ELUMIORU-LHORIEN / research /appendix.tex
PureOne's picture
Add ELUMIORU-LHORIEN research dataset release
80cb3ae verified
Raw History Blame Contribute Delete
13.3 kB
\section{A fully worked coupled obstruction}
\label{app:worked}
Take the prime $p=3$, precision $k=1$, two vertices $u,v$ and one edge $u\to v$.
Set
\[
F_u(X)=X,\qquad F_v(X)=X-3,\qquad
z_u=z_v=0\pmod3,\qquad d_{uv}=0\pmod9.
\]
Both local equations and the edge equation are valid at the current precision.
The local roots are simple. Nevertheless, the desired next-precision system
is inconsistent.
The local and edge residuals are
\[
a=\begin{pmatrix}0\\2\end{pmatrix},\quad
G=\begin{pmatrix}1&0\\0&1\end{pmatrix},\quad
\eta=0,\quad
D=
\begin{pmatrix}
1&0\\
0&1\\
-1&1
\end{pmatrix}.
\]
The equation $Db=-(a,\eta)$ first requires $b_u=0,b_v=1$, while its final
row requires $b_v-b_u=0$. Hence the cokernel class is nonzero.
The graph itself is a tree, with ordinary $H^1=0$. This small example
isolates why graph consistency alone is an incomplete test.
The forced local correction is $b^*=(0,1)$ and the edge defect is
$\chi=\eta+\delta_Kb^*=1$. If the next transport digit is genuinely free,
the unique permitted repair is
\[
d'_{uv}=d_{uv}+3\chi=3\pmod9.
\]
The lifted roots are $(0,3)$, which satisfy the repaired edge.
If the original transport zero modulo nine is an immutable measurement,
the output is instead \emph{infeasible}. The same numerical calculation
does not license changing that measurement.
For the real decoder choose $\rho=1/2$, so $c_p=1/4$, and action $A=1$.
The target coefficient functions at precision two are
\[
W_u(x)=0,\qquad W_v(x)=\frac14x.
\]
Keeping the old edge forces equal correction digits. A candidate with both
digits zero errs at $v$; a candidate with both digits one errs at $u$.
At $x=1/2$, at least one prediction has magnitude error
\[
\frac18e^{-1/\varepsilon}.
\]
At $\varepsilon=1$ this is approximately $0.045985$.
Every finite Taylor jet of that error at $\varepsilon=0$ is zero.
The scale is analytically invisible to a finite local power expansion,
although the obstruction and its positive-scale consequence are exact.
The associated $\delta_3$ carry check is also explicit:
\[
\delta_3(3)-\delta_3(0)
=\frac{3-27}{3}=-8\equiv1\pmod3.
\]
It records the appended digit. It does not establish that the digit solves
$F_v$; the polynomial check remains a separate condition.
\section{Actual intersection maps}
\label{app:zigzag}
For the first tested intersection span, the left components are
$\{0,1,2\},\{3\}$, the intersection components are
$\{0,1\},\{2\},\{3\}$, and the right components are
$\{0,1\},\{2,3,4\}$. In those bases, the measured inclusion matrices are
\[
L_0=
\begin{pmatrix}1&1&0\\0&0&1\end{pmatrix},
\qquad
R_0=
\begin{pmatrix}1&0&0\\0&1&1\end{pmatrix}.
\]
For the next span the intersection components are
$\{1\},\{2,3\},\{4\}$, with right components
$\{1,2,3\},\{4,5\}$. The matrices are
\[
L_1=
\begin{pmatrix}1&0&0\\0&1&1\end{pmatrix},
\qquad
R_1=
\begin{pmatrix}1&1&0\\0&0&1\end{pmatrix}.
\]
Every matrix has rank two over each tested prime field. A column has one
entry equal to one because a connected component of the intersection lies
in one connected component of the containing graph.
These matrices preserve more information than the dimension sequence alone:
they specify which component classes are merged under each inclusion.
They still omit the affine transports and anchors that determine a future
root branch. The implementation therefore keeps topological maps and
semantic branch data as different objects.
\section{Reference implementation and data layout}
\label{app:implementation}
The complete executable source is supplied with the manuscript. The modules
have narrow interfaces to make the mathematical assumptions inspectable.
\begin{longtable}{>{\raggedright\arraybackslash}p{0.32\textwidth}>{\raggedright\arraybackslash}p{0.60\textwidth}}
\toprule
File & Responsibility\\
\midrule
\endhead
\texttt{core.py} &
Exact polynomial arithmetic, operation meter, Hensel strategies, digit decoder,
rational pole reconstruction, affine graph memory and actual $H_0$ maps.\\
\texttt{synthesis.py} &
Typed expression enumeration, trace pruning, Laurent-polynomial normalization,
exact certificate and execution of the accepted expression.\\
\texttt{experiment.py} &
Public task generator, separated hidden target, learner AST gate, baselines,
inventor updates, frozen confirmation protocol and raw lineage logs.\\
\texttt{analytic\_experiments.py} &
Stokes quadrature, rational-pole trials, $\delta$ identities and exact factorial
information controls.\\
\texttt{obstruction\_experiments.py} &
Complete lifting obstruction, admissible repair, independent nonlinear
enumeration, finite-prefix error witnesses and nonnested graph maps.\\
\texttt{test\_core.py} and \texttt{test\_system.py} &
Algebraic failure cases, independent finite roots, grammar execution,
compiler equivalence, and test/train state separation.\\
\texttt{summarize\_results.py} &
Generate tables, seed-level summaries and the three scientific figures
from existing measurement files.\\
\texttt{reproduce.py} &
Run the complete experiment without requiring external packages.\\
\texttt{results/} &
Configuration, source hashes, raw rows, complete lineages, exact verification
outputs, CSV table, and derived summaries.\\
\bottomrule
\end{longtable}
The grammar-level types can be summarized as
\[
\begin{aligned}
\texttt{MomentPair}&\longrightarrow \texttt{PositiveAction},\\
(\texttt{Graph},\texttt{Anchors},\texttt{Query})
&\longrightarrow \texttt{Branch}\ \text{or ambiguity},\\
(\texttt{Polynomial},\texttt{Branch},k)
&\longrightarrow \texttt{RootWord}\ \text{or obstruction},\\
(\texttt{PositiveAction},\texttt{RootWord})
&\longrightarrow \texttt{LearnerProgram}.
\end{aligned}
\]
No function silently converts an $\F_p$ obstruction class to a real number.
Real outputs are produced only through the declared digit decoder.
A finite expression is certified in a ring where $g$ is a unit.
The verifier permits inverses only for expressions known to be unit
monomials in that fragment. It does not assume that an arbitrary expression
such as $e$ is invertible. Canonical polynomial normalization makes
$e+gh(e,g)$ an exact zero test, independent of the finite training traces.
The executable evaluator then reduces the expression at the requested
modulus and checks the resulting polynomial root.
The top-level protocol is:
\begin{lstlisting}
initialize learner grammar, graph relations, one anchor, inventor
for measured generation = 1,...,12:
evaluate a copy on the fixed held-out task stream
stop the probe when the operation budget is exhausted
record solved tasks, failed work, program version, and costs
if this was generation 12: stop
solve eight quadratic training tasks
retain validated anchors and residual traces
propose the next eligible inventor transformation
certify its algebraic or memory invariant
measure the four calibration conditions
adopt only if unweighted capability and coupled score improve
record both the old and new executable procedure
\end{lstlisting}
The full system's canonical procedure sequence is:
\begin{lstlisting}
I0 = digit enumeration + graph reconstruction
I1 = synthesized residual expression + graph reconstruction
I2 = synthesized residual expression + precision doubling
+ graph reconstruction
I3 = synthesized residual expression + precision doubling
+ certified component-potential memory queries
\end{lstlisting}
These are four actual program versions. They are not twelve distinct
invention mechanisms, and their semantics remain inside the initial
metalanguage.
\subsection{How to inspect a lineage}
Each learner-discovery event records the task family, exact polynomial,
rational moments, acceptance result, root and action when accepted,
instrumented work, invention-program signature, local root trace and
resulting anchor state. Each inventor-update event records its before/after
signature, synthesis or schema certificate, four calibration gains and costs,
raw and normalized interaction, objective values and meta-work breakdown.
Failed expression candidates are included in the synthesis cost.
A failed training attempt remains in the lineage with its cost.
A budget-interrupted probe remains an attempt with its consumed work.
The final generation's recorded cumulative training and meta-costs end
at the last intervention that could affect that measurement.
When comparing lineages, distinguish three counts. An accepted learner
instance has a new target parameter pair. A structural learner shape is
the arrangement of its executable operators. An inventor revision changes
the procedure used to produce future instances. The primary benchmark has
many accepted instances, one outer learner shape, and at most three
accepted inventor revisions per full-system seed.
\section{Coverage of the proposed mathematical domains}
The selected mathematical objects are those for which an exact executable
interface was constructed. Adding a field name without an operational role
would not strengthen the result.
\begin{longtable}{>{\raggedright\arraybackslash}p{0.28\textwidth}>{\raggedright\arraybackslash}p{0.64\textwidth}}
\toprule
Domain element & Status in this study\\
\midrule
\endhead
Borel--Laplace summation, a Stokes automorphism and an elementary bridge
relation & Defined and proved for rational Borel germs; numerical contour
and exact pole-recovery checks executed.\\
General alien calculus, moulds, ARI/GARI, acceleration and
\'Ecalle--Voronin invariants & No claim of implementing these general
theories. The controlled elementary Stokes generator has a specified domain.\\
Bounded prism and $\delta$-ring & A genuine $(\Z_p,(p))$ foundation is proved;
carry and precision-loss identities are executed over finite representatives.\\
Prismatic cohomology, derived comparisons, perfectoid tilting,
Nygaard filtrations and prismatic $F$-crystals & Not computed.
Finite Hensel reconstruction is explicitly distinguished from them.\\
Affine graph sections, $H^0$, $H^1$, and actual intersection $H_0$ maps &
Defined and executed. Labels and anchors carry the predictive information.\\
Dynamic lineage and higher-order coupled repairs & Executed as separate
exact finite experiments; not active causes of the twelve-generation gains.\\
Ran spaces, factorization homology, intersection homology and general
exit-path categories & Not needed for the proved restricted construction
and not asserted as implemented contributions.\\
Five independent indices & A family of typed certificate fibers, reductions,
grammar inclusions and graph spans is defined. No smooth manifold or new
cohomology theory is inferred from the number of axes.\\
Changes to the invention procedure & One synthesized expression and two
measured selections of supplied compiler schemas.\\
Changes to analytic model discovery and unrestricted outer invention &
No autonomous expansion of the rational model class or of the complete
meta-selector is demonstrated.\\
General intelligence growth & Not measured. The reported capability is
exact structured program synthesis under specified observations.\\
\bottomrule
\end{longtable}
\section{Seed-level primary results}
This table permits direct checking of the means and paired comparisons.
All values are solved tasks under the same deployment cap, except the final
column, which is the full system's cumulative meta-work before its last
measurement.
\input{seed_table.tex}
\section{Compact claim ledger}
The following statements are the strongest conclusions supported by the
delivered mathematics and experiments.
\begin{enumerate}
\item A rational resurgent diagnostic can identify a hidden exponential
scale under an explicit analytic model, while leaving its arbitrary boundary
amplitude undetermined.
\item Anchored affine memory can supply branch information missing from that
diagnostic; unlabelled topological persistence cannot guarantee preservation
of the branch meaning.
\item The combined derivative/transport operator gives a complete
first-order obstruction. Ordinary graph $H^1$ can miss it, even on a tree.
\item When next transport digits are permitted choices, the obstruction
specifies one exact repair. When they are fixed facts, it specifies failure.
\item A finite-prefix digit decoder turns the failed compatibility into a
real nonzero exponential-sector error with all finite local power jets zero.
\item A small symbolic system can synthesize and execute a certified
replacement for its future digit enumeration, then adopt additional
proved compilation schemas after measured comparison.
\item Those changes increase held-out discovery rate relative to the system's
own initial inventor in this task class, but the resulting rate plateaus
and stronger baselines remain faster.
\item The exact one-tenth objective weight has a measured null effect here.
A separate, explicitly constructed reciprocal-information task has positive
interaction; the two findings are compatible.
\end{enumerate}
No statement in this ledger claims a universal learning theorem, historical
first use of the component mathematics, or demonstrated superintelligence.
The package supplies the actual theorem hypotheses, counterexamples,
code and measurements needed to assess the more limited positive result.