# Quaternion-order height10,000 and the inheritance pattern

[project] Completed November 10, 2026, 01:48 PM PST. Extends the accepted100/1000 windows without changing the source-pinned mathematical evaluator. The original1000 archive is retained; the full10000 certificate is distributed as bounded multipart archives.

## Certified counts [project]

Positive heights mean0<Im(s)<=10000, anywhere in the complex plane. The centre is excluded from positive totals. Every nonreal zero in this window is simple.

|D|quarter|middle|three-quarter|positive total|centre order|symmetric total including centre|
|---:|---:|---:|---:|---:|---:|---:|
|2|22491|2206|22491|47188|0|94376|
|3|22491|3496|22491|48478|0|96956|
|30|22491|10824|22491|55806|2|111614|

Local counts are2206 for p2,3496 for p3,5122 for p5. D30's middle count is2206+3496+5122=10824. Sum the prime-by-prime floors, not the floor of their sum. No unexpected zeros, repeated nonreal zeros or complementary-factor cancellations appeared.

## Literal Turing completeness, not a table comparison

[known] Trudgian, *Improvements to Turing's Method*, Math.Comp.80(2011), Theorem2.2, gives

    |integral_b^c S(t)dt| <= 2.067+0.059 log(c), c>b>168*pi.

Primary source: https://arxiv.org/abs/0903.1885 ; DOI10.1090/S0025-5718-2011-02470-1. The retrieved original PDF/text and source hashes are audited in order-height/sources/README.md. The full integer-upper-bound derivation is retained in order-height.md, sections2–3.

[project] Here b=20000,c=20064. With n located sign brackets below b, and h_j the upper endpoints of disjoint brackets in(b,c),

    N(b) <= [B(c)+integral_b^c(theta(t)/pi+1)dt-sum_j(c-h_j)]/(c-b).

Each tail root contributes at least c-h_j to the integrated staircase count; no completeness assumption about the tail is used. The accepted83 tail brackets and1024 full interval log-gamma integration boxes give

    U = 22491.04704202994... +/-0.0616 <22491.109<22492.

The lower count is22491. Hence N_zeta(20000)=22491, and all22491 sign-bracket roots are simple and on the ordinary critical line. The Turing proof leaves an integer margin greater than0.89. The continuous analytic log-gamma branch in Re(z)>0 is used, not a principal argument of Gamma. Interval Riemann ranges cover every integration cell; there is no unbounded numerical quadrature remainder.

[project] Arb indexing is used only to choose seed coordinates. Fresh160-bit Hardy-Z endpoint signs plus our explicit Turing upper count are the proof. A224-bit streaming replay recomputed all signs, averages and complementary factors. No zero table or built-in N(t) value is a proof input. Stored22592 brackets include padding above20064; only22491 are claimed complete below20000. The old100/1000 certificates are retained and replayed too.

## Exact transport and why there are no extra zeros

[known] Voight, *Quaternion Algebras*, Theorem26.4.9 / equation26.4.10, https://jvoight.github.io/quat-book.pdf :

    Z_D(s)=zeta(2s) zeta(2s-1) product_(p|D)(1-p^(1-2s)).

An ordinary nontrivial zero rho contributes rho/2 and(rho+1)/2. Local nonreal roots are exactly1/2+i*pi*k/log(p),k nonzero. Nonreal roots cannot cancel against poles, which here are real. The other zeta factor at inherited zeros is nonzero by the functional equation / Euler-product half-plane; on local roots the zeta factors lie on Re0 and1 at nonzero height and are nonzero. Distinct-prime local roots cannot coincide because p^a=q^b is impossible for distinct primes. All other zeta zeros are real trivial zeros. These facts classify the entire nonreal divisor.

[project] Completeness is exact transport PLUS the fresh ordinary-zeta completeness proof—not a direct giant order-zeta contour. The replay additionally verified 135552 inherited complementary products and 16526 D-specific local complementary products. All exclude0. Central orders remain0,0,2; the D30 centre is a local-factor/pole cancellation, not a Q8 Artin forced centre. The coordinate is ideal-index s; reduced-norm coordinate w=2s doubles these heights to20000.

## Inheritance, not interaction

[known] In this finite verified window, an ordinary zero1/2+i*gamma appears on BOTH outer lines at exactly the same height gamma/2. The real coordinates become1/4 and3/4. This is a deterministic copying-and-rescaling rule, not an observed statistical resemblance.

[known] For each ramified prime p, middle-line heights pi*k/log(p) form an arithmetic progression with spacing pi/log(p): approximately4.532360142 for p2,2.859600867 for p3,1.951981266 for p5. The D30 union of three progressions is NOT itself equally spaced.

[project] To ask “how close in height?”, define the finite minimum

    delta_D(T)=min |gamma_j/2 - pi*k/log(p)|,
    0<gamma_j/2<=T, 1<=k<=floor(T log(p)/pi), p|D.

The minimum is the same for either outer line. Centre heights and local points above T are excluded. Strict certified enclosures at T10000 are:

|D|lower bound|upper bound|witness p|ordinary zero index|local k|
|---:|---:|---:|---:|---:|---:|
|2|0.000224748746|0.000224748750|2|9432|1036|
|3|0.000055637662|0.000055637666|3|21796|3402|
|30|0.000009241204|0.000009241208|5|18961|4412|

The D30 closest pair is ordinary zero#18961 divided in height by2, near8612.141335605921, versus local p5,k4412, near8612.141344847128. Their difference is about0.000009241206819. These displayed midpoint approximations are illustrations; the rational interval bounds above are the certificate.

### Why the gap check is exhaustive [project]

For period h=pi/log(p), enclose q=t/h for the whole inherited-root bracket. A floating estimate n is used ONLY as a hint; interval arithmetic must verify n-1<q<n+2. Check all integers n-1,n,n+1,n+2 in the allowed finite range. Every omitted integer is more than one full period away. The global lower bound is the minimum of all candidate-distance lower endpoints and h.lower; a candidate-distance upper endpoint supplies an attained upper bound. Every stored decimal bound is first made an exact rational and then rechecked by a STRICT interval comparison.

The runner-up lower endpoint exceeds the selected witness upper endpoint for each prime. Cross-prime lower bounds isolate the D30 witness as well. Thus the closest ordinate pair is certified unique for each D (but copied to two outer lines). Five unit tests cover integer/half-integer candidates, finite endpoints, agreement with brute force, exact vertical coincidence, and ties that must NOT be called unique. A separate288-bit full gap replay gives the same rational bounds and witness indices. Exact coincidence would return a zero lower bound, not manufacture a positive one.

[known] The Euclidean complex-plane separation from either outer line to the middle is at least1/4 because their real coordinates differ by1/4. Tiny vertical gaps therefore do not merge zeros. The exact product superposes families; these data establish no attraction, repulsion or dynamical interaction. [project] Positive vertical lower bounds here are finite-window facts, NOT a uniform lower bound for all heights, and no all-height noncoincidence theorem is claimed.

## Reproducibility, costs and figures [project]

Generator: 374.408s; peak52944KiB (51.7MiB).
Full224-bit replay: 114.698s; peak47756KiB (46.6MiB).
Combined accepted generation+required replay: 489.106s (8.15minutes). Gap pass/replay: 1.911s / 2.038s. These exclude development, figures, archives, publication and any extra archive replay. The earlier15–30minute sample-based forecast was conservative; it was not a guarantee.

The run resumed all70 accepted1000 chunks and added656 ordinary chunks plus155 local chunks. In total706 ordinary chunks and175 local chunks are immutable, atomically checkpointed, source-fingerprinted and hash-checked. Both main passes ran under nohup; progress and PIDs are in order-height/. Mathematical sources were unchanged. Python-flint0.9.0 / FLINT3.6.0; PARI is not needed.

Commands from this directory (run each only after the preceding dependency finishes):

```sh
nohup /workspace/tools/pari-venv/bin/python -u order_height.py --targets 100,1000,10000 > order-height/run-10000.log 2>&1 &
nohup /workspace/tools/pari-venv/bin/python -u order_height_verify.py > order-height/verify-10000.log 2>&1 &
/workspace/tools/pari-venv/bin/python order_inheritance_gap.py
/workspace/tools/pari-venv/bin/python order_inheritance_gap.py --precision 288 --output inheritance-gap-10000-replay.json
```

Publication uses bounded multipart archives listed in order-height/manifest-10000.json. Extract ALL parts into one empty folder, then run python order_height_verify.py --run-dir order-height and the gap replay. Individual chunks are also published. The original1.9MB order-height-certificate.tar.gz is preserved as the1000 snapshot; it is not the10000 archive. Legacy manifest-files.json/resume tests/benchmark records describe that earlier run; current full hashes are manifest-10000.json and manifest-files-10000.json.

Figures (SVG plus PNG, each with separate mobile variant): figures/order-height10000-growth; figures/order-height10000-density; figures/order-inheritance-gap. Density cells are exact100-height bins, not individual dots. Growth compares certified samples with familiar main terms, not an independently predicted zero count.

[known] Platt–Trudgian, *The Riemann hypothesis is true up to3*10^12*, BLMS53(2021),792–797, https://arxiv.org/abs/2004.09765 , DOI10.1112/blms.12460, already gives a vastly larger finite ordinary verification. It is context, not our proof input. This is a reproducible transport of known finite information, NOT RH progress or a new joint field.
