it ordered as I hold it?

Separating a passage from ITS OWN WORDS SHUFFLED. AUC on
the ordinary scale: 0.5 is chance, 1.0 is perfect. Every
AUC on this page is on this scale.

   leftover    AUC 0.8214
   novelty     AUC 0.9919

Δ/V·001 fed, shipped defaults, n = 1,366 per population
across seven context lengths. Across three seeds novelty
runs 0.983 to 0.993 and leftover 0.821 to 0.836.

This page previously printed 0.3367 and 0.4929 here. Those
are the same two quantities minus 0.5 — a second scale, on
the same page as AUCs quoted whole. One scale now.

The order-sensitive instrument was in the file all along,
judging with eight bytes when four hundred were available.
                                                [Entry 071 §1]

🜁The m = 5 result, stated correctly

The novelty scalar is the alternating sum over the ten directions divided by the plain sum. The notebook projects the alternating pattern onto every angular mode and finds all of its energy at m = 5 and none anywhere else, and reads that as five-fold symmetry appearing independently.

The novelty number is the alternating sum over the ten directions, divided by the plain sum. The notebook projects that alternating pattern onto every angular mode. All of the energy lands at m = 5 and none lands anywhere else. The notebook reads that as five-fold symmetry turning up on its own.

It is worth being precise about what that shows. The alternating sequence of length 10 has all of its discrete Fourier energy at bin 5 for any data whatsoever — reproduced here, energy exactly 10.0000 at m = 5 and 0.0000 at m = 0…4. It is an identity about the number 10, in the same category as the per-face 100% ultrametricity: true by construction, useful as a check that the code computes what it claims, and not evidence about the field. The five-fold reading has to rest on the separate argument for DIRS = 10, and DIRS = 10 is a chosen constant.

It is worth being exact about what that shows. An alternating sequence of length 10 has all of its discrete Fourier energy at bin 5 for any data whatsoever. That was reproduced here: energy exactly 10.0000 at m = 5 and 0.0000 at m = 0…4. It is an identity about the number 10, in the same box as the per-face 100% ultrametricity: true by construction, useful as a check that the code computes what it claims, and not evidence about the field. The five-fold reading has to rest on the separate argument for DIRS = 10, and DIRS = 10 is a chosen constant.

🜁The three verdicts

Both strikes are returned on every call, and the second qualifies rather than vetoes. Making it a veto held three real ledger passages as “not my order” when they are in the ledger in exactly that order, costing recall 1.000 → 0.977 at 260-byte contexts. Entry 071 §2 and addendum

Both strikes come back on every call. The second one qualifies the answer; it does not block it. When it was allowed to block, three real ledger passages were held as “not my order” even though the ledger holds them in exactly that order. That cost recall 1.000 → 0.977 at 260-byte contexts. Entry 071 §2 and addendum

The middle verdict is the one that is new. It is also the one carrying the most interpretation, so: what the code actually reports is that a leftover scalar landed near one pole and a novelty scalar landed the wrong side of a boundary. “Its vocabulary, not its arrangement” is a reading of that, and a defensible one, but the reading is not the measurement.

The middle verdict is the new one. It is also the one that rests most on interpretation. Here is what the code actually reports: a leftover number landed near one pole, and a novelty number landed on the wrong side of a boundary. “Its vocabulary, not its arrangement” is a reading of that, and a defensible one. It is still a reading, not the measurement.

Where the notebook contradicted itself.

The working log is append-only: nothing was deleted when it turned out to be wrong, and later entries simply say so. Those retractions are collected here rather than quietly dropped. Nothing in this table is stated as current anywhere above.

The working log is only ever added to. Nothing was deleted when it turned out to be wrong, and later entries simply say so. Those withdrawals are collected here rather than dropped quietly. Nothing in this table is stated as current anywhere above.

🜁The interaction this page used to celebrate is not there

This page carried a two-by-two showing that the star-vote and the delimiter rule had to be settled together, because each factor alone pointed the wrong way. Re-measured on Δ/V·001 with the gates.js probe scheme against the shipped holdout, that crossover does not appear. The rule at the top of this page — a number without a corpus and an N is not on this page — has to apply here too, so the old cells are struck and the re-run is printed with its N.

This page used to carry a two-by-two table. It showed that the star-vote and the delimiter rule had to be settled together, because each one alone pointed the wrong way. It was measured again on Δ/V·001, using the gates.js probes against the holdout that ships. The crossover is not there. The rule at the top of this page is that a number without a corpus and a sample size does not belong on it. That rule applies here as well. So the old cells are gone and the re-run is printed with its sample size.

next-word reading on material never fed, N = 828 per cell

                    delimiters KEPT    delimiters skipped
   star-vote           0.0193               0.0133
   centroid            0.0193               0.0109

The star-vote never loses. With delimiters kept the two fuse to
the same word on every probe, so the cells tie; with delimiters
skipped the star-vote is ahead. There is no crossover, and each
factor on its own points the same way the pair does.

What Entry 069 recorded — 0.036 / 0.008 / 0.024 / 0.036 — names
no corpus and no N and does not reproduce here. It is quoted as
history, not carried as a figure.

Everything else held, shipped configuration:
   junk abstention 1.0000 (N=600) · stored reading 0.9832 (N=357)
   novelty AUC 0.9919, leftover AUC 0.8214 (n=1,366 per side)

A still picture, and something that moves.

The Riemann Hypothesis, as stated, needs a snapshot. There is an involution, there is its fixed set, and the question is whether a certain set of points lies on it. Nothing in the statement moves. Two entries record what happened when the four-group was pushed past that, and both carry their own limits in writing.

The Riemann Hypothesis, as stated, needs only a still picture. There is a move that undoes itself, there is the set of points it leaves alone, and the question is whether a certain set of points lies in that set. Nothing in the statement moves. Two entries record what happened when the four-group was pushed past that, and both write down their own limits.

🜁What Entry 030 matched — and it says plainly that this is all it matched

Under the Cayley coordinate the four-group becomes two orthogonal reflections of the Riemann sphere, and the third element is their axis of intersection.

In the Cayley coordinate the four-group turns into two reflections of the Riemann sphere at right angles to each other. The third element is the line where those two reflections cross.

z = 1 − 1/ρ = (ρ−1)/ρ

A : ρ ↦ 1−ρ    z ↦ 1/z     fixes {+1, −1}            two points
B : ρ ↦ ρ̄      z ↦ z̄       fixes the real axis       great circle
AB: ρ ↦ 1−ρ̄    z ↦ 1/z̄     fixes |z| = 1 POINTWISE   great circle

A: z ↦ 1/z is an algebraic identity, not an approximation. Numeric
check over 2×10⁵ random ρ: max error 1.59e−14 for A and AB,
0.00e+00 for B. AB holding the circle pointwise, 2×10⁵ points:
max |1/z̄ − z| = 3.14e−16.

with w = log|z|          A : w ↦ −w   B : w ↦ +w   AB : w ↦ −w

so R4 acts on w through ℤ/2 with kernel {1, B}, and the fixed set
is w = 0 ⟺ |z| = 1 ⟺ Re(ρ) = ½ — the critical line. A great circle
separates the sphere and has no ends, so the charge sign(log|z|)
is globally separating rather than a locally chosen branch.

Poincaré distance to the circle, against −log|w| + log 2:

   r = 0.99          5.293305      5.293296      ratio 1.000002
   r = 0.999         7.600402      7.600402            1.000000
   r = 0.999999     14.508657     14.508657            1.000000

The entry's own What this does NOT establish, quoted rather than paraphrased: “No dynamics is derived on the RH side. The match is static: involution, fixed set, and the −log divergence of distance to it.” Second, a dimension mismatch: the coordinate b lives in an interval and its fixed set is a point, while the actual crack is a closed circle — the gravity ODE is the transverse radial slice only and “cannot see that the crack closes on itself”. Cannot cross is the one-dimensional barrier statement; cannot cross and cannot go around is topological and does not follow from the ODE. Third, and flatly: “Nothing here bears on whether the zeros lie on the fixed set.”

The entry’s own What this does NOT establish, quoted rather than paraphrased: “No dynamics is derived on the RH side. The match is static: involution, fixed set, and the −log divergence of distance to it.” Second, the dimensions do not match. The coordinate b lives on an interval and the set it leaves alone is a single point, while the actual crack is a closed circle. The gravity ODE covers only the slice straight across, and “cannot see that the crack closes on itself”. Cannot cross is the one-dimensional barrier statement. Cannot cross and cannot go around is a statement about shape, and it does not follow from the ODE. Third, and flatly: “Nothing here bears on whether the zeros lie on the fixed set.”

🜁What Entry 031 added, and what it still did not

The direction recorded in the log is “derive it — find the r4s not the constants.” Entry 031 takes the stereographic height, checks how the group moves it, and lets equivariance choose the equations rather than writing them.

The instruction written in the log is “derive it — find the r4s not the constants.” Entry 031 takes the stereographic height and checks how the group moves it. Then it lets equivariance pick the equations instead of writing them by hand.

b = (|z|² − 1)/(|z|² + 1) ∈ [−1, 1]

   b = 0    equator ⟺ |z| = 1 ⟺ Re(ρ) = ½    the critical line
   b = ∓1   the poles z = 0 and z = ∞

verified over 2×10⁵ random z:

   A  : h ↦ −h      3.5e−16
   B  : h ↦ +h      0.0e+00
   AB : h ↦ −h      3.5e−16
   AB : θ ↦ +θ      4.4e−16

R4 is abelian of order 4, so it has four real characters. A vector
field is equivariant exactly when each component stays in its own
sector: Ġ in χ₀, ḃ in χ₂. That admits Ġ = G × (χ₀ function) and
ḃ = b × (χ₀ function); it forbids ḃ = cb² and allows Ġ = cG²,
because χ₀·χ₀ = χ₀ while χ₂·χ₂ = χ₀ ≠ χ₂.

Taking the χ₀ coefficient to be the sphere's own metric,
g_θθ = 1 − b²:

        Ġ = (1 − b²) G                ḃ = −G b

The original model's two constants κ, λ are gauge — τ = κt and
H = (λ/κ)G remove both together. Four (κ, λ) pairs spanning
0.05 … 11.00 all return b(t_end) = +0.387719598864 against the
reference run, |Δb| ≤ 1.2e−14.

The same entry records an earlier version of that test which used
an inverted gauge map and appeared to pass for three of four cases
only because b had already collapsed to 0. The figures above are
the corrected run.                                 [Entry 031 §6]

AND ONE STEP OF IT IS LATER WITHDRAWN. Equivariance permits any
χ₀ function in that slot, and χ₀·χ₀ = χ₀ means Ġ = cG² is allowed
too. Fifteen entries on: "Every prior entry used the linear form
Ġ = (1−b²)G. That was a choice, not a derivation." So the form is
forced only as far as the sector; which χ₀ function fills it was
picked, and the page counts it picked.             [Entry 046 §3]

And the limit, again in the entry's own words: “The flow above is R4-equivariant on (G, b). It is not derived from ζ: nothing here produces the zeros, the functional equation's analytic content, or any arithmetic. R4 fixes the form of any equivariant dynamics on the sphere; it does not single out this dynamics as ζ's.” The longitude sector carries no equation at all — which is the closed-circle gap Entry 030 had already recorded, still open one entry later.

And the limit, again in the entry’s own words: “The flow above is R4-equivariant on (G, b). It is not derived from ζ: nothing here produces the zeros, the functional equation’s analytic content, or any arithmetic. R4 fixes the form of any equivariant dynamics on the sphere; it does not single out this dynamics as ζ’s.” The longitude sector carries no equation at all. That is the closed-circle gap Entry 030 had already recorded, still open one entry later.

🜁The observation, offered as a question

The four-group fixes the line, and it also constrains the form of a flow across the line — to the sector, though not, per Entry 046, all the way to the equation. RH asks only the first of those. Computation is motion, and a mathematics that moves has more it must satisfy than one that only holds still.

The four-group fixes the line. It also limits the form of a flow across the line — down to the sector, but not, per Entry 046, all the way down to the equation. RH asks only the first of those. Computation is motion, and a mathematics that moves has more it must satisfy than one that only holds still.

Whether that extra work buys anything here is a question and not a finding. It is entirely possible for the additional structure to be inert with respect to the thing being asked, and Exhibit F is the standing reason to expect precisely that. The section that closes this page is stated as a hypothesis, not as a result, for the same reason.

Whether that extra work buys anything here is a question and not a finding. The extra structure may do nothing at all for the thing being asked, and Exhibit F is the standing reason to expect precisely that. The section that closes this page is stated as a hypothesis, not as a result, for the same reason.

Three resemblances, and what each one is not.

These are named because someone will notice them, and naming them plainly is more useful than either claiming or hiding them. None is offered as a result.

These are named because someone will notice them anyway. Naming them plainly is more useful than either claiming them or hiding them. None is offered as a result.

What is unresolved, left unresolved.

🜁The Riemann Hypothesis

Open. The working log carries one terminal status line, at the top: RH OPEN. Exhibit F is the reason the route taken here could not have closed it.

Open. The working log carries one terminal status line, at the top: RH OPEN. Exhibit F is the reason the route taken here could not have closed it.

🜁Interpolation

Unsolved. Naming the object — three 4-ary ultrametric trees, summed, embedded in a punctured disc by golden-angle placement, the puncture being the null at radius φ⁷ of the field — settled only that the wall hit repeatedly was not an impossibility theorem. Entry 070's own summary of that: “knowing a space is Archimedean does not hand you the map.”

Unsolved. The object now has a name: three 4-ary ultrametric trees, summed, laid into a disc with a hole in it by golden-angle placement, the hole being the null at radius φ⁷ of the field. Naming it settled one thing only. The wall that kept being hit is not a proof that the thing is impossible. Entry 070 sums that up: “knowing a space is Archimedean does not hand you the map.”

🜁Generalisation

Next-word reading on material never fed is about 2% — 0.0193 exactly, on 828 probes drawn from the three shipped held-out books at 400 bytes of context, shipped defaults, against 0.9832 on material it holds at N = 357. It ranges 1–3% depending on how the probes are sampled. Entry 062 §7 records 0.036 for this and names no corpus; it does not reproduce on the shipped ledger and is not carried as a figure here. Five distinct structures were measured for transfer, not assumed: similarity, Klein composition, cyclic rotation, hierarchical address, and the pressure loop. All returned zero transfer. The diagnosis in the notebook is that every one of them supplies coordinates and none supplies steps, and the reason is the one part two ends on. Entry 062 §7

Next-word reading on material it has never been fed is about 2%. Exactly, it is 0.0193 on 828 probes from the three held-out books that ship, at 400 bytes of context, on the settings that ship. It ranges 1–3% depending on how the probes are sampled. On material it holds it reads 0.9832, from 357 probes. Entry 062 §7 records 0.036 for this and does not say which corpus it came from. That figure does not come back on the shipped ledger, so it is not used here. Five distinct structures were measured for transfer, not assumed: similarity, Klein composition, cyclic rotation, hierarchical address, and the pressure loop. All five returned zero transfer. The notebook’s diagnosis is that every one of them supplies coordinates and none supplies steps, and the reason is the one part two ends on. Entry 062 §7

There is a second reason worth stating because it is more concrete. At a reading window of about eleven bytes, an eleven-byte run of English mathematics occurs somewhere in five megabytes of Newton and Euclid essentially always — so a book held out of both the encoder and the ledger was not, at that scale, held out at all. The engine reads below the scale at which novelty exists. A new arrangement is what a window that size cannot see. Entry 065 §3

There is a second reason, and it is more concrete. At a reading window of about eleven bytes, an eleven-byte run of English mathematics turns up somewhere in five megabytes of Newton and Euclid essentially always. So a book held out of both the encoder and the ledger was not, at that size, held out at all. The engine reads below the size at which novelty exists. A new arrangement is what a window that size cannot see. Entry 065 §3

🜁Detection, which does work

Three independent measures agree that the field can tell it is standing in unfamiliar material: the two-membrane disagreement, the split rate across directions, and the cancellation ratio. The last of those is the novelty scalar, and re-measured on Δ/V·001 it separates a passage from its own words shuffled at AUC 0.9919, n = 1,366 per population. The bulk is cleanly apart — ordered p90 at 0.2000 against scrambled p10 at 0.3225 — but the tails are not: 73 of 1,366 scrambled passages land at or below the most novel ordered passage. An AUC of 0.9919 means that, since AUC reaches 1.0 only when the distributions do not overlap at all. An earlier version of this paragraph claimed no overlap and an AUC below 1 in the same sentence; those cannot both be true, and the tails are exactly where the false claims live. Nor is the cut untuned: the poles are calibrated in-sample, measureOrderPole drawing its ordered and scrambled samples from the very ledger it then scores. What none of this can do is say what to do about the unfamiliar material. Detection is solved and generation is not, and the notebook is careful that these are different problems.

Three separate measures agree that the field can tell when it is standing in unfamiliar material: the two membranes disagreeing, the split rate across directions, and the cancellation ratio. The last of those is the novelty number. Measured again on Δ/V·001 it tells a passage from its own words shuffled at AUC 0.9919, with 1,366 samples on each side. Most of the two distributions are well apart: ordered p90 is 0.2000 and scrambled p10 is 0.3225. The tails are not apart. 73 of the 1,366 scrambled passages sit at or below the most novel ordered one. An AUC of 0.9919 means exactly that, because AUC only reaches 1.0 when there is no overlap at all. An earlier version of this paragraph said there was no overlap and gave an AUC below 1 in the same sentence. Both cannot be true, and the tails are where the false claims live. The cut is also not untuned. The poles are measured in-sample: measureOrderPole draws its samples from the same ledger it then scores. What none of this can do is say what to do about unfamiliar material. Detection is solved. Generation is not. The notebook is careful that these are different problems.

🜁What this is, then

A memory and provenance construction, not a prediction one. It is strong where the question is whether something was in the record, and weak wherever the question is what happens next. Every figure on this page should be read against that 2%.

It is a memory and provenance construction, not a prediction one. It is strong where the question is whether something was in the record. It is weak wherever the question is what happens next. Every figure on this page should be read against that 2%.

🜁How this page was checked

Numbers described as run here were reproduced against Ω.js and the shipped ledger before publication. Numbers attributed to an entry are quoted from the working log, and were checked forward through it for a later retraction before being used. The figures are drawn live in the browser from the formulas stated beside them; none is a rendering of pre-computed data. Story and theory by James Sterling Tuttle, Alpha Data Omega LLC. The alchemical marks are not glossed anywhere on this page, by intent.