# Think harder: Heilbronn bookkeeping, its certificates, and its ceiling

10 November 2026, PST. Companion and critical revision to `think-hard-transfer.md`.
No deployment; no claim of RH progress or of a new subject.

**[project; subsequent independent check, 10 November 2026 PST]** The first
next-step proposal in §8 is now completed for all four proper subfields, the
four local targets and N(30): see `think-hard-independent.md`. That later
pipeline uses own prime-ideal data and a Gaussian functional equation, not the
factored evaluations discussed below. It recovers ordinary orders 0,1,1,1 and
N_ζ(30)=3. The degree-eight direct evaluator and global analytic gap remain open.

**Labels.** [known] = established mathematics, with sources in §9. [project] = a
proof reconstruction, certified finite computation, or interpretation here, not a
new number-theory theorem. [speculative] = an unproved additional input/proposal.

## 0. Verdict before the machinery

**[project]** Heilbronn characters really do link orders of zeros in a larger
field to ordinary zeta. They are especially useful when an Artin factor might
have a pole: a small independently known order of the whole field zeta can rule
that out. But **our Q8 factors are already known to be holomorphic**. In this
example the basic norm inequality adds no restriction beyond holomorphy and
factorization. Calling it our most promising *RH strategy* would be an overreach.
It is our most auditable *multiplicity-extraction mechanism*.

**[project; certified results, §4]** At the actual first three ordinary zeros,
near heights 14.13472514, 21.02203964, and 25.01085758, the specified degree-eight
field zeta and all its nontrivial proper subfield zetas have order **one**. Every
nontrivial irreducible Artin factor is nonzero there. At 1/2, ordinary zeta and all
proper subfield zetas have order **zero**, the two-dimensional Artin factor has
order **one**, and the field zeta has order **two**. These are interval results,
not guesses from rounded evaluations. They refine, and were already implied by,
the existing factored height-30 certificate; they do not extend its global scope.

**[project; essential limitation]** These new field contours are evaluated
**through the factorization**. The resulting counts are valid, but applying
Stark afterwards is not an independent discovery of simplicity downstairs. An
independent degree-eight/subfield evaluator would change the direction of the
verification, not turn three examples into RH.

## 1. General Galois case: every identity and hypothesis

### 1.1 Signed orders and induction

**[known; H73, FM, MR §3–4]** Let K/k be finite Galois, G=Gal(K/k), and fix a
complex point s0 other than 1. Brauer's theorem gives meromorphic continuation of
Artin L-functions. For each irreducible complex character χ of G set

    nχ = ord_s0 L(s,χ,K/k) in Z,
    Θ_G = Σχ nχ χ.

A positive order means a zero; a negative order means a pole. Thus Θ is an
integral **virtual** character, not automatically the character of an actual
representation. We use uncompleted L-functions here. In our boxes 0<Re(s)<1,
gamma factors have no zeros or poles, so completion does not change orders.

**[known]** For any subgroup H≤G, E=K^H, and irreducible character ψ of H,
Artin formalism and Frobenius reciprocity give, step by step,

    L(s,Ind_H^G ψ,K/k) = L(s,ψ,K/E),
    ord L(s,ψ,K/E)
       = Σχ <Ind_H^G ψ,χ>_G nχ
       = Σχ <ψ,Res_H^G χ>_H nχ
       = <Res_H^G Θ_G,ψ>_H.

Consequently

    Res_H^G Θ_G = Θ_H,                                      (1)
    ord_s0 ζ_E = <Res_H^G Θ_G,1_H>
                 = Σχ nχ dim(Vχ^H).                        (2)

Taking H={1} and H=G respectively gives

    M := ord_s0 ζ_K = Θ_G(1) = Σχ χ(1)nχ,
    m := ord_s0 ζ_k = n_1.                                  (3)

These are identities of signed orders, not assumptions of Artin holomorphy.
They apply to non-Galois intermediate E as well. In particular, no inference
that a factor is nonzero is justified merely by a floating-point product value.

### 1.2 The norm bound, reconstructed without Artin's conjecture

**[known; FM; explicitly quoted as MR Theorem 4.4. Proof reconstructed here.]**
For any g∈G, let C=<g>. This subgroup is cyclic. Every irreducible character of C
is one-dimensional, and its Artin L-function over K^C is a Hecke L-function.
Nontrivial ones are entire; the trivial one is ζ_(K^C), holomorphic away from 1.
Thus, at our s0, all coefficients of Θ_C are **nonnegative integers**.

By (1), Res_C Θ_G=Θ_C is a genuine character of dimension Θ_C(1)=M. A finite-order
matrix can be made unitary; its eigenvalues have absolute value one. Therefore

    |Θ_G(g)| = |Θ_C(g)| ≤ M,     and M≥0.

Orthogonality of irreducible characters now gives

    Σχ nχ² = <Θ_G,Θ_G> = (1/|G|)Σg |Θ_G(g)|² ≤ M².           (4)

**Important scope:** restrictions to cyclic subgroups are genuine characters;
Θ_G itself need not be. This is exactly why the proof is useful without the
Artin holomorphy conjecture. No zero-location assumption entered the proof.
The original FM proof has not been retrieved here; MR's original statement and
proof framework have been inspected, and the argument above supplies the needed
inequality rather than asking the reader to trust an inaccessible citation.

### 1.3 What Stark's simple-zero lemma actually says

**[known; S74 Theorem 3, cross-checked in MR introduction and W19 Proposition
3.3; proof from (3)–(4) here.]**

* M=0 forces every nχ=0: every irreducible Artin factor is holomorphic and
  nonzero at s0.
* M=1 forces exactly one nχ to be nonzero. It must equal +1 and have χ(1)=1:
  the norm permits only one ±1, and the weighted sum in (3) excludes a minus
  sign or larger degree. All other irreducible factors have order zero.
* Hence a simple ζ_K zero is a simple zero of **exactly one degree-one factor**.
  It need not be an ordinary-zeta zero.
* If ζ_k also vanishes there, that factor is the trivial character: m=1 and
  all nontrivial irreducible factors are nonzero there.

**[known; further consequence]** Put E0=K^(ker χ) for the unique character in
the M=1 case. E0/k is cyclic. Formula (2) says an intermediate E has a zero at
s0 exactly when E contains E0; in that case its order is one. At a *real* s0,
conjugation would pair any non-real χ with a different character. Uniqueness
therefore forces χ to be real: E0 is k or a quadratic extension. This explains
the real-simple-zero descent in Heilbronn's 1973 Theorem 1. It does not turn a
nonreal simple zero into a quadratic zero, and does not locate its real part.
The 1972 Boulder contribution is separately listed in §9; its original proof
was not retrieved. A 1972 receipt date on the 1973 paper is not a substitute.

**[known; immediate budget]** For k=Q, m≥0,

    Σχ≠1 nχ² ≤ M²−m².                                      (5)

Thus **independent** information M=m forces all nontrivial orders to be zero,
even without a global Artin-holomorphy theorem. Neither (4) nor (5) tells us M
at a hypothetical ordinary zero. Aramata–Brauer already gives M≥m.

### 1.4 Murty–Raghuram's stronger solvable-group inequalities

**[known; MR Theorems 4.1, 4.5, 4.6, 5.1, 5.4 and Corollary 5.6; original
pages 236–243 inspected.]** If G is **solvable**, and χ0 is any degree-one
character, their quotient-holomorphy theorem permits deletion of its contribution:

    Σχ≠χ0 nχ² ≤ (M−nχ0)².                                  (6)

For χ0=1 this is stronger than (5): its right side is (M−m)², **not** M²−m².
Do not promote this solvable-group refinement to arbitrary G without a theorem.

Here is the key extra step. For every cyclic C and ψ∈Ĉ, their Theorem 4.1 gives
holomorphy away from 1 of

    L(s,Ind_C^G ψ) / L(s,χ0)^<Res_C χ0,ψ>.

Its order is the ψ-coefficient of Res_C(Θ−nχ0χ0), and is nonnegative. Thus the
truncated virtual character again has genuine cyclic restrictions, now of
dimension M−nχ0. The same norm proof yields (6). Deleting a coefficient from a
virtual character *without* this theorem would leave a gap.

**[known; MR Theorem 5.4]** Let S be all degree-one irreducible characters and
K^ab=K^[G,G]. One can delete them together:

    Σχ∉S nχ² ≤ (M−ord_s0 ζ_(K^ab))².                         (7)

The stronger quotient theorem in MR 5.1 makes this truncation legitimate.
In particular ζ_K/ζ_(K^ab) has no simple zeros or poles (MR Corollary 5.6).
This is a real restriction on the **nonabelian quotient**, not on ordinary zeta.
It is stronger than a bare assertion that the quotient is entire.

**[known; scope audit]** FM and subsequent Foote–Wales/Wong results also yield
local Artin holomorphy from sufficiently small field or quotient order, with
solvability and prime-divisor qualifications specified in the earlier note.
None says that all field zeros are simple. None supplies a uniform upper bound
on M at every ordinary zero. For our monomial Q8 example Artin holomorphy is
already unconditional, so these holomorphy theorems are not a new input there.

## 2. Q8 in full, including the redundant part

**[known; elementary character theory]** Class sizes are 1,1,2,2,2:

| character | 1 | −1 | ±i | ±j | ±k | degree |
|---|---:|---:|---:|---:|---:|---:|
| 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| a | 1 | 1 | 1 | −1 | −1 | 1 |
| b | 1 | 1 | −1 | 1 | −1 | 1 |
| c | 1 | 1 | −1 | −1 | 1 | 1 |
| ρ | 2 | −2 | 0 | 0 | 0 | 2 |

Write Θ=m0·1+ma·a+mb·b+mc·c+r·ρ and S=m0+ma+mb+mc. Direct evaluation gives

    Θ(1)=S+2r=M,     Θ(−1)=S−2r,
    Θ(i)=m0+ma−mb−mc,
    Θ(j)=m0−ma+mb−mc,
    Θ(k)=m0−ma−mb+mc.

Substituting these values into the weighted character norm gives exactly

    [M²+(S−2r)²+2Θ(i)²+2Θ(j)²+2Θ(k)²]/8
       = m0²+ma²+mb²+mc²+r² ≤ M².                           (8)

**[known; explicit restriction]** On C_i=<i>≅C4, let ψ(i)=i. Then

    Res 1=Res a=1,   Res b=Res c=ψ²,   Res ρ=ψ+ψ^−1,
    Res Θ=(m0+ma)1+(mb+mc)ψ²+rψ+rψ^−1.

The other two cyclic order-four subgroups give the analogous formulas. This
both checks (1) concretely and exhibits ρ=Ind_(C_i)^Q8 ψ. Hence its L-function
is a nontrivial Hecke L-function of a quadratic field and is entire. The linear
characters give Dirichlet/Hecke factors. All five orders in our strip are
therefore **nonnegative**, without assuming any RH.

**[project; the important self-critique]** Once all nχ≥0 are already known,

    Σ nχ² ≤ (Σ nχ)² ≤ (Σ χ(1)nχ)²=M²

is elementary. Thus (8) adds **no new constraint** in this Q8 case. Even (6) and
(7) are automatic for these already-holomorphic factors. For example, (7)
reduces to r²≤(2r)². The character identities remain useful as exact extraction
maps; the norm bound is not an extra analytic weapon against ζ in this example.

### 2.1 The specified field and its exact projectors

**[project; exact field identification, known Artin formalism]** Our identifier
`2.42025.8t5.a.a` is the Artin-representation label used by the project. The
associated totally real degree-eight Q8 field is defined here by

    x^8−3x^7−63x^6+90x^5+1311x^4−20x^3−7702x^2−5524x−1009,
    disc(K)=205^6,    conductor(ρ)=42025=205²,    gamma shifts=(0,0).

Let A_D=L(s,χ_D), B=L(s,ρ), F_D=Q(√D), and E=Q(√5,√41). Then

    ζ_K=ζ A5 A41 A205 B²,        ζ_E=ζ A5 A41 A205,
    ζ_(F_D)=ζ A_D.

For arbitrary s0≠1 the subgroup-invariant dimensions give

    m_(F5)=m0+m5,   m_(F41)=m0+m41,   m_(F205)=m0+m205,
    m_E=m0+m5+m41+m205,    m_K=m_E+2r.

In particular

    2m0=m_(F5)+m_(F41)+m_(F205)−m_E,
    2r=m_K−m_E.                                             (9)

Every proper nontrivial subgroup of Q8 is one of its three C4's or its center C2;
these are all its nontrivial proper subfields. The character identity behind the
first projector is

    2·1_G = Σ_(three C4) Ind_C4^G 1 − Ind_C2^G 1.

It has a **negative coefficient**. Positivity of each subfield count alone does
not make the projected ordinary component positive on arbitrary tests.

## 3. Can many fields make a hypothetical ζ zero impossible?

### 3.1 A formal countermodel to the bookkeeping-only strategy

**[project; exact algebraic obstruction, not a construction of actual L-functions]**
Suppose ζ has a nontrivial zero at a noncentral s0, 0<Re(s0)<1, of multiplicity m≥1. For **every**
finite Galois group in a family or tower, assign

    Θ_G=m·1_G,       n_1=m,       nχ=0 for χ≠1,       M=m.    (10)

This assignment satisfies every restriction identity: Res_H(m1_G)=m1_H. Every
intermediate field zeta has order m by (2). It satisfies factorization,
Aramata–Brauer, and the norm inequality with equality m²=M². The refined
trivial-component inequality is 0≤0. Other admissible truncations and all
subfield projectors are consistent too. Apply the same assignment at reflected
and conjugate points to respect functional equations.

This works for m=1 or m>1, and whether Re(s0)=1/2 or not. It is compatible across
composita and an arbitrarily large family. No number of repetitions of those
identities can reject it. At the forced central zeros of sign-minus factors
there are additional constraints, but ζ(1/2) is already known nonzero; those
central constraints do not force coincidence at a noncentral hypothetical ζ zero.

**[project]** This is a limitation of these premises, **not** a theorem that no
future argument using number fields can prove RH. It shows precisely what an
argument must add to escape a consistent bookkeeping model.

### 3.2 What extra input would actually bite?

**[known implications; speculative premises]**

1. An independent zero-free region for one ζ_K descends to ζ. If it excluded all
   off-line zeros it would prove RH, by the existing no-cancellation bridge.
2. An independent M≤1 at a given ordinary zero forces simplicity and factor
   separation there. It does **not** force Re(s0)=1/2: an off-line simple zero
   passes Stark's test perfectly well.
3. An independently forced nontrivial-factor order at the *same* s0, combined
   with an incompatible independent upper bound on M, could give a contradiction.
   In the holomorphic Q8 case the linear inequality M≥m+2r is already stronger
   than the norm budget for this purpose. We have no such globally forced
   coincidence or small upper bound at hypothetical off-line ordinary zeros.
4. For unscaled off-line counts, N_ζ(D)≤N_K(D). A proved family average <1 in a
   fixed forbidden region would exclude even one ordinary zero there. The hard
   theorem is that upper bound, not averaging the shared factor itself.

**[known; scaling correction]** Fixed-window low-lying density is different.
At analytic conductor C, a fixed ordinary ordinate γ contributes at scaled
coordinate γ log C/(2π), which tends to infinity. Ordinary ζ is a fixed factor,
not a zero pinned at the center of the family's rescaled window. Standard
one-level density therefore cannot be substituted for the fixed-region bound
in item 4. Bulk pair correlation and prime-coefficient orthogonality likewise
are not pointwise noncoincidence or location theorems.

## 4. Fresh interval certificates: exactly what was proved

### 4.1 Outputs and local orders

**[project; certified, conditional only on the standard exact arithmetic/
analytic library and mathematical input trust described below]** Script:
`scripts/heilbronn_local_certificate.py`. Results:
`results/heilbronn-local/certificate.json`; full log: `run.log` in that directory.
Figure: `figures/heilbronn-local-orders.png` and `.svg`.

For each row, the contour is the rational square with half-width 10^−6 in both
coordinates, centered at 1/2 plus the listed rational seed height. Decimal
heights are **not claimed to be exact zeros**. For the noncentral rows, fresh
opposite-sign enclosures for ξ on the critical line in a rational bracket of
half-width 10^−12 prove an actual zero inside the square. An ordinary contour
count of one proves uniqueness; the existing height-30 ordinary certificate
also establishes these are precisely the first three positive ordinary zeros.

| target point | ζ | ζ_F5 | ζ_F41 | ζ_F205 | ζ_E | ζ_K | A5,A41,A205 | B |
|---|---:|---:|---:|---:|---:|---:|---|---:|
| 1/2 | 0 | 0 | 0 | 0 | 0 | 2 | each 0 | 1 |
| actual zero near 1/2+14.13472514173469i | 1 | 1 | 1 | 1 | 1 | 1 | each 0 | 0 |
| actual zero near 1/2+21.02203963877155i | 1 | 1 | 1 | 1 | 1 | 1 | each 0 | 0 |
| actual zero near 1/2+25.01085758014569i | 1 | 1 | 1 | 1 | 1 | 1 | each 0 | 0 |

These are **orders**, not function values. The Artin/quadratic factors are
certified nonzero on the entire three noncentral squares, not merely at the
root midpoints. At the center, the ordinary/quadratic factors are nonzero on
the whole square. The exact root number −1 forces B(1/2)=0; a total field count
of two then forces that B zero to be simple. A ball containing zero alone was
never accepted as evidence of a zero.

**[known implication; project application]** The noncentral field order one,
together with ordinary-root existence, triggers Stark's lemma and gives order
one for ζ and order zero for all other irreducible factors. Formula (9) agrees:
2m0=1+1+1−1=2 and 2r=1−1=0. At the center it gives m0=0 and 2r=2−0=2.

### 4.2 Enclosures, explicit tails, and the argument principle

**[project; reproducible certificate construction]** We used python-flint Arb
at 320 bits; 5000 exact Artin coefficients; integration cutoff U=4096; twelve
dyadic bands [1,2],…,[2048,4096]; 96-point Gauss–Legendre on each band; and a
local Taylor polynomial of degree 12. All geometric endpoints are rational.
PARI verifies the field discriminant/signature, Artin degree, conductor and
gamma parameters. The 5000 integers are freshly regenerated and match the
previous audited list. The coefficient bound |a_n|≤d(n)≤n follows from the
finite-image unitary two-dimensional Euler factors, including the ramified
factors. Correct identification of this Artin representation is an explicit
trust boundary, not something established by a winding number.

The completed quadratic factors are evaluated by the **finite** Hurwitz sums

    Λ_D(s)=(Dπ)^(−s/2) Γ(s/2) Σ_(a=1)^(D−1) χ_D(a) ζ(s,a/D).

Ordinary ξ uses Arb's Riemann-zeta and gamma implementations. No floating
quadratic zero lists are used to prove nonvanishing.

For the two-dimensional factor, put c=2π/205 and

    θ(u)=4Σ_(n≥1) a_n K0(cnu),
    Λ_B(s)=∫_1^∞ θ(u)[u^s−u^(1−s)] du/u.                   (11)

This is the standard Mellin identity plus its known Artin functional equation;
the root sign is independently selected as described in §4.3. K0(x)≤sqrt(π/(2x))
exp(−x), |a_n|≤n, u≥1, and −1≤Re(s)≤2 give a deliberately coarse integrand bound

    |integrand| ≤ C u Σ_(n≥1) n² exp(−cnu),
    C=8 sqrt(π/(2c)).

It follows that the omitted-coefficient and omitted-integral tails are bounded by

    E_n=C(1/c+1/c²) Σ_(n>5000) n² exp(−cn),
    E_u=C(U/c+1/c²) Σ_(n≥1) n² exp(−cnU).

Both geometric power sums are evaluated explicitly, not estimated by empirical
convergence. Their resulting upper bounds are approximately 1.426×10^−53 and
2.321×10^−48. The actual balls, rather than these rounded summaries, are used.

For a dyadic interval [l,2l], analytically continue the integrand to the circle
center 3l/2 and radius l. There Re(u)≥l/2, |u|≤5l/2, and
|arg u|≤α=arcsin(2/3). The integral representation of K0 gives
|K0(cnu)|≤K0(cn Re u). For −1/2≤Re(s)≤3/2 and |Im(s)|≤26, a valid bound is

    B_l = (16/l)(5l/2)^(3/2) exp(26α)
                     Σ_(n≤5000) |a_n| K0(cnl/2).

Cauchy's estimate bounds the Taylor tail on [l,2l] after degree 2q−1 by
2B_l·2^(−2q). Gauss q-point exactness through that degree, its positive weights,
and the integral length l bound the quadrature error by

    4l B_l 2^(−2q),       q=96.

Summed over the bands, this is less than 6.729×10^−45. The bound encloses the
quadrature error for the **full complex box**, not only sampled heights.
These are analytic truncation/quadrature error bounds, not total radii of all
computed balls. Arb separately propagates special-function, root/weight and
rounding uncertainty; those radii can dominate the analytic tails. Working at
320 bits does not assert that every final value has 320 accurate bits.

To avoid interval cancellation, the paired exponential in (11) is Taylor
expanded at each square's center. If h bounds |s−s_center| and ℓ=log u, each
node's exponential remainder is bounded by

    2|B_node|u exp(hℓ)(hℓ)^13/13!,
    B_node=(quadrature weight) θ_truncated(u)/sqrt(u).

The extra u≥1 is conservative; the centers have real part exactly 1/2. This
nodewise remainder plus E_n+E_u+the quadrature error encloses every full contour
segment. There is no unproved derivative-error estimate in the certificate.

Each completed product segment is enclosed in a convex complex Arb rectangle
excluding zero. Subdivision continues until this holds for every field product
and the endpoint ratio has positive real part. On such a convex zero-free image
the continuous argument lift has the same increment as the principal argument
of the endpoint ratio. Summing these interval increments and dividing by 2π
isolates a unique integer. Analyticity inside each small square then gives the
zero count by the argument principle. Accepted segment counts are 16,256,128,128
for the center and the three ordinary zeros respectively. The JSON retains
all segment images and increments, not just a final Boolean.

### 4.3 An additional root-number certificate, not a numerical sign guess

**[project; certified]** Self-duality makes the exact Artin root number W one
of +1 and −1. The two possible Mellin formulas at s=2 can be evaluated using
the same θ kernel, with a separate analytic-circle quadrature bound appropriate
to Re(s)=2 and Im(s)=0. Their enclosures are near

    W=−1: 3947.07144984834135,
    W=+1: 3952.03320957863029.

For an independent check at this point, use the Artin Euler product over primes
p≤5000, with determinant-one good factors

    (1−a_p/p²+1/p^4)^−1.

At p=5,41 the inertia subgroup is nontrivial. No nonidentity element in the
faithful quaternionic representation has eigenvalue one, so the inertia-fixed
space is zero and these local factors are 1. The good eigenvalues have modulus
one. The omitted Euler product therefore has logarithm of absolute value at most

    2/[5000·(1−1/5001²)] < 0.000400001.

After multiplication by (205/π)², its rigorous completed-value enclosure is

    3947.07423269763131 ± 1.58   (rounded outward summary).

This intersects the W=−1 candidate and is disjoint from the W=+1 candidate.
Since exactly one sign is allowed by the functional equation, W=+1 is excluded
and W=−1 is proved within the stated library/arithmetic trust. This removes a
previously implicit root-sign input from the new local calculation. The JSON
contains the unrounded balls and the explicit Euler-tail bound.

### Reproduction and saved-witness audit

**[project]** From this directory:

```sh
/workspace/tools/pari-venv/bin/python scripts/heilbronn_local_certificate.py
/workspace/tools/pari-venv/bin/python scripts/verify_heilbronn_local.py
python3 scripts/plot_heilbronn_local.py
```

The generator completes all four cases, including the sign certificate. The
saved-witness audit passes for **528 full segment images**, all four winding
vectors, source hash and root-sign separation. It additionally checks exact Q8
orthogonality/projectors on 3125 signed order vectors and the compatible
trivial-character model for m=1,…,100. The general argument for arbitrary m
and arbitrary G is §3.1, not extrapolation from that finite test. Library versions:
python-flint 0.9.0, cypari2 2.2.4. Verification outputs are
`results/heilbronn-local/verification.json` and `verification.log`.

### 4.4 What this computation is not

**[project; trust and independence audit]** It is not a direct degree-eight
Dedekind-zeta analytic evaluator; the field products share factor oracles. It
is not a proof about a literal rounded decimal point. It does not prove
noncoincidence at every ordinary zero, any global simplicity conjecture, or RH.
The previous `combined-certificate.json` already separated all factor zeros
through height 30. This rerun supplies especially transparent local witnesses,
a root-sign check, and a concrete implementation of (9), rather than a new
height theorem. The checker below verifies saved interval arithmetic/geometry
and hashes; independently reproducing the analytic special-function enclosures
requires rerunning the generator.

## 5. Simultaneous zeros: distinguish four very different statements

**[known/project; logical separation]**

1. **Simple zeros:** each zero of a given primitive factor has order one.
2. **No common zero:** two distinct primitive factors never vanish at exactly
   the same complex point (with forced central exceptions treated separately).
3. **Different zero multisets:** infinitely often their zero multiplicities
   differ. This allows both functions to vanish at the same point with orders
   one and two. It is much weaker than item 2.
4. **Statistics / linear independence:** correlation laws concern averaged
   spacings. A conjecture of rational linear independence of the *indexed*
   positive ordinates, with multiplicity, is much stronger and would imply
   simplicity and no coincidences in the class it actually covers. A set with
   repetitions silently discarded would not imply simplicity.

**[known conjecture formulation; RSa94 p.176, original read]** Rubinstein–Sarnak
state the Grand Simplicity Hypothesis for critical-zero ordinates of primitive
Dirichlet L-functions and explicitly note its intended consequences of simplicity
and central nonvanishing. Their paper assumes GRH separately; this conjecture
is not a proof of GRH. In §5, pp.194–195, they **already warn** that nonabelian
Dedekind products have repeated factors, and that even a primitive Artin factor
can have a forced central zero. They cite Armitage's example and its odd
functional equation, arising from Serre's conductor/root-number calculations.
Our caution is therefore established territory, not a newly discovered meeting
of quaternionic and ordinary-zeta theories.

**[known; B24 Theorem 1.1, original read]** Banks proves that distinct primitive
Dirichlet characters **of the same modulus q** have different zero multisets in
0<Re(s)<1, T<Im(s)<T+c2 q^θ log T for every sufficiently large T≥c1 q^θ, θ>1/3;
θ>1/4 suffices for cubefree q. This is unconditional, but the theorem explicitly
counts multiplicity. It does not rule out any prescribed common zero. It does
not directly compare ζ with our degree-two B. Banks's introduction also records
Fujii's positive-proportion multiplicity differences and Raghunathan's infinitely
many poles of quotients of distinct primitive Dirichlet factors. Those original
older proofs were not retrieved here; they are not used in our certificates.

**[known literature, proof not retrieved here; BP98]** Bombieri–Perelli study
quantitative distinctness of L-function zero multisets under stated analytic,
Euler-product/orthogonality and density assumptions. One must preserve those
hypotheses and the multiplicity-difference definition; their title alone does
not prove universal pairwise disjointness. The publisher's actual PDF link
returned an error in this audit, so no detailed theorem from that paper is used
as a premise here. See the source audit for rejected wrong-title downloads.

**[known; RS96, scope rather than a new application]** Rudnick–Sarnak's
n-level correlation work relates principal automorphic L-functions to unitary
random-matrix statistics under its hypotheses and restricted Fourier support;
it does not prove the grand simplicity/linear-independence conjecture or
zero-by-zero disjointness. The original proof was not retrieved in this pass.
We do not import its conclusions outside that theorem's class or hypotheses.
For example a blanket independence conjecture for *all Dedekind zero lists*
cannot be right: they deliberately share ζ, and Q8 field zeta has a B² factor.
Any primitive-factor conjecture must also handle the root-number-forced central
zeros of self-dual families. A “primitive” degree-two quaternionic Artin factor
and a full Dedekind product must not be conflated.

**[project; implication check]** Even proving both simplicity and noncoincidence
for every ordinary zero against every B in our family would allow off-line
ordinary zeros. It would give M=m=1 at those zeros, which satisfies Stark and all
budgets. Location and multiplicity are logically different questions. The
conjugate/reflected symmetry of a hypothetical off-line simple zero is not a
multiple zero at a single point.

## 6. All-index Li positivity: confirmed, with the quantifier intact

**[known; L97; BL99 Theorem 1, Corollary 1 and introduction, original rendered
pages 1–4 inspected]** For the completed Dedekind ξ_K, with its pole-removal
factor s(s−1), define

    λ_n(K)=1/(n−1)! [d^n/ds^n {s^(n−1) log ξ_K(s)}]_(s=1).

The nontrivial zero multiset has the necessary reflection/conjugation symmetry,
excludes 0 and 1, and satisfies the convergence condition in BL99. Their abstract
multiset criterion therefore applies. In the usual symmetric limiting convention,

    λ_n(K)=Σ_ρ [1−(1−1/ρ)^n],
    [λ_n(K)≥0 for EVERY integer n≥1] ⇔ GRH for ζ_K.           (12)

BL99's Theorem 1 uses a half-plane formulation with inverse powers. Its Corollary
1 explicitly adds ρ↔1−ρ and conjugation symmetry and yields the line criterion
with the displayed positive-index convention. Thus there is no sign/convention
gap in using it for Dedekind zeta. Their introduction expressly states Li's
Dedekind-zeta extension.

**[known implication]** For Galois K/Q, Aramata–Brauer and (12) give

    all-index field Li nonnegativity
        ⇒ every ζ_K nontrivial zero lies on Re(s)=1/2
        ⇒ every ordinary ζ zero lies there
        ⇒ all-index ordinary Li nonnegativity.              (13)

This is not conditional on Artin holomorphy. It *is* conditional on the unproved
all-index positivity premise. For our holomorphic Q8 factors,

    λ_n(K)=λ_n(ζ)+λ_n(A5)+λ_n(A41)+λ_n(A205)+2λ_n(B).

At a fixed n, positivity of this sum does not isolate λ_n(ζ). But if all indices
are nonnegative, (13) applies; saying “a positive sum says nothing about the
pieces” without this distinction would be false. No finite N, including the
20 coefficients previously computed, verifies (12). Those old Taylor-derived
coefficients remain high-precision numerical values, **not** interval-certified
by the zero certificates in §4. A certified truncated zero sum would require its
own Li-tail bound.

## 7. Hard self-critique and corrections to the earlier pass

**[project; audit of `think-hard-transfer.md`]**

| Earlier wording or temptation | Correction / precise disposition |
|---|---|
| “Substantially sharper bridge,” “single most promising” | Correct as a general local Artin-holomorphy/extraction framework, but not evidence of a stronger Q8 RH attack. Q8's norm inequality is redundant once its known monomial holomorphy is used. |
| Applying Stark to a factored field contour sounds independently downward | Valid implication, but the input has already used ordinary/factor evaluations. The new certificate says this prominently; the old height-30 factor certificate already implied the result. |
| More fields might overconstrain an ordinary multiple or off-line zero | The compatible assignment Θ_G=m1_G defeats all the bare bookkeeping identities, even across a family/tower. A separate analytic bound is indispensable. |
| Family average and one-level density can be interchanged | Only counts in the same unscaled fixed region have the common-factor lower bound. Fixed-window low-lying scaling sends each fixed ordinary ordinate to infinity. |
| “Simple zero” might mean simple ordinary zero | Stark starts with a simple **field** zero. The unique linear character may be nontrivial; an ordinary-zero existence input is needed to identify it as trivial. |
| Deleting the trivial coefficient is just algebra | The sharper RHS (M−m)² requires the solvable quotient-holomorphy theorem; norm subtraction alone gives M²−m². The two are not interchangeable. |
| Heilbronn 1972 and 1973 can be cited as one inspected proof | The Boulder 1972 contribution is unretrieved; the 1973 journal proof is retrieved. Stark and FM original proofs remain unretrieved; MR's original is now retrieved and the needed elementary proofs are supplied. |
| Finite positive field Li coefficients might certify a component | Finite-index inference fails, but all-index implication (13) is valid. Original BL99 now confirms the Dedekind case and convention. |
| A root number near −1 is automatically a certificate | New Euler/theta comparison excludes the exact alternative +1, rather than rounding a numerical root-number estimate. |
| A downloaded PDF named for an author is a source | Three wrong-title attempts were quarantined; titles/pages, not filenames or HTTP success, are the acceptance test. |

The earlier direction of Aramata–Brauer descent and the all-index Li conclusion
were already correct. This revision narrows interpretive optimism, adds the
stronger solvable truncations with their real hypotheses, and makes proof and
computation independence explicit. It does not advertise a corrected RH proof.

## 8. Ranked next steps, with pass/fail tests

1. **[project proposal; best verification next step] Independent subfield input.**
   Implement one direct subfield-zeta enclosure, initially for the small quadratic
   fields, using ideal/theta data rather than ζ·A_D. Then use the biquadratic
   subfield and projector (9) on one existing square. **Test:** the independent
   evaluator makes no ordinary-zeta/Artin-factor calls, supplies explicit tails,
   and recovers order one through the signed projector. Agreement tests a genuine
   downward data path. This is bounded analytic-software work, not currently done;
   direct degree-eight certification is harder and should not be promised first.
2. **[project proposal; doable finite extension] Extend factor-separation witnesses
   to the next ten ordinary zeros.** Increase the uniform height in the analytic
   quadrature bound rather than blindly recycling H=26. **Test:** certify ordinary
   root existence, field count one, and nonzero nontrivial factors on disjoint
   boxes, or record unresolved boxes with their error budget. The output is a
   finite noncoincidence dataset, not a theorem of global simplicity or RH.
3. **[project proposal; best analytic gate before calling this an RH approach]**
   Audit one proposed family multiplicity/zero-density upper bound against (10),
   retaining its full discriminant, degree, height and scaling costs. **Test:**
   identify an independently proved bound that actually excludes (10) at an
   off-line ordinary zero; if no such bound exists, record the failed gate and
   do not invest in a larger family merely to repeat compatible identities.

**[project; final ranking]** (1) No-cancellation field factorization and exact
subfield projectors: genuinely load-bearing. (2) Heilbronn/Stark local
multiplicity budgets: genuinely load-bearing in general, useful diagnostics but
analytically redundant in holomorphic Q8. (3) Universal Li/Weil positivity:
exact, but its missing universal input remains RH-sized. (4) Family statistics
and geometric analogy: no pointwise RH descent without a new bridge theorem.

## 9. Primary sources, exact scope, and retrieval status

* **H72.** H. Heilbronn, *On real simple zeros of Dedekind ζ-functions*,
  Proceedings of the Number Theory Conference, University of Colorado, Boulder
  (1972), 108–110. **Original not retrieved.** Bibliographic attribution checked
  in later literature; no claim to have inspected its proof.
* **H73.** H. Heilbronn, *On Real Zeros of Dedekind ζ-Functions*, Canadian J.
  Math. 25 (1973), 870–873, Theorem 1 and its proof.
  <https://doi.org/10.4153/CJM-1973-090-3>.
  **Retrieved original:** `sources/think-hard/heilbronn-1973.pdf/.txt`.
* **S74.** H. M. Stark, *Some effective cases of the Brauer–Siegel theorem*,
  Invent. Math. 23 (1974), 135–152, Theorem 3.
  <https://doi.org/10.1007/BF01405166>.
  **Original proof not retrieved.** Statement cross-checked in MR and W19;
  the particular simple-zero consequence needed here is proved in §1.3.
* **FM.** R. Foote and V. Kumar Murty, *Zeros and poles of Artin L-series*,
  Math. Proc. Cambridge Philos. Soc. 105 (1989), 5–11.
  <https://doi.org/10.1017/S0305004100001316>.
  **Original proof not retrieved:** publisher PDF route returned HTML, not PDF.
  Norm theorem read in MR 4.4; explicit proof reconstructed in §1.2.
* **MR.** M. Ram Murty and A. Raghuram, *Some Variations on the Dedekind
  Conjecture*, J. Ramanujan Math. Soc. **15 no.4** (2000), 225–245.
  <https://mrammurty-ai.github.io/murty-raghuram.pdf>.
  **Original retrieved and inspected**, particularly 4.1–4.6, 5.1, 5.4 and 5.6.
  PDF is a sideways two-up scan; readable renders are archived. Do not replace
  this bibliographic data by incorrect secondary volume/issue/page entries.
* **W19.** P.-J. Wong, *Zeros of Dedekind zeta functions and holomorphy of Artin
  L-functions*, J. Number Theory 197 (2019), 282–295.
  Original already archived at `sources/think-hard/wong-loworder.pdf/.txt`;
  Propositions 3.1–3.3 and Theorem 3.5 cross-check the main statements here.
* **L97.** X.-J. Li, *The Positivity of a Sequence of Numbers and the Riemann
  Hypothesis*, J. Number Theory 65 (1997), 325–333.
  <https://doi.org/10.1006/jnth.1997.2137>. Dedekind extension also explicitly
  confirmed in BL99's introduction. No new full retrieval of L97 in this pass.
* **BL99.** E. Bombieri and J. C. Lagarias, *Complements to Li's Criterion for
  the Riemann Hypothesis*, J. Number Theory 77 (1999), 274–287.
  <https://doi.org/10.1006/jnth.1999.2392>;
  <https://websites.umich.edu/~lagarias/doc/bombieri.pdf>.
  **Original retrieved:** introduction, Theorem 1, Corollary 1 and relevant
  proof inspected on rendered pages. Raw PDF text extraction has broken glyphs;
  it must not be mistaken for readable theorem text.
* **B24.** W. Banks, *Comparing zeros of distinct Dirichlet L-functions*,
  arXiv:2302.07073v2 (21 May 2024), Theorem 1.1 and introduction.
  <https://arxiv.org/abs/2302.07073>.
  **Original retrieved and read**, `sources/heilbronn-deep/comparison-2023.pdf/.txt`.
* **BP98.** E. Bombieri and A. Perelli, *Distinct zeros of L-functions*, Acta
  Arith. 83 (1998), 271–281. <https://doi.org/10.4064/aa-83-3-271-281>.
  **Original PDF not retrieved:** actual publisher download product 110104
  returned HTTP 500. No detailed theorem is used as a premise above.
* **RSa94.** M. Rubinstein and P. Sarnak, *Chebyshev's Bias*, Experimental
  Mathematics 3 no.3 (1994), 173–197, p.176 and §5 pp.194–195.
  <https://doi.org/10.1080/10586458.1994.10504289>;
  <https://personal.math.ubc.ca/~gerg/teaching/592-Fall2018/papers/1994.Rubinstein.pdf>.
  **Original retrieved and read:** GSH formulation and the explicit caveats for
  repeated Dedekind factors and Armitage's forced primitive Artin central zero.
* **RS96.** Z. Rudnick and P. Sarnak, *Zeros of principal L-functions and
  random matrix theory*, Duke Math. J. 81 (1996), 269–322.
  <https://doi.org/10.1215/S0012-7094-96-08115-6>.
  **Original proof not retrieved in this pass.** Cited for the distinction
  between correlation theorems and conjectural independence, not for an added
  zero-disjointness or RH theorem.

**[project; rejected-source audit]** Under `sources/heilbronn-deep/misidentified/`
are deliberately quarantined wrong-title downloads: a Murty–Pathak special-value
paper, a Butz–Moerdijk logic paper, and Coulter's Weil-sum paper. None supports a
claim here. `README.md` in the source directory records accepted titles and
failed retrievals. A missing/failed fetch is not evidence that a paper or proof
does not exist.
