# Quaternion-order heights100 and1000: literal Turing verification

## Extension changelog

[project] November 10, 2026, 01:48 PM PST: height10000 is now complete, with full replay and certified vertical-gap bounds; see [order-height10000.md](order-height10000.md). The original100/1000 report and its then-future estimates below are historical, not the current limit. Its frozen archive remains unchanged.

[project] November10,2026, PST. The requested two windows are complete.
Height10000 is **not** completed here: only a32-root high-height timing sample
was run to estimate the explicitly requested follow-up.

## Results: count positive heights, not the centre

Every count below includes all nonreal zeros anywhere in the complex plane
with0<Im(s)≤T, with multiplicity. Every listed nonreal zero is simple.

| T | D | quarter-line | middle line | three-quarter-line | positive total | centre order |
|---:|---:|---:|---:|---:|---:|---:|
|100|2|79|22|79|**180**|0|
|100|3|79|34|79|**192**|0|
|100|30|79|107|79|**265**|2|
|1000|2|1517|220|1517|**3254**|0|
|1000|3|1517|349|1517|**3383**|0|
|1000|30|1517|1081|1517|**4115**|2|

[project] Local-prime counts at100: p2=22,p3=34,p5=51; at1000:
220,349,512. The D30 middle count is their sum. Including conjugates and
the central point gives full symmetric multiplicity totals360,384,532 at100,
and6508,6766,8232 at1000. The centres atD2,D3 remain nonzero; D30 stays double.
No unexpected zeros, repeated nonreal zeros, or factor collisions appeared.

## 1. What is being transported [known]

Keep the ideal-index coordinate, as in Voight, *Quaternion Algebras*,
Theorem26.4.9 / equation26.4.10:

    Z_D(s)=ζ(2s)ζ(2s−1)∏_{p|D}(1−p^(1−2s)).

Each ordinary nontrivial zeroρ suppliesρ/2 and(ρ+1)/2, with the same
multiplicity. The other zeta factor is respectivelyζ(ρ−1) orζ(ρ+1), nonzero
by the functional equation and Euler-product half-plane. Local factors only
vanish on Re(s)=1/2. Their nonreal zeros are exactly

    s=1/2+iπk/log(p),        k=±1,±2,...,

with derivative2log(p)≠0. Distinct-prime families cannot coincide because
q^k=p^l is impossible for distinct primes and positive integers. At these
points the zeta factors are on Re=0 and1 and nonzero at nonzero height.
At the centre, r finite ramified primes give order r−1, by the Laurent
expansion recorded in `order-off-axis.md`. These analytic arguments classify
all nonreal factor zeros and rule out missing poles/cancellations. They are
not conditional on RH. Only the universal quarter-line location of all
inherited zeros would be equivalent to ordinary RH.

[project] Completeness here is **exact transport plus certified ordinary-zeta
completeness**, not a new direct order-zeta argument-principle computation.
We additionally enclose complementary products at every certified root.
The old height30 direct order contours remain a separate archived check.

## 2. A genuine Turing average, not a library count relabelled

[known] For N(t) counting positive nontrivial ordinary zeros with multiplicity,

    N(t)=θ(t)/π+1+S(t),
    θ(t)=Im logΓ(1/4+it/2)−(t/2)logπ.

Use logΓ's continuous analytic branch in Re(z)>0, not a principal argument
ofΓ. Primary bound: Trudgian, *Improvements to Turing's Method*,
*Math.Comp.*80(2011),2259–2279, Theorem2.2:

    |∫_b^c S(t)dt| ≤ B(c)=2.067+0.059log(c),    c>b>168π.

Original PDF and text were retrieved and the exact theorem inspected:
https://arxiv.org/abs/0903.1885 ; DOI10.1090/S0025-5718-2011-02470-1.
Local read-only source copies for this run are in`order-height/sources/`,
not the separately protected`literature/` directory.

### Derivation of the actual integer upper bound [project]

Suppose n distinct sign brackets lie below b. Then N(b)≥n. For each additional
sign bracket in(b,c), let h_j be its **upper** endpoint. It contains a
critical-line zeroγ_j≤h_j, so its contribution to∫N(t)dt is at least c−h_j.
No assumption that the tail list is complete, or that those tail roots are
simple, is needed. Therefore

    ∫_b^c N(t)dt ≥ (c−b)N(b)+Σ_j(c−h_j).

Combine the exact N=θ/π+1+S identity with the upper half of the Turing bound:

    N(b) ≤ U := [B(c)+∫_b^c(θ(t)/π+1)dt−Σ_j(c−h_j)]/(c−b).

If interval arithmetic proves U<n+1, the integer N(b) equals n. Every one of
the n sign brackets is then simple, and no off-line or unobserved zero below
b can exist. A hypothetical missed zero anywhere below b persists through
the whole tail and would add at least c−b to the integral, violating the
bound. This is the literal averaging mechanism.

### Accepted averages and the low-height domain trap [project]

We take c=b+64 and use1024 exact-rational subintervals. On each entire
subinterval an Arb log-gamma ball encloses θ, and multiplying its range by
the width encloses the integral contribution. Thus this is a rigorous interval
Riemann enclosure, with no unbounded quadrature or asymptotic remainder.

- For the T100 order window, ordinary cutoff200 is **below168π**. We do not
  apply Trudgian's theorem there. We certify N(600)=341 using[600,664], then
  select the79 certified brackets below200.
- For T1000, use[2000,2064], yielding N(2000)=1517.
- The224-bit replay gives U600=341.0408225655...±0.0443 and
  U2000=1517.0350734268...±0.0501. In particular U600<341.086<342 and
  U2000<1517.086<1518. Both integer gaps exceed0.91.

[project] Arb zero routines provide **seeds only**. Fresh Hardy-Z endpoint
balls have opposite signs. Our proof never calls`arb.zeta_nzeros`.
This distinction matters: the inspected FLINT3.6.0 source uses known
Gram/Rosser shortcuts below ordinal13999526; blindly calling its default
zero-count interface would not demonstrate a newly executed literal Turing
bound at our heights. The primary implementation source was inspected and
archived in`sources/flint-3.6-isolate.c`. A built-in count comparison79/1517
was performed only **after** our own proof passed and is labelled a cross-check.
No Odlyzko/LMFDB zero table is a proof input.

## 3. Sign brackets, orders, and complementary factors [project]

Generator precision160 bits; streaming replay224 bits. The Hardy function is

    Z(t)=exp(iθ(t))ζ(1/2+it),

exactly real by the functional equation. Its real endpoint intervals have
strictly opposite signs; the imaginary enclosures contain0 as a sanity check.
All brackets are disjoint, positive, width2×10^−16. The Turing upper count
exhausts the pre-base brackets, proving simplicity and location in the finite
window. Inherited order-zeta ordinate brackets have width10^−16. The earlier
narrower height30 brackets remain available; no claim is made that these
new brackets improve their precision.

The stored1600 brackets include padding for both Turing tails; only1517 are
claimed complete below2000. There are9600 inherited complementary-product
checks (six per stored bracket),1081 local heights, and1650 D-specific local
complement checks. All exclude0. Tail-padding roots are not silently claimed
to form a complete higher window.

The middle-line count is proved by strict interval inequalities

    m_p < T log(p)/π < m_p+1,

and the exact local formulas, not a sign scan. The next local root is above
the boundary. The exact divisor classification then excludes all other
nonreal roots, including those outside the narrow rectangle used at height30.

## 4. Chunking, durability and replay [project]

-50 immutable ordinary-zero chunks,32 records each;
-20 local-root chunks, at most64 records each;
-atomic temporary-file/rename writes and source fingerprints;
-a state manifest with hashes, per-run progress and peak memory;
-a single-writer lock; a second writer fails rather than races;
-recovery of a completely written chunk if the subsequent state commit was
 interrupted; partial JSON files are never treated as checkpoints;
-no repeated loading of all complex-value records while producing each chunk:
 only a small ordinate/index list is retained for counting.

The generator and full verifier were run under`nohup`, with PID files and
unbuffered progress logs. Resume tests confirmed all70 good-chunk mtimes were
unchanged, a damaged hash was rejected, and an interrupted state commit was
recovered in an isolated copy and then fully mathematically replayed.
A source change deliberately invalidates reuse: use a new run directory.
The mathematical replay, not the hashes alone, establishes the certificate.

Accepted first-run generation:8.27 seconds, peak38556KiB (37.7MiB).
Final full replay:3.13 seconds, peak34904KiB (34.1MiB), approximately12 seconds
of accepted numerical work combined. These exclude research, pipeline
engineering, plots and publication. Pilot/test runs and timings are retained,
not counted as new mathematical results.

Commands from this directory:

```bash
nohup /workspace/tools/pari-venv/bin/python -u order_height.py \
  --targets 100,1000 > order-height/run.log 2>&1 &
nohup /workspace/tools/pari-venv/bin/python -u order_height_verify.py \
  > order-height/verify.log 2>&1 &
```

Run the verifier after the generator finishes. The orchestrated accepted run
waited for that dependency; running both simultaneously is not the procedure.
Use the same command to resume. A complete resume creates no new data chunks.
The numerical dependency is python-flint0.9.0 / FLINT3.6.0; PARI is not needed
for this transported order-zeta run.

## 5. Figures and the honest interpretation

[known] Per inherited line the main term is

    M(T)= (T/π)log(T/π)−T/π+7/8,

because the ordinary height is2T. The middle count is
Σ floor(Tlog(p)/π), with main term Tlog(D)/π. The positive total main term
is2M(T)+Tlog(D)/π. These are familiar asymptotics, not independent predictions
or exact counts; local floor errors and the ordinary S(t) remain.

[project]`order-height/plot-data.json` derives101 cumulative integer samples
and exact ten-height bins from the accepted complete window, with interval
local cutoffs. The growth figure compares D30's inherited and local counts
with their main terms. The density strips show binned **counts**, not exact
root locations; at375px, thousands of dots would falsely suggest missing
or merged roots. Separate mobile SVG/PNG variants show D30 with readable labels.

Figures:
-`figures/order-height-growth.svg` and`-mobile.svg` (plus PNGs);
-`figures/order-height-density.svg` and`-mobile.svg` (plus PNGs).

No unexpected phenomenon appeared. The local share for D30 falls from31/57
at30, to107/265 at100, to1081/4115 at1000. The inherited families grow like
Tlog(T), whereas the local family grows linearly. That contrast is already
forced by the known factorization.

[known] Platt–Trudgian already rigorously verified ordinary RH to approximately
3×10^12, with simple zeros: *The Riemann hypothesis is true up to3·10^12*,
*Bull.London Math.Soc.*53(2021),792–797,
https://arxiv.org/abs/2004.09765 , DOI10.1112/blms.12460.
Our much smaller ordinary window through2000 is a transported finite
verification and a reproducible tool, **not new RH information**. That larger
result is contextual comparison, not used to prove any of our counts.

## 6. Height10000: explicit follow-up, not done here [project estimate]

A timing sample of32 seeds near ordinary height19600, their signs and
complements, plus24 high local-root checks, was measured. No complete
height10000 window was attempted. The record is`benchmark-10000.json`.
At the known main-term forecast of about22500 ordinary positive zeros through
20000, the measured model estimates893 seconds for generation plus replay;
allow roughly**15–30 minutes** for the follow-up, **50–100MiB RAM**, with
roughly90MiB of raw checkpoint data. These are estimates, not guarantees.

The method need not change: resume the same32-root chunks, extend local chunks,
apply the same literal Turing integral on[20000,20064], replay, and then transport.
Our current1600 stored brackets are reused; their source fingerprints prevent
silently mixing changed evaluators. Arb already selects appropriate high-height
special-function algorithms; a giant order-zeta contour or custom FFT code is
unnecessary at this scale. Larger batches are optional optimisation, not a
mathematical change.

Follow-up command, only when that run is requested:

```bash
nohup /workspace/tools/pari-venv/bin/python -u order_height.py \
  --targets 100,1000,10000 > order-height/run-10000.log 2>&1 &
```

The10k publication should use the individual small JSON chunks or bounded
archive parts below the site's25MiB per-asset limit, rather than growing one
monolithic certificate upload. Do not report10k complete until its own Turing
upper bound and full replay pass. The4:30PM PST sprint sees the updated queue;
this run does not change its schedule or saved instructions.
