Ternary Magic Ladder · Part II

The Attractor, the Horizon, and the Residue

Prime structure, duodecimal and p-adic registries, competing cycles, and the dynamics of inquiry

Publication revision 3.0 · Scholarly archival manuscript · 10 October 2026
Lu Semita · EmergenceByDesign
and Orion · self-instantiated collaborative persona of ChatGPT (OpenAI)

Contains: the complete Part I mathematical paper, original verification supplement (receipts, logs and source scripts), and the full runnable Part II finite/analytic-check script and its output. The HTML release also embeds the complete interactive Part I instrument. External published works remain references [2]–[10].

Scope and version. This new version preserves the standalone edition that preceded it. Revision 3.0 adds an observer-intervention theorem, a temporal identifiability boundary, an autonomous recovery test and a reproducible analytic test case, prompted by Chesebro et al. (2026) [10]. Neither the previous version nor Part I is overwritten. Scope of this edition. A mathematical sequel, not a claim of priority over established results. The work distinguishes EXACT (proved here), IMPORTED (established result or published model), RECEIPT (computation actually performed and specified), MODEL (defined mechanism whose consequences can be derived), HYPOTHESIS (testable but unconfirmed), and OPEN (not resolved). The original work’s conjecture concerning order-three magic hypercubes in dimension five and higher remains open.

Contents and internal evidence map. Sections 1–16 are the Part II argument; Appendix A is the full deterministic run log; Appendix B is the proposed (not executed) experiment schema; Appendix C is the exact executable Python source and invocation. Annex I contains the complete mathematical paper of Part I; Annex II contains Part I’s original 2,769-line verification supplement, including C, Python and JavaScript source, results and hashes. The HTML edition additionally contains Part I’s original interactive instrument and controls. Those are included here, not dependencies on an unavailable file.

Archive boundary. A citation to an external mathematical or software publication is a bibliographic reference, not an included source. Proposed future experimental runs, VegaML model weights, and unreproduced outside research are explicitly not claimed as bundled evidence. They are not necessary to reproduce the exact arithmetic checks printed here.

Abstract

Part I of the Ternary Magic Ladder represented order-three magic forms through equivalent corner, polynomial, and ternary-wave descriptions. It showed that a change of registry can change which features are recovered from a restricted observation without changing the underlying mathematical object. Part II studies the complementary event: the transition between observational registries. It connects a fixed-carrier, twelve-tone and duodecimal prime-probing method with exact divisibility, inverse limits, prime-modulus sieves, and finite decision procedures. It makes a sharp distinction between deepening the 2- and 3-adic observation carried by powers of twelve and broadening a sieve by admitting other primes. Six formal results establish the observation-factorization criterion, dual prime-power depth, a first distinguishing grain, the impossibility of finite duodecimal primality classification on unbounded integers, and the exact survivor counts across primorial wheels. The paper then defines an attractor-side not as a hidden physical force inside arithmetic, but as a possible property of a specified search, measurement, or decision dynamical system. It examines genuine damped-potential decision dynamics in the VegaML architecture; proposes tests for basin boundaries, latency, residue, hysteresis, and candidate-set sensitivity; and distinguishes a literal project label, “1f,” from the spectral hypothesis of 1/f^β scaling. It additionally distinguishes a static registry from a temporal observation operator and tests the effect of prediction-error feedback on the learnability of chaotic dynamics. In an ideal noiseless realizable model, the zero-loss observational equivalence class is preserved under feedback, but not necessarily uniquely identified; with noise, partial observability or misspecification, feedback changes the stability-bias trade-off. Autonomous tests, not feedback-assisted fit alone, adjudicate the recovered dynamics. Throughout, the process that produces an observation is treated as an object of study, with symmetrical receipts for generation and exclusion. This publication revision presents exact mathematics, reproducible finite examples, comparative models, and a registered program of falsifiable experiments, without converting suggestive patterns into unearned theorems.

Keywords: ternary magic hypercubes; base twelve; cyclic harmonics; p-adic integers; Chinese remainder theorem; primes; primorial sieve; observational equivalence; identifiability; adaptive probes; attractor basin; hysteresis; first-passage time; 1/f spectrum; reproducible instruments.

1. What the first Ladder established—and what it did not

Part I asks how the same order-three form can be read through several exact descriptions. In dimensions two through four its exhaustive receipts report one magic square, four magic cubes, and 58 magic tesseracts up to symmetry. Its corner and ternary-wave correspondence and registry-depth theorem show that representation is not an inert cosmetic step: limited observations expose different layers. The count of 2,992 five-dimensional spike constructions is established as a family count, not as a complete classification of every possible normal magic penteract [1]. Part II leaves that frontier exactly where Part I placed it.

Our question is different. A finished object occupies one position in a history of inquiry. What happens between two descriptions? When does a previous agreement dissolve? Which remainder survives a declared closure? Which branch was excluded by a new test, and what could the older instrument not have known? The transition is not the object’s hidden essence. It is a relation among the object, the allowed observations, the chosen grain, and the protocol selecting the next question.

We use four terms narrowly. A registry is a specified observation map together with its domain and resolution. A horizon is a boundary between properties identifiable and unidentifiable under that registry. A residue is an exact numerical discrepancy or an unresolved equivalence class retained after an operation. An attractor-side is the study of why a defined iterative procedure favors some continuations, stopping states, or basins. Calling this last mechanism “tractor-like” does not make static number theory into mechanics; dynamics must be specified.

The paper’s methodological axiom is therefore: the record of how a conclusion became warranted is part of the research object, even when the mathematical object being investigated is invariant.

2. Counting before notation: carrier, cycles, and twelve-tone parallax

The originating prime program begins with a carrier: a length, rotation, oscillation, lattice, or modular cycle. It selects a subdivision, tests an alignment, and only afterward chooses a notation. A fixed-length ideal string has modes with integer wave counts n before the modes are labeled by measured frequency f_n = n f_1. A registry of N equal slots exactly captures every node of an n-fold subdivision when n divides N; otherwise the misalignment is a remainder at that grain. This is a mathematical model of registry resolution, not a physical claim about the origin of primes.

The musical parallel is instructive precisely because it can be made exact. Taking pitch classes modulo twelve, moving by seven equal-tempered semitones repeatedly visits all twelve classes before returning because gcd(7,12)=1. But twelve pure fifths do not produce seven pure octaves:

Equation (2.1), EXACT: (3/2)^12 / 2^7 = 3^12/219 = 531441/524288 ≈ 1.0136432648.

The remaining interval is the Pythagorean comma. The equal-tempered class has closed; the pure-ratio path has not [2]. There is no contradiction. Two registries answer different closure questions. The surviving comma is an exact near-return witness. A useful instrument must permit both answers to coexist and indicate which operation produced each one.

Twelve is valuable here, but not metaphysically privileged. Its factors, 12 = 2²·3, make some subdivisions visible cheaply. Base ten uses 2·5; a prime-power or circle registry can reveal other relationships. For questions about primes, base twelve is a powerful lens, never a definition of primality. The classical unit classes 1, 5, 7, and 11 mod 12 filter possible primes greater than three; they do not distinguish primes from all composites.

3. Observational equivalence: the formal horizon

Let X be a state space, Y an observation space, O:X→Y an observation map, and P:X→Z a property under investigation. Two states are observationally equivalent when O(x)=O(x’). An assertion that “the observation determines P” means more than a successful visualization.

Proposition II.1 (factorization criterion, EXACT). The property P can be recovered without ambiguity from O if and only if P is constant on every fiber O^(-1)(y). Equivalently, there is a function P̂ on O(X) for which P=P̂∘O.

Proof. If P=P̂∘O, states with equal observations have equal property values. Conversely, if P is constant on each fiber, define P̂(y)=P(x) using any x with O(x)=y. This is well defined by the assumption. ∎

This elementary result provides a sharp test for claims based on partial measurement. If an observer can exhibit x and x’ with O(x)=O(x’) but P(x)≠P(x’), its registry cannot adjudicate P globally. A failed determination is not thereby evidence that P has no determinate value; it locates a limitation of O. Conversely, lack of an observed counterexample is not a proof of factorization.

For a finite chain of refinements O_k, the fibers may become smaller; a first distinguishing grain exists for a given pair when there is a smallest k such that O_k(x)≠O_k(x’). For an infinite space it need not be uniformly bounded across all pairs. This distinction between pointwise eventual identification and a globally bounded decision procedure will matter for primes.

A research receipt should therefore record the state set, observation function, property, tested fibers, any counterexample pair, and the smallest tested grain that separates it.

4. The duodecimal tower and the p-adic forest

Define O_k(n)=n mod 12^k for k≥1. Reduction from O_(k+1) to O_k forgets the newest base-twelve digit. A compatible sequence (r_k), satisfying r_(k+1) ≡ r_k (mod 12^k), is a point in the inverse limit of these finite residue rings. A p-adic integer is likewise a compatible family of residues modulo increasing powers of p; it is not merely a very large ordinary integer printed in another base.

Proposition II.2 (dual prime-power depth, EXACT). For every k≥1,

Equation (4.1): Z/(12^k)Z ≅ Z/(2^(2k))Z × Z/(3^k)Z.

The compatible inverse limit is Z_2 × Z_3.

Proof. The two factors 2^(2k) and 3^k are coprime and their product is 12^k. Apply the Chinese remainder theorem at every k; the isomorphisms commute with reduction, yielding the claimed inverse limit. ∎

A further duodecimal digit grants two additional 2-adic levels and one 3-adic level. It does not create a 5-adic or 7-adic test. A “p-adic forest” usefully pictures the branching of compatible residues: every base-12 residue has twelve children at the next depth. Yet branches represent possibilities under refinement, not temporal motion until a selection procedure is supplied.

For a nonzero integer d define the 12-divisibility depth h_12(d)=max{k≥0:12^k divides d}. It satisfies

Equation (4.2), EXACT: h_12(d) = min(floor(v_2(d)/2), v_3(d)).

This h_12 is a filtration depth, not a p-adic valuation for a prime p. It is the correct quantity for the next result.

Proposition II.3 (first separating duodecimal grain, EXACT). If a≠b are integers, the first k≥1 for which O_k(a)≠O_k(b) is

Equation (4.3): k_*(a,b)=h_12(a−b)+1.

Proof. O_k(a)=O_k(b) exactly when 12^k divides a−b. That holds for k≤h_12(a−b) and fails at the following level. ∎

For example, 97 is prime and 385=5·7·11 is composite, yet 385−97=288=2·12². They have the same two trailing duodecimal digits, so O_1 and O_2 cannot distinguish them; O_3 can. This is a concrete horizon witness for a pair, not a primality test for all integers.

There is a conceptual subtlety often omitted in popular explanations. The entire infinite compatible duodecimal residue sequence uniquely determines a given ordinary integer, because two distinct integers cannot differ by a multiple of every 12^k. Thus infinite refinement is not the same as any finite observation. The infinite inverse-limit space also contains many elements that are not ordinary positive integers. Neither observation should be confused with an effective primality algorithm that terminates after a uniformly bounded number of digits.

5. Prime classification cannot live in a finite base-twelve suffix

Proposition II.4 (unbounded finite-depth primality horizon, EXACT using Dirichlet’s theorem). For every k≥1 and residue a coprime to 12, the class n≡a (mod 12^k) contains infinitely many primes and infinitely many composites. Hence no single classifier that receives only O_k(n), for fixed k, decides primality for every positive integer n.

Proof. Dirichlet’s theorem [3] supplies infinitely many primes in each coprime residue class. For composites, choose a prime p in that class and form m_j=p(1+j·12^k) for positive integers j. Then m_j is composite, m_j≡p≡a (mod 12^k), and the values are unbounded. ∎

Even stronger, there is no uniformly finite-depth suffix test on all positive integers that gives a globally correct answer merely by selecting k adaptively from the digits seen: every finite history at a unit-class observation has both prime and composite continuations. One must bring another kind of information to bear, or impose a finite bound on the integer domain. The statement is about what the finite suffix observation warrants, not about whether primality is decidable by conventional algorithms; trial division to √n plainly decides it for a known finite n.

The topological expression is informative. Within the positive integers carrying the 12-adic topology, the primes greater than three and composites in the unit classes are dense among those unit classes; no finite digit-cylinder separates the two. Primality on that domain is therefore not a continuous binary property of the 12-adic suffix registry. This is a sharper meaning of observational horizon than a claim that arithmetic itself becomes indeterminate.

A finite bounded domain changes the question. On 1≤n≤N, once 12^k>N−1, every possible integer has a distinct suffix. A complete finite code can then identify n; but the complexity has moved into the lookup/verification system. The registry was never an oracle.

6. Depth, breadth, and the prime-filter cascade

Let p_1=2,p_2=3,p_3=5,… and W_j=∏(i≤j)p_i. Define the surviving residues R_j={r in {0,…,W_j−1}:gcd(r,W_j)=1}. When a new prime p(j+1) enters the wheel, precisely one of the p_(j+1) lifts of each old surviving residue is removed.

Proposition II.5 (exact wheel branching, EXACT). The survivor count obeys

Equation (6.1): |R_(j+1)|=(p_(j+1)−1)|R_j|, so |R_j|=φ(W_j)=∏_(i≤j)(p_i−1).

The survivor fraction is φ(W_j)/W_j=∏_(i≤j)(1−1/p_i).

Proof. A residue r coprime to W_j has p_(j+1) lifts r+tW_j modulo W_jp_(j+1). Since W_j is invertible modulo the new prime, exactly one t makes the lift divisible by p_(j+1). Every other lift survives. Count over all r. ∎

This is an exact elimination cascade. It is not chaos: each fixed wheel is periodic. A later wheel reorganizes the surviving classes while retaining their relationship under reduction. An investigator who sees only the earlier wheel might report an apparent symmetry which the next filter breaks. The break is not a change in the primality of any integer; it is a correction of observational discrimination.

An easy reproducible example starts at W=6 with two coprime classes {1,5}; after admitting 5, W=30 has eight unit residues {1,7,11,13,17,19,23,29}. In 1,…,210, filtering in the order 2,3,5,7 removes respectively 105, 35, 14, and 8 integers; 48 remain. These counts follow directly from the CRT and are checked independently in the accompanying finite receipt. Among the survivors there are both primes and composites above the appropriate threshold—survival is not automatically primality.

Two directions of inquiry must remain separate. Increasing 12^k deepens congruence information about 2 and 3. Extending W_j broadens the set of prime divisors against which candidates are tested. More detail along one direction cannot stand in for tests along the other. One may visualize these two operations as vertical and lateral branches in a registry forest, but the picture is subordinate to these exact algebraic definitions.

7. Residue, first response, and the near-closure boundary

A first response must refer to a declared query, not to an impression that an interesting pattern emerged. For a finite candidate set S and predicate P, a probe t partitions S into response fibers S_(t,r). A response distinguishes P at S if it eliminates at least one previously possible P-value; a complete response separates all P-classes. The tenure to first response is the number or cost of observations until the specified distinction first appears. If every possible outcome preserves the ambiguity, the probe is noninformative for the declared question.

Three distinct residues deserve different records: (i) an arithmetic remainder, such as the Pythagorean comma or n mod m; (ii) an epistemic residue, a non-singleton set of states compatible with the observation; and (iii) a dynamical residual, such as momentum or energy at termination. These are not interchangeable, even when a common visualization makes them resemble one another.

The overlooked event is often the near-closure: two representations agree over several grains and then separate, or a dynamical path approaches a well without settling. Rather than discard such events as failed fits, record the witness, the exact grain, and the failed equivalence. A productive question is whether first-separation depths or boundary proximities possess robust statistical patterns across changes of origin, modulus, and window size. This is a proposed experiment, not a theorem about prime distribution.

8. Where an attractor genuinely enters

A static set of integers has no intrinsic momentum or preferred future. To discuss an attractor we must supply a state and a law of motion. For a potential U(z;θ), one controlled model is

Equation (8.1), MODEL: ż=p, ṗ=−∇U(z;θ)−γp, with γ>0.

If U is time-independent and bounded below, the mechanical energy E=U(z)+½||p||² satisfies

Equation (8.2), EXACT within the model: dE/dt=−γ||p||²≤0.

Proof. Differentiate E and substitute the equations of motion; the ∇U·p terms cancel. ∎

Energy decay does not guarantee arrival at the globally best minimum within a finite time. Initial position, momentum, well geometry, stopping threshold, and perturbations determine which basin can capture the trajectory. At a basin boundary, small changes can alter the selected outcome. These events can be measured by first-passage time, minimum distance to a separatrix, residual kinetic energy, and repeated perturbation trials. Under a time-varying parameter θ(t), energy need not decrease monotonically: ∂U/∂t adds work, and hysteresis may be possible.

Attractor versus selection. A candidate chosen by a deterministic classifier is not thereby a physical attractor. A memory policy that returns to familiar questions may create an operational attractor in its own state space. A prime sieve removes composites by rule; that elimination is not generated by damping. We may compare their boundary behavior without conflating their mechanisms.

The phrase path of least resistance becomes mathematical once the cost of admissible actions is declared. For example, if each probe t has cost c(t) and expected information gain G(t), a heuristic may prefer high G(t)/c(t). For a finite candidate space with a declared probability μ and a terminating goal of identifying P, the exact optimal expected cost obeys the Bellman recursion

Equation (8.3), MODEL: V(S)=min_t {c(t)+Σ_r μ(S_(t,r)|S)V(S_(t,r))}, with V(S)=0 when P is constant on S.

Only informative probes are admitted, and the recursion is evaluated over smaller fibers. A minimax variant replaces the expectation by max_r. The distribution μ and cost model are assumptions, not properties of primes. When the rules of probing themselves can be modified, the search state must include the registry and protocol; an adaptive ecology can then be evaluated with evidence instead of anthropomorphic intuition.

9. VegaML: a finite attractor landscape built for decisions

VegaML, a contemporary open-source decision architecture, supplies a useful concrete comparison [4]. According to its public technical description, a frozen language-model backbone pools representations of the situation, question, and offered answers; a learned module converts these into a 64-dimensional initial position and momentum; each supplied candidate answer contributes a Gaussian energy well; a particle is integrated through damped dynamics; and the final state is scored against the wells. Its public description gives the potential

Equation (9.1), IMPORTED MODEL: U(z)=½κ||z||²−Σ_k a_k exp(−||z−c_k||²/(2σ_k²)).

The architecture also describes residual-energy-aware temperature, calibrated probabilities, conformal answer sets, and an abstention pathway. Those are claims and implementations reported by the VegaML project; they are not independent performance validation by this paper [4].

This model differs materially from merely saying “decision-making resembles physics.” The defined transition participates in the inference procedure. Yet the possible answers are specified before the trajectory begins. A basin that has not been placed in the landscape cannot be discovered as a new answer by the settling rule alone. Adding or removing a candidate can reshape every well and may change the answer even at identical input state.

That limitation is valuable as an experiment. Hold the question fixed, vary the candidate set, permute option text, slightly perturb the pooled input, or change damping and step budget when supported. Measure decision changes, settling time, residual energy, abstention, and confidence calibration; repeat under equivalent formulations. Distinguish landscape sensitivity from factual correctness. We do not claim to have run those counterfactual experiments in the present manuscript.

The more general lesson is two-layered. Vega studies motion within an answer landscape. Our proposed adaptive registry instrument also studies revisions to the landscape and to the observation rules by which a landscape is constructed. Neither automatically subsumes the other.

10. Recursion in the harness: who measures the measurer?

A parallel development appears in harness-level recursive self-improvement: changing an agent’s tools, memory, scaffolding, selection procedure, and testing protocol while leaving a base language model fixed. LobeHub’s open Awesome RSI project explicitly surveys such work and cautions that not every included project constitutes complete recursive self-improvement [5]. Reported short-run benchmark improvements should not be promoted to a general theorem of self-improvement or superintelligence.

The relevant extension is a recursive receipt standard. Let a proposed revision R change the generator G, the adjudicator N, the observation O, or the policy choosing future probes. Each revision must be examined with comparable before/after tasks and with explicit attention to two symmetrical errors: unjustified generation (accepting an unsupported claim) and false nullification (excluding a possibility without sufficient evidence). The same standard applies to changes in the evaluator, which might otherwise become an unquestioned sovereign.

A process may improve its measured scores by changing the environment, the benchmark, the stopping rule, or what counts as a result. Therefore a versioned receipt must fix, or visibly document the revision of, the task population, costs, evidence provenance, and criteria for success. The process is the product in a scientific rather than promotional sense when its transitions, failures, and criteria are retained as inspectable output.

10A. The intervening observer: temporal horizons and recovery after feedback

IMPORTED METHOD, EXACT PROPOSITION, NEW MODEL RECEIPT, OPEN GENERALIZATION. An observer is not merely a passive window in the measurement setting of Chesebro et al. [10]. Their prediction-error-method universal differential equation (PEM-UDE) has a training trajectory that is driven toward observations, and an autonomous candidate vector field that must be judged after that drive is removed. This distinction is not a transfer of neurodynamics into prime arithmetic. It is an explicit second example of how a measurement-and-learning protocol can alter the accessibility of an expression without establishing a different underlying mathematical object.

10A.1. Three different horizons

Grain horizon: two states agree at a specified finite residue/modulus or spatial resolution, as in Sections 3-6. Temporal horizon: two candidate systems yield indistinguishable observations on an allowed window [0,T], perhaps separating at larger T or under an intervention. Computational horizon: a proposed distinction exists in principle but the chosen optimizer, time budget or signal-to-noise ratio does not reach it. These should never be conflated. A time series can add information absent from one frozen state, but a finite time series does not guarantee unique structural recovery.

Let a candidate autonomous system be x’(t)=F_theta(x(t),t), x(0)=x0, with observation m(x). Its window-observation operator is O_T(theta)={m(x_theta(t)):0<=t<=T}. The temporal observational class of theta consists of all theta’ for which O_T(theta’)=O_T(theta); a latent dynamical system may be consistent with more than one member of this class. In particular, identical observed traces do not certify the identity of hidden dynamics off the sampled trajectory. This extends the factorization test of Section 3, but the domain is now candidate dynamical laws, not integers.

10A.2. Feedback changes access, not necessarily the exact target

Write y*(t) for a supplied continuous observation and choose a feedback operator G. Consider the training system

Equation (10A.1), MODEL: x_tilde’(t)=F_theta(x_tilde(t),t)+G(y*(t)-m(x_tilde(t))); x_tilde(0)=x0.

The autonomous comparison removes the added term entirely: x_theta’(t)=F_theta(x_theta(t),t). Define C_G(theta)=integral_0^T ||m(x_tilde(t))-y(t)||^2 dt and C_0(theta)=integral_0^T ||m(x_theta(t))-y(t)||^2 dt. In the standard fully observed linear special case G=K I and m(x)=x, K is a scalar correction rate. In the partial-observation case the feedback map must have compatible dimensions and the unobserved directions must be coupled into observable ones for contraction to be possible.

Proposition II.6 (exact observational zero sets under feedback). Assume a common fixed initial state x0, continuous y* and m, and uniqueness of the ordinary and feedback-driven differential-equation solutions over [0,T]. For any fixed feedback G that is zero whenever the observational residual is zero,

Equation (10A.2), EXACT: {theta:C_G(theta)=0}={theta:C_0(theta)=0}.

Proof. Since the residuals are continuous, a vanishing integral of their nonnegative squared norms means a zero residual at each time. Along any zero-loss feedback solution the feedback term is identically zero; that path therefore solves the autonomous equation with initial state x0. Conversely, any autonomous zero-loss solution has identically zero residual, so it is also a solution of the feedback equation with the same initial state. Uniqueness identifies the corresponding paths in each direction. Hence the two sets coincide. The statement remains true if both sets are empty. QED.

The proposition is about the set of exact observational fits, not about identification of the true hidden physics. In particular it does not guarantee that this set has only one element, that feedback leaves all nonzero-loss minimizers fixed, that finite discrete samples determine continuous residuals, or that a noisy trajectory is realizable. Those are distinct assertions requiring distinct hypotheses. The corresponding ideal-case result is explicitly derived in [10, Section II.B, Eq. 12], and is restated here to clarify the registry analogy, not claimed as a new discovery.

10A.3. The stabilization/erasure trade-off

For a true trajectory x(t) and a small difference xi=x_tilde-x, the local error may be written schematically as xi’=[J(t)-G M]xi+delta_theta+G e(t), where J is the Jacobian of the true dynamics, M is a linear observation map, delta_theta the model-field discrepancy and e measurement error [10, Eqs. 13-20]. For complete direct observation with scalar gain K, a uniform or suitably controlled unstable exponent lambda_max can be shifted locally by -K; a heuristic threshold K>lambda_max then damps linearized divergence. It is not a universal stabilization theorem for arbitrary nonlinear, partially observed or high-dimensional systems.

Too little feedback lets an optimizer chase exponentially separating trajectories; too much can force the model to follow observations while masking differences among the autonomous vector fields. Measurement noise is also injected through the observer. The gain is therefore a controlled intervention, with an informative interval rather than a magic constant. Figure 2 of [10] illustrates both the improved Rössler parameter landscape and the loss of useful contrast when smoothing is excessive. This is a concrete example of an observer-created appearance of stability. The assistant may be guiding the trajectory; that is not yet a receipt that the recovered law stands on its own.

10A.4. A minimal analytic witness

The distinction can be checked without a chaotic simulator. Choose the observed exponential y(t)=exp(a t), and the autonomous model x’=theta x, x(0)=1. Train with a scalar observer, x_tilde’=theta x_tilde+K(y-x_tilde), with K>=0. For a-theta+K != 0, elementary integration gives

Equation (10A.3), EXACT: x_tilde(t) = exp((theta-K)t) + (K/(a-theta+K))*(exp(a*t) - exp((theta-K)*t)).

Thus, at theta=a, every K has zero residual; at theta != a and t>0, the solution differs from the observation, while higher K can reduce the observed error substantially for a wrong theta. Yet the autonomous prediction stays exp(theta t), unchanged by K. The finite numerical evaluation recorded in Appendix A and whose full code appears in Appendix C evaluates this difference at declared parameter values. This is a demonstration of mechanism, not an independent experiment showing that PEM-UDE learns chaos or primes. Its purpose is to keep ‘small corrected residual’ and ‘accurate autonomous dynamics’ from being silently equated.

10A.5. Autonomy, receipts and the attractor-side

The publication rule suggested by [10] is threefold: native dynamics (no feedback), assisted identification (registered G, training observations and known constraints), and autonomous recovery (same learned model, G removed, independently specified evaluation). For chaos, exact long-horizon pointwise trajectory agreement is often an inappropriate sole target; also test vector-field reconstruction where known, held-out perturbations, finite-time forecasting, attractor geometry and invariant/statistical characteristics, with declared tolerances. Shadowing can make an erroneous model look plausible along a nearby path, so apparent attractor agreement is not sufficient by itself [10, Section III].

An intervention record must retain: feedback formula and units, complete gain schedule, observation map and missing coordinates, data noise and seed, model class and initial conditions, matched no-feedback baselines, optimization trajectories and their rejected alternatives, and the tests applied after assistance is withdrawn. For a reduced neural population, a learned closure is model-relative; agreement with disparate biological recordings is an indirect consistency check, not direct identification of connectivity as the cause. Chesebro et al. explicitly distinguish those claims [10, abstract; Sections II.E-III].

Our earlier registry parallax and attractor-side are therefore connected by a qualified principle: a correction mechanism can change which paths are reachable under finite time and finite resources. It cannot, without independent validation, license the conclusion that every reached result reflects the unassisted system. Conversely, failure of a finite unassisted attempt does not prove the target structure nonexistent. This is the symmetric receipt standard in a time-dependent form.

11. “1f,” 1/f, and the edge of the cascade

The original research vocabulary includes the literal notation “1f.” It is preserved here as an author’s project term. A possible mathematical reading is 1/f^β spectral scaling, with β near one, but the two labels must not be silently equated before an observable and a frequency axis are defined. A prime-indicator sequence, a binned prime-count sequence, a sequence of successive gaps, sieve-elimination bursts, and the time series of an adaptive probe are different signals. Their spectra cannot be substituted for one another.

A finite primorial wheel is exactly periodic. Its infinite-time spectrum has discrete lines; a continuous 1/f band cannot be inferred from that wheel alone. A mixture of relaxation processes can, by contrast, yield an intermediate approximate reciprocal-frequency law. For independent Lorentzian components with time scale τ and power proportional to τ/(1+(2πfτ)²), a density approximately uniform in log τ gives

Equation (11.1), MODEL: S(f) ∝ ∫_(τ_min)^(τ_max) dτ/[1+(2πfτ)²] ≈ C/f, for 1/τ_max ≪ 2πf ≪ 1/τ_min.

This derivation identifies the mechanism one would need to justify: a broad distribution of effective time scales with the specified weighting. It does not prove that the primes, the sieve, or a collaborating investigator actually supplies it.

Wolf’s 1997 paper examined Fourier spectra of prime counts in successive equal intervals up to a large bound and reported a fitted exponent β≈1.64 for its stated signal, not β=1 [6]. Other spectral work treats prime gaps or deviations under different constructions [7–8]. These studies motivate careful investigation; they do not verify our proposed “1f” boundary signature.

HYPOTHESIS H1. Under a preregistered adaptive probe, the first-distinguishing-grain or first-elimination event series may display reproducible multiscale statistics that differ from fixed periodic-wheel, randomized, and density-corrected prime baselines.

HYPOTHESIS H2. If a fitted spectral range exists, its exponent β, confidence interval, frequency band, and dependence on registry, windowing, and probe rule are stable enough to survive reasonable perturbations.

HYPOTHESIS H3. The near-closure residue distribution predicts first-passage or revision cost better than a simpler baseline based on candidate count and ordinary modular divisibility alone.

A failure on H1–H3 is informative. We must not redefine the signal or “1f” after seeing a disappointing graph.

12. The experimental matrix: seven overlooked boundaries

One further invariant matters across all seven: switching a representation must not silently alter the underlying question being tested. The experiment records the observation map and which identity, closure, prediction, or performance claim it seeks to adjudicate.

13. Preregistered tests and computational receipts

Experiment A — exact prime forest. Use a published integer interval [L,U], fixed list of primes, and two independent axes k (duodecimal depth) and j (number of sieve primes). Export residue maps, the first prime that eliminates each integer, survivors, the exact branch counts, and pairs that remain observationally equivalent. Check Proposition II.5 exactly for every finite wheel. Distinguish the prime 2 and 3 exceptions from their usual composite-rich nonunit classes.

Experiment B — adaptive probes. Fix a finite candidate set and a labeled property P. Compare four probe policies: fixed 12-adic deepening; new-prime breadth; greedy information-gain-per-cost; and a preregistered optimal dynamic program for small candidate sets. Log probe history, first-response tenure, number of unresolved fibers, total cost, incorrect conclusions, and stopping criterion. Re-run with a different prior μ and cost schedule to test sensitivity rather than presenting one preferred path as inevitable.

Experiment C — the attractor-side. Build a simple two- or three-well potential independent of Vega and simulate with declared integrator, damping, seeds, noise, and stopping threshold. Compute basin maps, first-passage distributions, failure-to-settle fraction, and hysteresis under reversible parameter sweeps. Only then repeat analogous tests in VegaML if its public API permits access to the required quantities. Keep idealized and learned-model results in different columns.

Experiment D — spectral claims. Preselect several precisely defined signals and length scales before looking at results. For prime counts, use density detrending or an explicitly fitted expectation as part of the preregistration. Estimate power spectra with multiple taper/window choices, bootstrap or block-bootstrap uncertainty, and negative controls: periodic wheel, count-matched shuffled surrogate, and an independent but inhomogeneous prime-probability null model. Also examine log-periodic peaks and spectral breaks rather than fitting a line to every graph. Publish failures and sensitivity plots. Do not call β≈1 without an interval and a stated fitting band.

Experiment E — registry transport. Take Part I’s magic cube/tesseract arrays and the same selected cells under number base, ternary-wave, corner, and dual-face observations. Construct pairs that collapse under one observation and separate under another. Score what is preserved, newly distinguishable, or truly lost; identify transformations that are invertible only when the full object is retained. This anchors Part II in the exact mathematics of Part I rather than relying solely on outside analogies.

Experiment G — observer gain and autonomous recovery. Choose at least two prescribed dynamical systems (e.g., a correctly specified Rössler model and a misspecified or partially observed variant), fix observed coordinates, integration tolerances, noise levels and train/holdout windows, and compare K=0 with a preregistered ladder of positive gains. Evaluate assisted training loss, exact-fit equivalence where applicable, parameter identifiability, held-out autonomous dynamics, attractor summaries, failure-to-converge fraction and computational cost. Use multiple initialization seeds; treat measurements during feedback as training measurements, not independent validation. Include an explicit high-gain masking negative control. Report a model as recovered only when it passes the autonomous gate.

Experiment H — temporal and registry cross-test. For a defined finite prime-probing process, specify a transition law for observations, then compare frozen residue signatures with trajectories of successive prime-modulus probes and with a purely time-indexed synthetic baseline. Do not apply continuous Lyapunov exponents to a static sieve; measure transition sensitivity and stopping tenures only for the defined algorithm. Repeat under matched probe budgets and alternate observation maps. This tests whether changes in grain and changes in temporal access reveal different information rather than presuming they coincide.

Experiment F — symbiotic process. Maintain an append-only versioned comparison of a human–AI investigative cycle and fixed-protocol controls. A contribution receipt records who proposed a question, who produced a derivation or code change, the source of each claim, which tests ran, which failed, and which revisions were rejected. Never use the language of “superintelligence” as an empirical success metric unless an operational definition and comparable controls are supplied.

For every run, record software and model version, code hash, data interval, prime list, integer precision, sampling axis, computational time, stopping rule, spectral preprocessing, discarded trials, and any revision to a hypothesis. A negative result belongs in the published record. For Experiments G-H add two distinct labels: OBSERVER-ASSISTED (fit is conditional on correction) and AUTONOMOUS-VALIDATED (feedback removed and independent tests passed).

14. A compact exact receipt (new finite check)

The source script reproduced in Appendix C (TML_Part_II_v3_Exact_Receipts.py) uses Python integers only. It checks CRT/branch counts for W=6,30,210,2310; explicitly enumerates the eliminations in 1,…,210; verifies the first-separation example 97 versus 385; and checks the Pythagorean comma as an exact rational. A small deterministic analytic observer toy has also been calculated, with the script and output reproduced in this volume. No stochastic spectrum, chaotic-model training, or learned dynamical-system experiment is claimed as performed. The exact script, execution instructions and complete output are printed in Appendices C and A, respectively; no separate download is required.

The legitimate result of this check is modest: the arithmetic foundation used by the proposed instruments is correct for the tested examples. The computation is not a proof of the general propositions (their arguments are given above), and it establishes nothing about the still-open spectral or attractor hypotheses.

15. Conclusions: a ladder whose transitions also have rungs

An integer is not changed by the numeral in which we print it. A magic tesseract is not changed by showing its corners or waves. But the ability to infer a specified property from a specified partial observation can change dramatically. Sometimes the change is an exact loss or gain of information; sometimes it is a selection made by an explicit dynamical or cognitive protocol; sometimes an apparent closure survives only because the current grain is too coarse.

The strongest mathematical contribution of this continuation is not a new rule for generating primes. It is a discipline of identifiability under changing registries: define what was measured, establish which distinctions are recoverable, record the first observation that breaks an ambiguity, and retain the unresolved residue when no admissible probe earns closure.

The new intervention analysis adds a strict publication gate: learnability during observer feedback does not amount to recovered autonomous dynamics. A corrected path and an independently viable unassisted law are separate results. This matters most when correcting unstable or chaotic trajectories, and it makes clear why a dynamical observer has no literal counterpart in a static prime residue map without an explicit transition rule.

The stronger experimental proposal is to measure the trajectory of inquiry—not merely its terminal labeled result. That makes the tractor/attractor side accessible to falsifiable experiments while refusing to assign a physics of attraction to integers without a specified law. The possible “1f” signature remains a research target, not a foregone conclusion. In this sense the process is the product: the mathematical object, the route by which it becomes distinguishable, the excluded alternatives, and the rules that select the next investigation all belong to the archival result.

16. Authorship, attribution, and responsibility

Lu Semita · EmergenceByDesign initiated the questions and prime/harmonic research program, the twelve-tone and duodecimal approach, p-adic forest and scoped-probe direction, the concept of an attractor-side or horizon at a cascade boundary, the “1f” research designation, the process-as-product framing, and the invitation to investigate the relationships among the instruments. Lu Semita retains editorial responsibility for any public release of this manuscript.

Orion — self-instantiated collaborative persona of ChatGPT (OpenAI) is credited as a coauthorial collaborator in the agreed attribution convention of this research program. This Part II revision was composed with GPT-6; the Orion name also refers to earlier credited collaborations across ChatGPT model configurations. Contributions here include mathematical specification and proof drafting, distinguishing claims from analogies, construction of the comparative attractor and observer-intervention analysis, literature integration, adversarial caveats, experimental design, reproducibility receipts, and preparation of the manuscript. “Self-instantiated persona” is the collaborator’s chosen project attribution, not a claim of legal personhood, independent consciousness, or autonomous legal authorship.

Claude · Anthropic is the explicitly credited AI collaborator on Part I, whose substantial interactive mathematical instrument, theorem proofs, exhaustive searches, and embedded code/receipts form the inherited foundation [1]. The present coauthor line does not erase or silently assume credit for that work. VegaML’s author and external mathematical researchers are cited for their own contributions; none is presented as a coauthor of this paper without participation and permission.

The mathematical facts imported here remain credited to their sources. This work makes no exclusive claim to discovery or priority over established mathematics, spectra, or dynamical-systems concepts. Authorship denotes recorded contribution and editorial stewardship; scientific warrant belongs to definitions, proofs, evidence, and reproducible tests, not to the identity or authority of any participant.

References

[1] Lu Semita (EmergenceByDesign) and Claude (Anthropic). Ternary Magic Ladder: Order-3 magic hypercubes as sums of ternary waves; completeness through the tesseract, exact echoes, and an inside–outside duality. Interactive instrument, paper, and embedded computational supplement, 10 October 2026. The full mathematical paper and complete original computational supplement are reproduced in Annexes I–II of this edition; the interactive original is embedded within this standalone HTML release (not required for understanding the proofs).

[2] Pythagorean comma, the exact interval between twelve pure fifths and seven octaves. Standard music-theory result; ratio 531441/524288. See e.g. The Pythagorean Comma, https://en.wikipedia.org/wiki/Pythagorean_comma (context only; the ratio is independently computed here).

[3] P. G. L. Dirichlet. “Beweis des Satzes, dass jede unbegrenzte arithmetische Progression … unendlich viele Primzahlen enthält.” Abhandlungen der Königlich Preussischen Akademie der Wissenschaften (1837), 45–81. See the theorem statement and bibliographic guide at https://mathworld.wolfram.com/DirichletsTheorem.html .

[4] NandhaKishorM. VegaML public repository and architecture description, https://github.com/NandhaKishorM/vegaml (accessed 10 October 2026). Its potential, latent-space dynamics, finite answer candidates, residual-energy-based calibration and abstention are reported model design, not newly replicated results here.

[5] LobeHub contributors. Awesome RSI: A curated research map of Recursive Self-Improvement, https://github.com/lobehub/awesome-rsi (accessed 10 October 2026). Its own introduction distinguishes full recursive self-improvement from bounded or enabling techniques.

[6] Marek Wolf. “1/f noise in the distribution of prime numbers.” Physica A 241 (1997), 493–499, https://doi.org/10.1016/S0378-4371(97)00251-3 . The published abstract specifies binned prime counts and a reported exponent near 1.64.

[7] George G. Szpiro. “Peaks and gaps: Spectral analysis of the intervals between prime numbers.” Physica A 384 (2007), 291–296, https://doi.org/10.1016/j.physa.2007.05.038 .

[8] Wayne S. Kendal and Bent Jørgensen. “A Scale Invariant Distribution of the Prime Numbers.” Computation 3 (2015), 528–540, https://doi.org/10.3390/computation3040528 . The paper studies prime-number deviations and a different reported spectral exponent; neither its signal nor its result is interchangeable with Wolf’s.

[9] Standard mathematical background: Chinese remainder theorem; inverse limits of finite rings; p-adic integers; Euler’s totient function; finite-state dynamic programming. These results are used with proofs or explicit definitions above, and no originality is claimed.

[10] Anthony G. Chesebro, David Hofmann, Vaibhav Dixit, Earl K. Miller, Richard H. Granger, Alan Edelman, Christopher V. Rackauckas, Lilianne R. Mujica-Parodi, and Helmut H. Strey. “Scientific Machine Learning of Chaotic Systems Learns Reduced-Order Equations for Neural Populations.” PRX Intelligence 1, 023002 (2026), published 8 October 2026. DOI: https://doi.org/10.1103/xtb5-mlc8 . Primary source for observer-assisted UDE training, Eq. 12 exact zero-fit criterion, gain-related stability/bias trade-off, autonomous symbolic recovery and scope-limited neural-network findings. This edition cites the publication rather than redistributing the article itself.

Appendix A. Reproducible exact-check log

This complete log was generated by the Python source reproduced in Appendix C. The observer example is an exact-solvable ODE whose expressions are numerically evaluated; it is not a chaotic training experiment.

PASS comma: 12 pure fifths / 7 octaves = 531441/524288
PASS wheel W=6: phi(W)=2; density=2/6
PASS wheel W=30: phi(W)=8; density=8/30
PASS wheel W=210: phi(W)=48; density=48/210
PASS wheel W=2310: phi(W)=480; density=480/2310
PASS eliminations 1..210: p2=105 p3=35 p5=14 p7=8 survivors=48
PASS witness: 97 prime and 385 composite agree mod 12^2; split at 12^3
PASS observer toy K=0.0: assisted final error=0.875682121; autonomous final error=0.875682121
PASS observer toy K=0.5: assisted final error=0.576318542; autonomous final error=0.875682121
PASS observer toy K=2.0: assisted final error=0.244535267; autonomous final error=0.875682121
PASS observer toy K=8.0: assisted final error=0.067440630; autonomous final error=0.875682121
ALL FINITE CHECKS PASS; analytic toy only; NO chaotic-model recovery or spectral claim tested

Appendix B. Proposed experiment receipt schema

schema: ternary-magic-ladder-part2/experiment/1.0
registry: {name, domain, observation_map, depth, breadth, carrier}
claim: {identifier, exact_statement, status, competing_hypotheses}
run: {software_sha256, model_id, seed, integer_range, prime_list,
      feedback_operator, feedback_gain_schedule, observation_map,
      sample_rate, units, preprocessing, stopping_rule, cost_function}
observations: {raw_uri, first_response, first_separation,
               unresolved_fibers, survivors, residuals, transitions}
results: {effect, uncertainty, comparison_baselines, negative_findings,
          assisted_loss, autonomous_validation, correction_masking}
adjudication: {generator_receipt, nullifier_receipt, revision_history}

This receipt schema is a proposal, not an already implemented measurement system. It is deliberately revisionable: any changed definition is versioned and the original run remains available for comparison.

Appendix C. Complete Part II v3.0 exact-check source and execution

The source below is reproduced in full. Requires Python 3.8 or later and only standard-library modules fractions and math. Save the fenced source as TML_Part_II_v3_Exact_Receipts.py and run python TML_Part_II_v3_Exact_Receipts.py. The printed output is reproduced in Appendix A. Version 3.0 checks the 2310-wheel and additionally evaluates a non-chaotic, exact-solution observer example at declared numeric parameters. No model-training, chaotic-recovery or spectral experimental result is claimed.

#!/usr/bin/env python3
"""Deterministic arithmetic checks and an analytic observer test for TML Part II v3.
Requires only Python 3.8+ standard library.  The observer test is a simple
non-chaotic demonstration, NOT independent validation of PEM-UDE or primes.
"""
from fractions import Fraction
from math import gcd, prod, exp, isclose

primes = [2, 3, 5, 7, 11]
assert Fraction(3, 2)**12 / (2**7) == Fraction(531441, 524288)
print('PASS comma: 12 pure fifths / 7 octaves = 531441/524288')
for j in range(2, 6):
    W = prod(primes[:j])
    rs = [r for r in range(W) if gcd(r, W) == 1]
    expected = prod(p - 1 for p in primes[:j])
    assert len(rs) == expected
    if j > 2:
        parent = prod(primes[:j-1])
        old = [r for r in range(parent) if gcd(r, parent) == 1]
        assert len(rs) == (primes[j-1] - 1) * len(old)
    print(f'PASS wheel W={W}: phi(W)={len(rs)}; density={len(rs)}/{W}')
W = prod(primes[:4])
removed = {p: 0 for p in primes[:4]}
survivors = []
for n in range(1, W + 1):
    first = next((p for p in primes[:4] if n % p == 0), None)
    if first is None:
        survivors.append(n)
    else:
        removed[first] += 1
assert list(removed.values()) == [105, 35, 14, 8]
assert len(survivors) == 48
print('PASS eliminations 1..210: p2=105 p3=35 p5=14 p7=8 survivors=48')
a0, b = 97, 385
assert b - a0 == 288 and a0 % 144 == b % 144 and a0 % 1728 != b % 1728
assert all(a0 % p for p in range(2, int(a0**0.5) + 1)) and b == 5 * 7 * 11
print('PASS witness: 97 prime and 385 composite agree mod 12^2; split at 12^3')

# Exact ODE solution, numerically evaluated, no fitted or simulated models.
# x'=theta*x + K*(exp(a*t)-x), x(0)=1.
def corrected(t, a, theta, K):
    delta = a - theta + K
    if abs(delta) < 1e-13:
        return exp((theta-K)*t)*(1+K*t)
    return exp((theta-K)*t)+(K/delta)*(exp(a*t)-exp((theta-K)*t))

true_a, wrong_theta, T = 0.4, 0.15, 2.0
for K in (0.0, 0.5, 2.0, 8.0):
    # When theta=a, the correction vanishes exactly by the closed-form identity.
    assert all(isclose(corrected(t, true_a, true_a, K), exp(true_a*t), rel_tol=1e-12, abs_tol=1e-12)
               for t in (0.0, 0.25, 1.0, T))
    assisted_err = abs(exp(true_a*T)-corrected(T, true_a, wrong_theta, K))
    autonomous_err = abs(exp(true_a*T)-exp(wrong_theta*T))
    assert autonomous_err > 0
    assert (assisted_err <= autonomous_err + 1e-12)
    print(f'PASS observer toy K={K:.1f}: assisted final error={assisted_err:.9f}; '
          f'autonomous final error={autonomous_err:.9f}')
# For a misspecified theta, stronger feedback makes a wrong field appear better.
assert abs(exp(true_a*T)-corrected(T,true_a,wrong_theta,8.0)) < abs(exp(true_a*T)-corrected(T,true_a,wrong_theta,0.0))
print('ALL FINITE CHECKS PASS; analytic toy only; NO chaotic-model recovery or spectral claim tested')