Published July 8, 2026 | Version v2
Other Open

Fractal Geometric Analysis of the Twin Prime Counting Function via the Fractal Correction Engine

Description

# A Fractal Correction Engine Analysis of the Twin-Prime Counting Function $\pi_2(x)$, With Surrogate-Gated Null Testing of Every Claim

**Author:** Adam L McEvoy

**Keywords:** twin primes, fractal correction engine, Hardy–Littlewood, Brun's constant, surrogate null models, phase randomization, Riemann zeta zeros, explicit formula, reproducibility

---

## Abstract

I present a complete, self-auditing computational study of the twin-prime counting function $\pi_2(x)$ using the **Fractal Correction Engine (FCE)** — a domain-agnostic predict → compare → correct framework that extracts a curvature-driven "fractal path" from an observed signal and forecasts it between re-anchoring events. The system maps $\pi_2(x)$ and eleven derived sub-waveforms as independently evolving wave surfaces, models the double sieve (the joint survival of $p$ and $p+2$) as a phasor interference problem, and runs a closed adaptive control loop that classifies its own prediction residuals.

The distinguishing feature of this work is that **every inferential claim is tested against a matched statistical null before it is reported.** Descriptive geometry (fractal dimension, Hurst exponent, Brun-constant convergence) is separated from inferential claims (predictive superiority over Hardy–Littlewood, golden-ratio dominance, zeta-zero correlation, sieve interference). Under proper surrogate and baseline nulls, **none of the four headline inferential claims survive.** What does survive is grounded in real analytic number theory: a parameter-free reconstruction of the prime-counting Chebyshev function $\psi(x)$ from Riemann zeta zeros (a theorem), and a closed-loop residual classifier that independently rediscovers the leading-order Hardy–Littlewood bias as a secular drift. I report the full run, the exact numbers, and the methodology so that the negative results are as reproducible as the positive ones.

---

## 1. Introduction

The twin-prime conjecture asserts that there are infinitely many primes $p$ such that $p+2$ is also prime. It remains open. This paper does **not** attempt a proof. Instead, it asks a narrower, empirical question:

> When the twin-prime counting function is treated as a signal and analyzed with a general-purpose fractal/wave prediction engine, do any of the "interesting" structural signatures — golden-ratio self-similarity, zeta-zero resonance, predictive gain over the classical estimate, sieve interference asymmetry — reflect real properties of $\pi_2(x)$, or are they artifacts of the measurement procedure?

The motivation is methodological. Fractal- and wave-based analyses of number-theoretic sequences routinely produce striking-looking numbers — a correlation of $48.6$ here, a "$1.49\times$ improvement" there — that evaporate once compared against a proper null distribution. This system is built so that the null is computed automatically, in the same run, and printed next to every claim. The result is a pipeline willing to refute its own headline.

Throughout, I use the twin-prime constant
$$
C_2 \;=\; \prod_{p>2}\left(1 - \frac{1}{(p-1)^2}\right) \;=\; 0.6601618158\ldots,
$$
the golden ratio $\varphi = \tfrac{1+\sqrt 5}{2} = 1.6180339\ldots$, and Brun's constant $B_2 = \sum_{(p,\,p+2)}\left(\tfrac1p + \tfrac1{p+2}\right) = 1.9021605831\ldots$, where the sum runs over twin-prime pairs.

---

## 2. What the System Is

The system is a single Python pipeline that runs eight stages end to end:

1. **Sieve engine** — a sieve of Eratosthenes builds all primes up to $n_{\max}$ and identifies twin pairs, deserts, and clusters.
2. **Analyzer** — computes $\pi_2(x)$, gaps, Brun partial sums, densities, and the Hardy–Littlewood baseline.
3. **FCE core** — extracts the fractal/curvature path of $\pi_2(x)$ and runs forward/backward prediction.
4. **Sieve interference engine** — maps the $p$ and $p+2$ sieves as interacting waves.
5. **Spectral analyzer** — FFT / power-law / zeta-frequency decomposition.
6. **Statistical validation** — bootstrap confidence intervals, permutation tests, OEIS cross-check.
   - **6.5 Rigor / null-model layer** — computes a matched null for every inferential claim.
   - **6.6 Explicit-formula reconstruction** — the parameter-free $\psi(x)$ test and the twin test.
7. **Visualization** — twelve diagnostic figures plus a summary dashboard.
8. **Report generation** — a human-readable report that leads with the gated verdicts.

### 2.1 Hardware-safety substrate

The reference machine (AMD Ryzen 9 5900X) has a hardware fault: physical core 1 throws an L2-cache Machine Check Exception under sustained load (`mce: CPU 13 Bank 5`), halting the operating system below any software guard. Because a machine check cannot be caught in user space, the system does not try to survive it — it avoids the faulty silicon. A CPU-safety module reads the topology, pins the process to a safe subset of physical cores (logical CPUs $[0,4,5,8]$), forces single-threaded BLAS/OpenMP so numeric libraries cannot spawn workers onto the bad core, and an optional root helper offlines the faulty cores at the kernel level. Every result in this paper was produced on this pinned configuration with no crash.

---

## 3. What the Fractal Correction Engine Is

The FCE is a **domain-agnostic prediction-and-correction engine.** It does not model twin primes specifically; it models *any* signal that can be expressed through a common eight-slot interface, with the domain's physics plugged in as data.

### 3.1 The core idea

Given an observed path (a sequence of points), the FCE:

1. Extracts a **fractal path** using $\pi$ and the local **curvature** of the signal.
2. **Predicts** that path forward (and, where reversible, backward) *between perturbations*.
3. At each perturbation, **re-anchors**: it takes the newly measured state, recomputes curvature, and re-extracts the path.

The engine runs **closed-loop**, not open-loop: a forecast is provisional until an observation confirms or corrects it. Each cycle is **predict → compare → correct**, and the residual (forecast minus observed) is treated as a *signal*, not merely an error.

### 3.2 The universal plug interface

Every domain is described by the same eight slots:

| Slot | Meaning |
|------|---------|
| **STATE** | the coordinates that define one point on the path |
| **BASELINE** | the unperturbed "free" path law (the geodesic) |
| **CURVATURE** | what bends the path locally |
| **PERTURBATIONS** | events that re-anchor the path — each a `{trigger, update}` pair |
| **INVARIANTS** | conserved quantities that anchor prediction and sanity-check it |
| **SUPERPOSITION** | how two paths/waves combine (the interference rule) |
| **OBSERVABLES** | what is measured — the feedback channel |
| **COUPLING** | how perturbations from other parts enter |

Every domain reduces to one of three engines: a **trajectory** engine (a body in a state space), a **wave** engine (amplitude + phase field), or a **probabilistic/quantum** engine (complex amplitude evolved unitarily). The twin-prime problem is a **wave-engine** problem.

### 3.3 Curvature and the fractal path

For a signal $y(x)$, the local planar curvature is
$$
\kappa \;=\; \frac{\lvert y'' \rvert}{\left(1 + (y')^2\right)^{3/2}}.
$$
A critical subtlety: curvature is **not** scale-invariant unless both axes are nondimensionalized first. Computing $\kappa$ on raw counts (values up to $\sim 8000$ with an implicit $dx=1$) makes the reported curvature an artifact of the plot's aspect ratio. The system therefore z-scores both axes before differentiating:
$$
\tilde x = \frac{x - \bar x}{\sigma_x}, \qquad \tilde y = \frac{y - \bar y}{\sigma_y}, \qquad
\kappa = \frac{\lvert \tilde y'' \rvert}{\left(1 + (\tilde y')^2\right)^{3/2}},
$$
which guarantees $\kappa(cy) = \kappa(y)$ for any $c>0$. The $\pi$-normalized curvature $\kappa/\pi$ then feeds the correction operator, which adds a curvature-weighted, multi-scale $\zeta$-sum to the baseline:
$$
y_{\text{corr}} \;=\; y + \alpha\,\pi\,R\,\sum_{n\ge 1} \frac{d_n}{n^{3/2}}, \qquad R = \frac{\sigma_y}{1+\kappa},
$$
where $d_n$ are Gaussian multi-scale detail bands at widths $\sigma_n = \pi\,2^n$, and $R$ is the curvature-modulated osculating radius expressed in signal units.

### 3.4 Perturbation taxonomy

Every perturbation in every domain is one of seven classes, of which four matter here: **external forcing** (reversible), **coupling** (reversible), **boundary/interface** (reversible), and **dissipation/measurement** (irreversible → backward prediction becomes a reconstruction problem). Forward prediction always works; backward prediction is exact only back to the last irreversible event.

### 3.5 Interference algebra

When two waves combine, the resultant intensity follows phasor addition:
$$
I \;=\; A_1^2 + A_2^2 + 2 A_1 A_2 \cos(\Delta\varphi), \qquad \Delta\varphi = \varphi_2 - \varphi_1.
$$
The sign shorthand $++,\,--,\,+-,\,-+$ encodes the two extremes: same sign ($\Delta\varphi=0$) reinforces to $A_1+A_2$; opposite sign ($\Delta\varphi=\pi$) cancels to $\lvert A_1 - A_2\rvert$. For $N$ contributions the general phasor sum is $A_{\text{tot}} = \sum_j A_j e^{i\varphi_j}$ with intensity $I = \lvert A_{\text{tot}}\rvert^2$.

### 3.6 The adaptive closed loop

The residual's **shape** says where the model is incomplete:

- a **periodic** residual → a missing oscillatory mode;
- a **secular** (steadily growing) residual → a missing constant drift/force;
- a residual **correlated with another part** → a missing coupling.

The runtime re-anchors on a one-off perturbation but **escalates** (decomposes deeper) only when the residual returns with the *same* structure. Escalation is hysteresis-gated (escalate after $n$ over-tolerance segments, de-escalate after $m>n$ calm ones) so the analysis depth does not thrash.

---

## 4. How the FCE Is Applied to the Twin-Prime System

The FCE is applied **once per independently evolving part**, not once for the whole system. For this problem the decomposition yields **14 FCE instances**: the main $\pi_2(x)$ path, eleven derived sub-waveforms, and the two sieve-wave components. This is the wave-engine plugin, filled in:

- **STATE:** value, analytic-signal amplitude $A$, and phase $\varphi$ of the surface.
- **BASELINE:** for $\pi_2(x)$, the Hardy–Littlewood logarithmic-integral estimate; for other surfaces, a low-order fit.
- **CURVATURE:** the nondimensionalized $\kappa/\pi$ of §3.3.
- **PERTURBATIONS:** re-anchor when the residual exceeds tolerance (boundary class); couple the two sieve waves at twin loci (coupling class).
- **INVARIANTS:** OEIS A001359 membership, monotonicity of $\pi_2$, the $\mathrm{Li}_2$ asymptotic, and $B_2$-boundedness of the Brun sum.
- **SUPERPOSITION:** the phasor rule of §3.5, applied to the $p$-sieve and $(p+2)$-sieve waves.
- **OBSERVABLES:** forecast residual, MAE, $\varphi/\pi$ affinity, fractal dimension, Hurst exponent.

### 4.1 The classical baselines

The leading-order Hardy–Littlewood estimate is
$$
\pi_2(x) \;\sim\; 2 C_2 \, \frac{x}{(\ln x)^2},
$$
but this underestimates $\pi_2(x)$ by roughly $18\%$ at $x=10^6$. The correct, stronger baseline is the logarithmic-integral form
$$
\pi_2(x) \;\sim\; 2 C_2 \int_2^{x} \frac{dt}{(\ln t)^2} \;=\; 2 C_2 \,\mathrm{Li}_2(x),
$$
and **any predictive claim must be measured against $\mathrm{Li}_2$, not the leading-order form.** Using the weak baseline is the single most common way to manufacture an illusory "improvement."

### 4.2 The double sieve as interference

Twin primes survive when **both** $p$ and $p+2$ pass the sieve. Define indicator waves $s_p(n)=\mathbb 1[n\text{ prime}]$ and $s_q(n)=\mathbb 1[n+2\text{ prime}]$; then the twin indicator is $t(n)=s_p(n)\,s_q(n)$. The system takes the analytic (Hilbert) signal of each wave to obtain $(A_p,\varphi_p)$ and $(A_q,\varphi_q)$, forms the relative phase $\Delta\varphi=\varphi_q-\varphi_p$, and applies the phasor intensity of §3.5. Constructive loci are $\cos\Delta\varphi > +\tfrac12$; destructive loci are $\cos\Delta\varphi < -\tfrac12$.

### 4.3 The explicit-formula reconstruction

As an independent, parameter-free check of the phasor engine on *real* number theory, the system reconstructs the Chebyshev function via the Riemann–von Mangoldt explicit formula (a theorem):
$$
\psi(x) \;=\; x \;-\; \sum_{\rho} \frac{x^{\rho}}{\rho} \;-\; \ln(2\pi) \;-\; \tfrac12\ln\!\left(1 - x^{-2}\right),
$$
where $\rho = \tfrac12 + i\gamma$ ranges over the nontrivial zeta zeros. Because this has no free parameters, rising reconstruction accuracy as more zeros are added is a genuine validation that the engine composes real analytic number theory.

### 4.4 The statistical gates

Each inferential claim is compared against a **matched null**:

- **Phase-randomized surrogates** (Theiler 1992): randomize the Fourier phases while preserving the exact power spectrum. Formally, for a series with Fourier transform $\hat y(f) = \lvert\hat y(f)\rvert e^{i\theta(f)}$, a surrogate replaces $\theta(f)$ with i.i.d. uniform phases (keeping conjugate symmetry). This is the correct null for "is the phase structure — peak-spacing clustering at $\varphi$, or power landing on zeta frequencies — more than a linear Gaussian process with the same spectrum would give?"
- **Constant-grid ranking:** a constant is "dominant" only if it is the best fit among seven candidates $\{1, \varphi, \sqrt2, e, \pi, 2, 3\}$ — not merely better than $\pi$. The affinity score on peak-spacing ratios $r$ is $\exp\!\left(-\,\overline{\lvert r - c\rvert}\right)$.
- **Control-frequency ensembles:** zeta-frequency band power is compared against many random control frequencies of the same count.
- **Change-point test:** an "information horizon" is reported only if $\log(\text{MAE})$ vs $\log(\text{horizon})$ shows a genuine slope *break*, not merely an increase.

A claim survives only if it beats its null (typically $p<0.05$) **and**, for the constant test, ranks first.

---

## 5. Results and Findings (Latest Run)

**Run configuration:** $n_{\max}=100{,}000$, bootstrap samples $=1000$, permutations $=500$, fixed seed. Pinned to safe cores; exit code $0$; no machine-check event.

### 5.1 Descriptive facts (not in dispute)

| Quantity | Value |
|---|---|
| Twin pairs found, $\pi_2(10^5)$ | $1224$ (exact OEIS A001359 match) |
| Mean gap between twins | $81.75$ |
| Twin deserts / max desert / clusters | $101$ / $630$ / $433$ |
| Brun partial sum (vs $B_2 = 1.9021605831$) | $1.6727995848$ (deviation $-0.2294$, rate $0.157$) |
| Hardy–Littlewood $R^2$ / RMSE | $0.815$ / $141.95$ |
| Fractal dimension (Higuchi) | $1.7518$ |
| Hurst exponent | $0.9737$ |
| Curvature $\overline{\kappa/\pi}$ (nondim; core $=$ null layer) | $10.9353$ |

The curvature agreement is a consistency check: the FCE core and the independent null-model layer now compute identical nondimensionalized curvature ($10.9353 = 10.9353$), confirming the geometry is aspect-ratio-free.

### 5.2 The four headline claims — and their gated verdicts

**Claim 1 — "The FCE beats Hardy–Littlewood by $1.49\times$."**
Against the *leading-order* baseline (MAE $204.4$) the enhanced FCE ensemble (MAE $137.2$) does look better. But against the *correct* $\mathrm{Li}_2$ baseline (MAE $27.6$),
$$
\text{improvement vs }\mathrm{Li}_2 \;=\; \frac{27.6}{137.2} \;=\; 0.201 \;<\; 1,
$$
i.e. the FCE ensemble is about $5\times$ **worse** than the classical estimate. Every individual strategy also fails to beat $\mathrm{Li}_2$ (Bias Correction $89.99$, Multi-Scale $191.85$, Fractal Path $606.24$). The permutation test on the raw FCE prediction gives $p=1.0$. **Verdict: ARTIFACT** — the apparent gain is exactly the leading-order bias the FCE absorbs.

**Claim 2 — "Golden-ratio dominance."**
The reported $\varphi$-correlation ($0.668$) exceeds the $\pi$-correlation ($0.149$), but that is a two-horse race. Ranked against seven constants by peak-spacing affinity:
$$
\text{unity}(1.0)\!:\,0.787 \;>\; \sqrt2\!:\,0.715 \;>\; \varphi\!:\,0.668 \;>\; 2\!:\,0.467 \;>\; e\!:\,0.228 \;>\; 3\!:\,0.172 \;>\; \pi\!:\,0.149.
$$
$\varphi$ ranks **3rd of 7** (surrogate $p=0.040$, only $6$ ratios). The trivial attractor is ratio $\approx 1$ (evenly spaced peaks). **Verdict: ARTIFACT.**

**Claim 3 — "Zeta-zero correlation $=48.6$."**
The $48.6$ figure was a max-over-median power ratio, which exceeds $1$ for essentially any signal. The matched-statistic enrichment of band power at zeta frequencies versus random control frequencies is $0.683$ (i.e. *below* random), with control-frequency $p=0.921$ and phase-randomized surrogate $p=1.0$. **Verdict: ARTIFACT.**

**Claim 4 — "Sieve interference asymmetry."**
Under the density-ratio null the constructive/destructive asymmetry is $16$ against a spectrum-matched null of $20.6\pm15.0$ ($p=0.645$). Under the true phasor treatment (§4.2), constructive $=1567$, destructive $=1139$, but the mean $\cos\Delta\varphi = 0.083$ (phases essentially unrelated) and the surrogate $p=0.322$. The four-quadrant census is $++\!:205$, $--\!:2749$, $+-\!:1023$, $-+\!:1022$ — dominated by mutual compositeness, exactly as random. **Verdict: ARTIFACT.**

### 5.3 Exhaustive per-surface gating

Extending the golden-ratio gate to **all eleven** mapped sub-waveforms (H-L residual, gaps, Brun sums, density, acceleration, log-derivative, Brun deviation, cumulative gaps, $\pi_2/\text{H-L}$ ratio, spectral envelope, gaps $\bmod 6$):
$$
\textbf{0 of 11 surfaces survive the }\varphi\textbf{-gate.}
$$
Every surface either ranks $\varphi$ third or has too few autocorrelation peaks to test. There is no hidden surface on which the golden-ratio story holds. Coverage is total; significance is nil.

### 5.4 What survives

**Prime $\psi(x)$ reconstruction (real).** The parameter-free reconstruction from zeta zeros improves monotonically as zeros are added:

| # zeros | $1$ | $5$ | $10$ | $25$ | $50$ | $100$ |
|---|---|---|---|---|---|---|
| correlation | $0.44$ | $0.67$ | $0.78$ | $0.85$ | $0.90$ | $0.93$ |

reaching correlation $0.93$ at $100$ zeros. This is a theorem, and its success confirms the phasor engine composes genuine analytic number theory — **for primes.**

**The twin test is null.** Fitting the same zeta-frequency basis to the twin residual gives $R^2_{\text{zeta}}=0.065$ versus a matched random-control median of $R^2_{\text{control}}=0.067$ (enrichment $0.974$, surrogate $p=0.529$): *zeros no better than random frequencies of the same count.* The engine works; the twin signal simply has no special zeta structure.

**The closed loop rediscovers the missing drift.** Run on the H-L residual, the adaptive loop classified **all 15 segments as secular** (never periodic, never flat), triggering $15$ re-anchors and $5$ escalations to depth $5$. A secular residual is, by the classifier's definition, a *missing constant drift* — and that drift is precisely the leading-order Hardy–Littlewood bias that the $\mathrm{Li}_2$ term corrects. Two independent mechanisms — the baseline comparison of §5.2 and the §3.6 residual classifier — point at the same real analytic fact. This is the most satisfying result in the study: the framework diagnosing a genuine correction term rather than inventing a spurious one.

**A clean information horizon.** The stricter change-point test detects a real slope break in $\log(\text{MAE})$ vs $\log(\text{horizon})$ at horizon $\approx 200$: the slope rises from $0.78$ to $16.0$, explaining $99.9\%$ of the variance. The break is statistically clean; its interpretation is the mundane one that extrapolation error grows rapidly past a short horizon.

### 5.5 Sensitivity caveat

The descriptive fractal/Hurst statistics are computed on a cubic detrend. The residual standard deviation under a cubic detrend ($4.57$) differs from that under the $\mathrm{Li}_2$ detrend ($9.56$) by roughly a factor of two, so those descriptors are sensitive to an arbitrary modeling choice and should be read as such. The Higuchi point estimate ($1.75$) also lies outside its own bootstrap confidence interval $[1.99, 2.00]$, and the Hurst estimate ($0.97$) above its interval $[0.49, 0.64]$ — further reason to treat the descriptive geometry as method-dependent rather than intrinsic.

### 5.6 Summary of gated verdicts

| # | Headline claim | Gated statistic | Verdict |
|---|---|---|---|
| 1 | Beats H-L $1.49\times$ | $0.201\times$ vs $\mathrm{Li}_2$; perm. $p=1.0$ | **Artifact** |
| 2 | Golden-ratio dominance | $\varphi$ rank $3/7$, $p=0.040$ | **Artifact** |
| 3 | Per-surface $\varphi$ structure | $0/11$ surfaces survive | **Artifact** |
| 4 | Zeta correlation $48.6$ | enrichment $0.68\times$, $p=1.0$ | **Artifact** |
| 5 | Twin ↔ zeta structure | $R^2_\zeta 0.065 \approx R^2_{\text{ctrl}} 0.067$, $p=0.529$ | **Artifact** |
| 6 | Sieve interference (phasor) | $\overline{\cos\Delta\varphi}=0.083$, $p=0.322$ | **Artifact** |
| — | Prime $\psi(x)$ reconstruction | corr $0.44\to0.93$ over $1\to100$ zeros | **Survives (theorem)** |
| — | Secular-drift diagnosis | $15/15$ segments secular | **Survives (real)** |
| — | Information-horizon break | $h\approx200$, var. expl. $0.999$ | **Survives (clean)** |

---

## 6. Discussion

The pattern here is instructive and, I suspect, general. Every inferential claim that looked exciting — beating the classical estimate, golden-ratio self-similarity, zeta-zero resonance, interference asymmetry — was an artifact of a weak baseline, an unfair comparison, an unnormalized statistic, or an absent null. Each dissolved the moment a matched null was placed beside it. Making the coverage *more* thorough — gating all eleven surfaces, upgrading the interference model to true phasor composition — did not rescue any claim; it only made the refutation more complete.

Conversely, the two things that survived are exactly the two things anchored in established mathematics: a theorem (the explicit formula) and a known analytic correction (the $\mathrm{Li}_2$ term, rediscovered as a secular residual). The FCE is a capable signal-composition engine — it reconstructs $\psi(x)$ from zeta zeros with rising fidelity — but capability is not the same as discovery. Applied to twin primes, it reveals no geometry beyond what Hardy and Littlewood already gave.

The methodological takeaway stands on its own: a fractal/wave analysis pipeline is only as trustworthy as the nulls it computes for itself. Building the null-model layer into the same run, and forcing the report to lead with the gated verdict, converts an engine that could have produced four false discoveries into one that honestly reports four negatives and two positives.

---

## 7. Methods and Reproducibility

- **Language / stack:** Python 3.12, NumPy, SciPy (optional), Matplotlib (Agg backend).
- **Sieve:** boolean sieve of Eratosthenes to $n_{\max}=10^5$; twins identified by $s(n)\wedge s(n+2)$.
- **Curvature:** nondimensionalized planar curvature (§3.3); verified scale-invariant and identical to the independent null-layer implementation.
- **Nulls:** phase-randomized (Theiler) surrogates, $300$–$500$ realizations per statistic; constant-grid ranking over $\{1,\varphi,\sqrt2,e,\pi,2,3\}$; random control-frequency ensembles; log–log change-point test.
- **Explicit formula:** first $100$ nontrivial zeta zeros; prime validation parameter-free; twin test parameter-matched against random controls.
- **Determinism:** fixed seeds; the reported numbers are exact outputs of the run described in §5.
- **Hardware safety:** process pinned to non-faulty cores with single-threaded BLAS/OpenMP; no machine-check event during the run.
- **Test suite:** 39 unit/regression tests covering the sieve, curvature scale-invariance, per-surface gates, phasor limits (in-phase $\to$ constructive, antiphase $\to$ destructive), the residual classifier, and the closed loop.
- **Artifacts:** `metrics.json` (all statistics, including the `rigor` block), `analysis_report.txt` (leads with gated verdicts), and twelve figures including a summary dashboard whose verdict panel marks each claim PASS/FAIL.

**Data availability.** All code, the metrics JSON, the report, and the figures are included in the deposit. The negative results are as reproducible as the positive ones from a single command.

---

## 8. Conclusion

I built a fractal/wave prediction engine, applied it exhaustively to the twin-prime counting function, mapped every surface and every interaction, and gated every claim against a matched statistical null. The verdict is clean and honest: **the twin-prime "fractal / golden-ratio / zeta / predictive-superiority" signatures are measurement artifacts**, while the components grounded in real number theory — the parameter-free $\psi(x)$ reconstruction and the secular-drift diagnosis of the Hardy–Littlewood bias — hold up. The twin-prime conjecture is untouched by this work; what the work demonstrates is a disciplined way, built directly into the analysis pipeline, to tell an artifact from a discovery.

---

## References

1. G. H. Hardy and J. E. Littlewood, "Some problems of 'Partitio numerorum' III: On the expression of a number as a sum of primes," *Acta Mathematica* **44** (1923), 1–70.
2. V. Brun, "La série $\tfrac15+\tfrac17+\tfrac1{11}+\tfrac1{13}+\cdots$ où les dénominateurs sont nombres premiers jumeaux est convergente ou finie," *Bulletin des Sciences Mathématiques* **43** (1919), 100–104, 124–128.
3. B. Riemann, "Ueber die Anzahl der Primzahlen unter einer gegebenen Grösse," *Monatsberichte der Berliner Akademie* (1859).
4. J. Theiler, S. Eubank, A. Longtin, B. Galdrikian, J. D. Farmer, "Testing for nonlinearity in time series: the method of surrogate data," *Physica D* **58** (1992), 77–94.
5. T. Higuchi, "Approach to an irregular time series on the basis of the fractal theory," *Physica D* **31** (1988), 277–283.
6. Y. Zhang, "Bounded gaps between primes," *Annals of Mathematics* **179** (2014), 1121–1174.
7. OEIS Foundation, sequence A001359 (lesser of twin primes), *The On-Line Encyclopedia of Integer Sequences*.

Files

01_counting_function.png

Files (4.6 MB)

Name Size
md5:435df8866848da2936d894a85ac25d58
135.2 kB Preview Download
md5:5d852a81b8498aa7592e600ae7d89954
454.3 kB Preview Download
md5:ab954251716597e76c4cdc52d3775195
102.0 kB Preview Download
md5:51278981398ec1d2f175aa2d6c89c0e7
278.7 kB Preview Download
md5:330d922988b1240d6c3473ad9bf6df61
88.2 kB Preview Download
md5:b1b89224d57840b02e2a365ae5f3712f
171.5 kB Preview Download
md5:6d0b01f600b9cb5d87fd4d498c4533d3
145.0 kB Preview Download
md5:82e9ff56d681114481d8a6fa7261cbe6
258.6 kB Preview Download
md5:1eab82fa5913ef3ca4a956bf897695ed
98.5 kB Preview Download
md5:8add7c39773399d65632767a7c83fb64
117.1 kB Preview Download
md5:0611c81c8c21c6ec70b0fa0629d19b10
171.8 kB Preview Download
md5:9d5f0cde42bb92577558cbc98a058061
207.2 kB Preview Download
md5:8cfd76e594e633e4ebbe864e83203248
8.1 kB Preview Download
md5:a5c050cb89e690eb541dd3b71b5c0443
22.2 kB Preview Download
md5:135fbe685d41ca77c7eead203c4644cd
8.9 kB Download
md5:3b925010ae64232b25cae79ad9a7c7e0
916 Bytes Download
md5:bf272701aead323af1b0e3b5223a461d
16.4 kB Download
md5:7b842c93cdb79664d98fb909ed44b69a
4.0 kB Download
md5:d9f7f3a13a7c2214999975040496ab32
34.2 kB Download
md5:d2ccea2d6daf22090bff16238e07ce9f
1.7 kB Download
md5:b058414e7efbd05ede454c20343009c2
3.0 kB Download
md5:00d82060b9e743eeb7e696332c0f8288
18.1 kB Preview Download
md5:f2ab5790dc2f58a476a82e2181bbf4a1
4.4 kB Preview Download
md5:27ace5b07f03ee9d05c4ac524d46d062
238 Bytes Preview Download
md5:f1ce2bb86cf73c15c2d9eb7b48735d72
3.3 kB Download
md5:03b63eccead2e5525a5d66f7e30a5c9a
1.8 kB Download
md5:b9c08fe9a3deb04a7d36122257d57e11
3.8 kB Download
md5:62f303ae8178262ef1e5a94436ca2929
4.0 kB Download
md5:71a45398ef5532e86fa6f4fcd1fd55f7
5.5 kB Download
md5:10c505e2d245e8c2b0e7c83e67733f73
2.6 kB Download
md5:d6f9acdb480c3ca54bc3e191d3d3e5f2
10.1 kB Preview Download
md5:48dc100d06b21dad11824040b230bb13
145.0 kB Download
md5:043a8b21975be65937254899f8133c27
2.1 MB Preview Download
md5:6df76e2ee8eea2271f997a6930d81f2d
26.1 kB Preview Download