Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 26 additions & 0 deletions .github/workflows/h3-solver-pilot.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,26 @@
name: h3-search-and-certify-pilot
on:
push:
pull_request:
permissions:
contents: read
jobs:
scalar-replay:
runs-on: ubuntu-latest
timeout-minutes: 10
steps:
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
with:
persist-credentials: false
- name: Exact arithmetic and adversarial controls
run: |
python3 -m unittest discover -s tests -p test_h3_solver_pilot.py -v
python3 -O -m unittest discover -s tests -p test_h3_solver_pilot.py -v
- name: Cold candidate reproduction (not scientific acceptance)
run: |
expected=$(python3 -c "import json; print(json.load(open('engine/solver_pilots/h3_20260921/CERTIFICATE_IDENTITY.json'))['canonical_sha256'])")
python3 engine/solver_pilots/h3_20260921/h3_solver.py \
--source-archive research/campaigns/h3_rn_n6_20260920_v1.zip \
--context engine/solver_pilots/h3_20260921/CONTEXT.json \
--expected-certificate-sha "$expected" \
--out "${RUNNER_TEMP}/h3-scalar-${GITHUB_RUN_ID}.json" --search --cold
149 changes: 149 additions & 0 deletions engine/solver_pilots/h3_20260921/CERTIFICATE.json

Large diffs are not rendered by default.

10 changes: 10 additions & 0 deletions engine/solver_pilots/h3_20260921/CERTIFICATE_IDENTITY.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
{
"canonical_encoding": "UTF-8 JSON, sorted keys, compact separators, ensure_ascii=false, allow_nan=false",
"canonical_sha256": "dfc9c25f1ebdd6545b56bf1ecc3606cd3b950a8f2b5e779bcfe11b88f94c7dea",
"coefficient": "1747/1000",
"file_location": "Complete CERTIFICATE.json in the Drive delivery and downloadable package; not an authority token",
"raw_bytes": 18967,
"raw_sha256": "c26d9be8134de716c2f42378d09e7121855279fec1a068c2e4b5e7ef95823095",
"schema": "h3-certificate-identity-v1",
"status": "AUTHOR_SIDE_CANDIDATE_REVIEW_OPEN"
}
10 changes: 10 additions & 0 deletions engine/solver_pilots/h3_20260921/CONTEXT.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
{
"context_id": "H3-SOLVER-CANDIDATE-20260921",
"scientific_use": "REQUIRES_REVIEW",
"source_usable_for_candidate": true,
"basis": "Explicit field definition and frozen H3 analytic proof; source-exposed continuation, not current theorem admission",
"scope": "2D SIDE24 fixed x axis; canonical six-pin law; b=6/5; gap=r^3/6; 0<r<=1/20",
"source_archive_sha256": "73b9e63800f77c677c504f3097fe0bb2459ff6c94cdee5772b56062320aefc21",
"current_use_requires_revalidation": true,
"organizational_independence_credit": 0
}
200 changes: 200 additions & 0 deletions engine/solver_pilots/h3_20260921/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,200 @@
# H3 scalar optimization: a whole-band lower coefficient of 1747/1000

**Author-side proved candidate; independent-method arithmetic, not independently
authored review.** This is a numbered additive extension of the fixed-axis H3
argument in `h3_floor/PROOF.md`, frozen inside source archive SHA-256
`73b9e63800f77c677c504f3097fe0bb2459ff6c94cdee5772b56062320aefc21`.
The source proof is 10,423 bytes, SHA-256
`b48d8baa39b80be9a15af9833234d9339ea0afa4cdff2e1c60ae9ef79f5cb4fa`.
No source body or previous scientific disposition is replaced.

## Statement and exact scope

Let

\[
k(s)=\frac{\sum_{j\in\mathbb Z}e^{-(s+24j)^2/2}}
{\sum_{j\in\mathbb Z}e^{-(24j)^2/2}},\qquad K(x,y)=k(x)k(y).
\]

Use the centered stationary Gaussian field with covariance K, the fixed x-axis
points M=(-r/2,0), S=(r/2,0), and canonical Gaussian conditioning on
f(M)=6/5, f(S)=6/5-r^3/6 and both gradients zero. Define Z(r) as the conditional
expectation of |det H_M det H_S| times the indicator that M is a maximum and S
is a saddle. Then the argument below and the accompanying rational certificate give

\[
\boxed{Z(r)\ge \frac{1747}{1000}r^2,\qquad 0<r\le\frac1{20}.}
\]

The coefficient is 47/1700, approximately 2.7647%, larger than the prior stated
17/10 coefficient. This does not mean the true Z increased; the proved lower
estimate improved. At r=1/20 this gives 1747/400000=0.0043675. The already-recorded
sharper fixed-radius floor 0.0077592917375327855 remains preferable for the existing
RN consumer. Nothing here substitutes the new band estimate into that RN budget.

No all-angle, 3D, full-annulus, actual elder-selection, q0, prize or independent
acceptance claim is made. The original upper bound is not re-proved by this work.

## 1. Generalize the source's two free event parameters

Write R=1/20, choose epsilon in (0,1), let a=1-epsilon, and choose d>0. The new
selection is epsilon=103/500 and d=9/100. These are event cutoffs, not changes to
the field, pins, conditioning measure or radius domain.

Set m_(2j)=(-1)^j k^(2j)(0), sigma_e^2=m4-m2^2 and beta=m4*m2/4. Retain the source's
cancellation-free pin vector

\[
V=(f_M,f_{x,M},f_{y,M},(f_{x,S}-f_{x,M})/r,
(f_{y,S}-f_{y,M})/r,
(f_S-f_M)/r^3-(f_{x,S}+f_{x,M})/(2r^2)).
\]

Its conditioned value is v=(6/5,0,0,0,0,-1/6). The limiting vector is
(f,fx,fy,fxx,fxy,-fxxx/12). Its only nonsingleton covariance blocks are
[[1,-m2],[-m2,m4]] and [[m2,m4/12],[m4/12,m6/144]]. The other variances are m2,m2^2.
Exact positive shifted pivots establish G(0)>(3/80)I. Mean-square integral
identities give

\[
\|c\cdot(V(r)-V(0))\|_2^2\le r^2\|c\|_2^2
\left(m_2/4+m_4/4+m_2^2/4+m_6/16+m_4m_2/16+m_8/4096\right).
\]

For example the first three coordinate differences use endpoint displacement
r/2; the next two use average derivatives with mean |t|=1/4; the last uses
kernel (1/2)(1/4-t^2), whose absolute first moment is 1/64. Squared L2 bounds
sum by Cauchy-Schwarz. The displayed coefficient is below (31/20)^2.
Since (193/1000)^2<3/80, reverse triangle and Loewner inversion imply

\[
Q(r)=v^TG(r)^{-1}v\le Q_*=\frac{17/6}{[1-(31/20)R/(193/1000)]^2}
=\frac{1266466}{160083}<8.
\]

Here Q(0)=36m4/[25(m4-m2^2)]+4m2/(m2*m6-m4^2)<17/6, also recomputed. Thus every
positive radius in the band has a nonsingular regression law. No finite-radius
matrix grid is being used as a substitute for this transport argument.

## 2. Sector decomposition and axial event

As in the source, the complete line processes X=f(x,0), Y=fy(x,0), and
E=fyy(x,0)+m2*f(x,0) are mutually independent: separability and parity make all
cross-covariances zero. The pins condition X and Y separately and leave E
unchanged, with covariance sigma_e^2*k(x-z).

Put A_M=fxx(M)/r and A_S=fxx(S)/r. Integration by parts after subtracting the pins
gives residuals R_M=A_M+1 and R_S=A_S-1 with Peano kernels
(1/2-t)^2(1/2+t) and (1/2+t)^2(1/2-t), respectively, integrated over [-1/2,1/2]
against r*X''''(rt). Both are nonnegative with mass 1/12. Hence each unconditional
residual variance is at most s(r)^2=r^2*m8/144. Regression bounds the conditional
mean magnitude by sqrt(Q_*)s(r), and the conditional variance by s(r)^2.

For the event A={A_M<=-a,A_S>=a}, a union bound therefore gives

\[
P(A)\ge p_A=1-2\overline\Phi\left(\frac{\epsilon}{R\sqrt{m_8}/12}-\sqrt{Q_*}\right).
\]

The standardized argument is positive. A smaller r improves the bound; a
zero conditional variance is covered by the deterministic mean bound. The two
residuals are not assumed independent. Unlike the old fixed 14/15 floor, this
uses the full certified probability bound and the sharper Q_* rather than 8.

The centered mixed derivative B_M=fxy(M)/r belongs to the independent Y sector.
The pin-subtracted second-derivative kernel has mass 1/2, and conditioning on
zero Y values cannot increase variance, so E B_M^2<=beta. No B_S bound is needed.

## 3. Transverse event and moments

Let q_M=fyy(M), q_S=fyy(S), Qbar=(q_M+q_S)/2 and Delta=q_S-q_M. Then

\[
E Qbar=-m_2(6/5-r^3/12),\quad Var(Qbar)=\sigma_e^2(1+k(r))/2,
\]
\[
E Delta=m_2r^3/6,\quad Var(Delta)=2\sigma_e^2(1-k(r)).
\]

Stationarity gives zero covariance between midpoint and difference, hence their
Gaussian independence. Both are independent of X and Y. The mean-square
fundamental theorem gives 2(1-k(r))<=m2*r^2. Consequently

\[
Var(Qbar)\in[\sigma_e^2(1-m_2R^2/4),\sigma_e^2],
\quad sd(Delta)\le R\sqrt{\sigma_e^2m_2}.
\]

For D={|Delta|<=2d}, retain both tails and the positive nonzero mean:

\[
P(D)\ge p_D=1-2\overline\Phi\left(
\frac{2d-m_2R^3/6}{R\sqrt{\sigma_e^2m_2}}\right).
\]

Let T=(-Qbar-d)_+. Its underlying normal mean mu=m2*(6/5-r^3/12)-d and variance
sigma^2=Var(Qbar) lie in explicit whole-band intervals. With t=mu/sigma,

\[
M_1=E T=\sigma\phi(t)+\mu\Phi(t),\qquad
M_2=E T^2=(\sigma^2+\mu^2)\Phi(t)+\mu\sigma\phi(t).
\]

The program interval-evaluates both quantities over their entire parameter
ranges, including overestimation of correlated intervals. The approximation
used in parameter search is not used to certify these moments.

## 4. Typed determinant bound

Write H_i=[[rA_i,rB_i],[rB_i,q_i]], det(H_i)=rD_i with
D_i=A_i*q_i-rB_i^2. On A and D, whenever T>0, both transverse entries are at most
-T. Thus D_M>=aT-rB_M^2 and D_S<=-aT. A positive first expression ensures M is a
maximum and S a saddle. Pointwise, the rescaled typed integrand is at least
1_A*1_D*aT*(aT-rB_M^2)_+. The zero positive-part cases cause no difficulty.
Using x_+>=x and the established independence gives

\[
Z(r)/r^2\ge P(A)P(D)[a^2M_2-ar\beta M_1].
\]

First verify the bracket is positive; only then replace probabilities by lower
bounds and r by R in the negative term. The sufficient lower-bound expression
is enclosed by exact rational arithmetic, with outward decimal display

\[
[1.747106129274,\ 1.747994749902].
\]

**This interval encloses the sufficient-bound formula, not the true value of
Z(r)/r^2.** Its lower endpoint exceeds 1747/1000. All r dependence was bounded
uniformly before this scalar evaluation, proving the displayed candidate for
every 0<r<=R, not only tested radii.

## 5. Separate rational implementation and explicit tails

The scalar arithmetic is deliberately separate from the original repository
implementation. It uses Fraction endpoints, outward 2^-256 rounding, integer
square-root brackets, Machin's pi identity and alternating series with a proved
first-omitted-term remainder. CDF arguments outside [-8,8] and exponential
arguments outside [0,50] are refused, not extrapolated.

A positive Taylor partial sum proves e^12>100000. Hence e^-288<10^-120,
e^-1152<10^-480 and e^-1440<10^-600. For n<=8, writing C_n as the sum of absolute
Hermite coefficients, the |j|>=2 image tail is bounded by

\[
2C_n48^n10^{-480}/[1-(3/2)^n10^{-600}].
\]

The j=+/-1 contribution is bounded by 2|H_n(24)|10^-120. Keeping normalization,
the resulting moments differ from (1,3,15,105) by less than 10^-100. These are
rationally checked bounds, not deletion of small image terms. Rounding error is
then enclosed independently. Source hashes select the declared source, not an
extraction transform chosen to fit a desired hash.

The tests include both-tail omissions, understated moments, scope expansion,
wrong-quantity refutations, cache corruption and actual candidate recomputation.
The optional mpmath comparison is diagnostic only. Arb was not executed.
Nonauthor mathematical review, a second rigorous implementation, and any wider
research application remain separate tasks.
38 changes: 38 additions & 0 deletions engine/solver_pilots/h3_20260921/PROPOSAL_ASSESSMENT.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
# Six suggested mechanisms: disposition after source inspection

Owner authorization supplies permission to act, not a mathematical result. This
experiment advances a concrete H3 estimate instead of installing all six ideas
as workspace-wide rules. The suggestions are useful engineering patterns, not
established novel mathematical primitives.

| Suggestion | Decision and actual scope |
|---|---|
| Bitemporal/Merkle claims | Defer global migration. Valid time means the applicability interval of an assertion, review or operational status; it is not when a timeless theorem became true. Recursive hashes are implemented only for this experiment's declared formula DAG. They prove integrity under a trusted root, not logical soundness. Cyclic logical dependencies need an explicit treatment before any forest conversion. |
| Prover/falsifier lanes | Implement bounded parameter proposal -> rigorous scalar certification and adversarial negative controls. Actual result improves the source's band lower coefficient. No failure-count-to-proof or 10k-sample budget rule. The full-law Z counterexample evaluator remains unimplemented. |
| Execution cache | Implement local content-addressed scalar reuse with source/code/context/runtime keys, immutable outputs, hash checks and optional mandatory cold replay. It is not a hermetic distributed build and never becomes source of scientific authority. Shared cache authentication and Nix/Guix pinning are deferred. |
| Dual arithmetic | Optional mpmath cross-checks are diagnostic, not certified. Arb was inspected in primary documentation but not executed in this runtime. Agreement/overlap does not validate shared formulas, conditioning, tails or scope. A disagreement identifies a consistency problem, not automatically which backend is wrong. |
| Formula-cell lineage | Trace the actual mathematical formula dependencies here, with no invented Sheets ancestry. Full formula-cell lineage needs a source census, parser, named-range/INDIRECT/volatile/external/cycle handling and changed-cell impact tests. The supplied counts were not verified and are not adopted. |
| Proof market/energy | Record actual wall time, process CPU time and achieved bound improvement. Energy is unmeasured. No hypothetical 12h/200h cost estimates, spend, token market or automatic resource grants. Prioritization must consider all unresolved AND premises and alternative OR routes, not naive descendant count. |

## Highest-value next scientific actions

A nonauthor reviewer can check the general epsilon,d adaptation and the alternate
rational kernel using the exact source and new certificate. A genuinely second
rigorous backend can then reproduce the scalar inequalities; an independent
expectation oracle would be required for the proposed full-Z falsifier. Broader
all-angle or RN work must retain its own hypotheses and gates. No automatic
replacement of the existing stronger fixed-radius consumer is justified.

## Primary sources inspected

- MariaDB Bitemporal Tables: system-time and application-time versioning are distinct.
https://mariadb.com/docs/server/reference/sql-structure/temporal-tables/bitemporal-tables
- RFC 9162, section 2.1: Merkle inclusion and consistency concern authenticated data.
https://www.rfc-editor.org/rfc/rfc9162.html
- Bazel Remote Caching: action metadata and content-addressed output storage;
undeclared environment/input changes matter. https://bazel.build/remote/caching
- python-flint 0.9.0 arb documentation: ball representations and error tracking.
https://python-flint.readthedocs.io/en/latest/arb.html

These sources support the engineering distinctions, not the H3 theorem. The
mathematical argument and exact scalar computations are explicitly supplied here.
Loading
Loading