\documentclass[10pt,letterpaper,twocolumn]{article}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[
  letterpaper,left=.72in,right=.72in,top=.75in,bottom=.73in,
  headheight=14pt,headsep=14pt,footskip=26pt
]{geometry}
\usepackage{microtype}
\usepackage{mathtools,amssymb,amsthm}
\usepackage{xcolor}
\usepackage{tikz-cd}
\usepackage{titlesec}
\usepackage{fancyhdr}
\usepackage{caption}
\usepackage{booktabs,balance}
\usepackage{xurl}
\usepackage{hyperref}
\usepackage{bookmark}

\definecolor{paperink}{HTML}{202824}
\definecolor{paperaccent}{HTML}{145A51}
\definecolor{papermuted}{HTML}{51655F}
\definecolor{paperrule}{HTML}{B6CBC4}
\hypersetup{
  pdftitle={Evolutionary Selection for Agent Populations under Comparable Evaluation},
  pdfauthor={Anonymous},
  pdfsubject={Exact aggregation and robustness to score mismatch},
  colorlinks=true,linkcolor=paperaccent,citecolor=paperaccent,urlcolor=paperaccent
}
\AtBeginDocument{\color{paperink}}
\setlength{\columnsep}{.29in}
\setlength{\parindent}{1em}
\setlength{\parskip}{1pt plus .5pt}
\setlength{\emergencystretch}{1.7em}
\linespread{1.035}
\allowdisplaybreaks[1]
\raggedbottom
\clubpenalty=1000
\widowpenalty=1000

\titleformat{\section}{\large\bfseries\raggedright}
  {\color{paperaccent}\thesection}{.55em}{}
\titlespacing*{\section}{0pt}{14pt plus 3pt minus 2pt}{5pt}
\titleformat{\subsection}{\normalsize\bfseries\raggedright}
  {\color{paperaccent}\thesubsection}{.55em}{}
\titlespacing*{\subsection}{0pt}{10pt plus 2pt minus 1pt}{4pt}
\captionsetup{font=small,labelfont={bf,color=paperaccent}}

\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\footnotesize\sffamily\color{papermuted}Evolutionary selection for agent populations}
\fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}
\renewcommand{\headrulewidth}{0pt}
\fancypagestyle{plain}{%
  \fancyhf{}%
  \fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}%
  \renewcommand{\headrulewidth}{0pt}%
}

\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\DeclareMathOperator{\Map}{Map}
\DeclareMathOperator{\Tot}{Tot}
\DeclareMathOperator{\Fun}{Fun}
\DeclareMathOperator{\Hom}{Hom}
\DeclareMathOperator{\grad}{grad}
\DeclareMathOperator{\Ext}{Ext}
\DeclareMathOperator{\Res}{Res}
\DeclareMathOperator{\Kl}{Kl}
\DeclareMathOperator{\id}{id}
\DeclareMathOperator{\rank}{rank}
\newcommand{\E}{\mathcal E}
\newcommand{\Spc}{\mathcal S}
\newcommand{\R}{\mathbb R}
\newcommand{\C}{\mathbb C}
\newcommand{\Z}{\mathbb Z}
\newcommand{\overB}{\E_{/B}}
\tikzcdset{every label/.append style={font=\small}}
\newcommand{\AgenticTitle}{Evolutionary Selection for Agent Populations under Comparable Evaluation}
\newcommand{\AgenticAbstract}{An agent population can allocate work and select proposals by comparing their measured performance. The validity of a comparison depends on the evaluation context, while allocation and selection normalize the resulting weights. We model the inner execution by a scored probability kernel and the outer selection by exponential reweighting followed by variation. We characterize exact aggregation when the scores and variation descend to a common quotient. A projective total-variation bound controls one-step errors from stale or mismatched scores. Under a contracting variation kernel we propagate that bound over repeated updates. These results give an interface for auditing the outer update separately from the evaluation assumptions that make it meaningful.}

\begin{document}
\twocolumn[
\begin{@twocolumnfalse}
\noindent\textcolor{paperaccent}{\rule{\textwidth}{1pt}}\par
\vspace{11pt}
{\noindent\fontsize{21}{27}\selectfont\bfseries\raggedright
Evolutionary Selection for Agent Populations\\under Comparable Evaluation\par}
\vspace{10pt}
{\noindent\small Anonymous \quad\textbullet\quad November 2026\par}
\vspace{9pt}
\noindent\begin{minipage}{\textwidth}\small
\textbf{Abstract.} \AgenticAbstract
\par\medskip
\noindent\textbf{Keywords.}\enspace
Bilevel selection, agent populations, exponential reweighting,
quotients, evaluation delay, Dobrushin contraction.
\end{minipage}
\vspace{16pt}
\end{@twocolumnfalse}
]
\thispagestyle{plain}
\section{Execution and population selection}
\label{amj:execution}

An agent system can learn at two levels. At the inner level, a proposed
strategy carries out a task and returns a trace. At the outer level,
a population controller assigns evaluation or deployment mass to
available strategies. A useful outer algorithm should specify which
inner measurements it can compare, how it constructs the next
population, and how a change in those measurements affects that
population. This paper addresses these three questions for
evolutionary selection.

The model takes the initial population and the variation kernel as
inputs. It applies to routing among fixed specialists, to allocating
evaluation budgets to immutable proposals, and to selecting newly
admitted variants. In each application the same selection rule can
operate over a finite registry or a standard Borel proposal space.
Finite registries admit an algebraic realization by sums. The
measure-theoretic statements below describe the update independently
of a specific registry.

\subsection{An inner evaluation contract}

Let $S$ be a standard Borel space of strategy versions. A strategy
version fixes the admissible behavior to be evaluated, including
any state snapshot that affects its execution. A change of the
behavior-defining data creates a different version. An evaluation
context $c$ specifies a task space $D_c$, a task law $\lambda_c$,
an execution environment and budget, an outcome space $Z_c$, and
a bounded measurable scoring function $r_c:D_c\times Z_c\to\mathbb R$.
The outcome space includes statuses that the scoring policy uses
to treat failure or noncompletion. The inner semantics is a
probability kernel
\[
  K_c\colon S\times D_c\longrightarrow\mathcal P(Z_c).
\]
Here and below, probability kernels include both deterministic maps
and randomized evaluation. The expected score is
\begin{equation}
 \label{amj:contract}
 F_c(s)=\int_{D_c}\int_{Z_c}
     r_c(d,z)\,K_c(s,d,dz)\,\lambda_c(dd).
\end{equation}
Measurability of $F_c$ follows from the defining measurability of the
kernel and of the bounded integrand.

A context identifier in an audit record stands for this collection
of choices. Distinct realizations may represent the same contract
when they induce the same relevant scored law. Equality of a score
convention by itself need not imply equality of task laws or
execution budgets. A stated comparison across contexts therefore
uses transport carrying the score or a quantitative bound on
the relevant discrepancy. A missing trace and a trace scored as
failure may give different score estimands if they are pooled
without an explicit inclusion rule.

Endogenous evaluation is also possible. Let $a(s,y)$ be a bounded
measurable interaction payoff for evaluating $s$ against $y$ under
a fixed contract, and let $m$ be a probability measure on $S$.
Set
\begin{equation}
 \label{amj:endogenous}
 F_m(s)=\int_S a(s,y)\,m(dy).
\end{equation}
The opponent law or population at which the score was measured
is now part of the comparison. In an empirical implementation,
the exact integral and an estimate made from a sample are separate
inputs to the error analysis. All statements about scores are
relative to the declared bounded score function.

\subsection{An outer update}

For a probability measure $m$ on $S$ and a measurable weight
$w:S\to(0,\infty)$ with $0<\int_S w\,dm<\infty$, define the
probability measure
\begin{equation}
 \label{amj:tilt}
 \mathcal S_w(m)(B)=
 \frac{\int_B w(s)\,m(ds)}{\int_S w(s)\,m(ds)} .
\end{equation}
In the score-based algorithm the weight is $w=\exp(\eta F)$, with
$\eta\ge0$ fixed by the controller in each round. The score is
held fixed for that transition, including when it was obtained
from \eqref{amj:endogenous}. A variation kernel
$V:S\longrightarrow\mathcal P(S)$ follows selection. The
combined transition $m\mapsto\mathcal S_w(m)V$ sends a selected
strategy to a distribution over strategies for the next round.
The identity kernel leaves the selected population unchanged.
The semantics allows a controller to put exploration or
externally admitted replacements in $V$.

Multiplication by a positive constant in $w$ leaves
\eqref{amj:tilt} unchanged. Thus adding a common constant to
every score of a round leaves its transition unchanged.
A positive change of units $F\mapsto aF+b$ gives the same
transition when the controller replaces $\eta$ by $\eta/a$.
The support of selection in the measure-theoretic sense is
also preserved. A measurable set has selected mass zero
exactly when its original mass is zero, because $w>0$.
Variation determines any movement to previously unoccupied
regions. These identities help distinguish score conventions
that can be changed by a reparametrization from differences
that change allocations.

For bounded $F$ write $\bar F_m=\int_SF\,dm$.
For any bounded measurable test function $\varphi$,
dominated differentiation gives
\begin{equation}
 \label{amj:tangent}
 \left.\frac{d}{d\eta}\right|_{\eta=0}
 \int_S \varphi\,d\mathcal S_{e^{\eta F}}(m)
       =\int_S\varphi(F-\bar F_m)\,dm .
\end{equation}
Indeed, both integrals in the ratio are differentiable
at zero. Their derivatives insert a factor $F$, and
the denominator at zero is one. With $F=F_m$ this is the
weak form of the replicator vector field. For finite
strategy types, the right side tested against a coordinate
indicator is $m_i(F_m(i)-\bar F_m)$. This links the
outer controller to the continuous evolutionary dynamics
treated by Alexander \cite{amj-alexander}. The finite-step
rule belongs to multiplicative weights
\cite{amj-arora,amj-sorin}. When $F$ is fixed across all
rounds and $V$ is the identity kernel, the derivative of
the expected score along exponential reweighting is
$\operatorname{Var}_{\mathcal S_{e^{\eta F}}(m)}(F)$, by the same
differentiation at each $\eta$. Endogenous $F_m$ may vary
with the population, so that calculation is a statement
about a frozen evaluation.

\section{When a reduced population model is exact}
\label{amj:quotient}

A controller often compares populations across different
registries or versions of an agent catalogue. Let $\pi:S\to Q$
be a measurable map to a second standard Borel space, and write
$\pi_*m$ for pushforward. A reduced controller retains $\pi_*m$
rather than the original $m$. We first characterize exactly what
happens to one weighting and then give a sufficient condition
for repeated selection and variation.

\begin{theorem}[Pushforward of a selection]
\label{amj:pushforward}
With $m$ and $w$ as in \eqref{amj:tilt}, let
$\bar m=\pi_*m$. Let $\bar w$ be a version of
$d\pi_*(wm)/d\bar m$, where $wm$ is the finite measure with
density $w$ with respect to $m$. Then
\begin{equation}
 \label{amj:descent}
   \pi_*\mathcal S_w(m)=\mathcal S_{\bar w}(\bar m).
\end{equation}
If $w=v\circ\pi$ for a positive measurable $v$, the same
weight $v$ can be used for every $m$.
\end{theorem}

\begin{proof}
For a $\bar m$-null set $C$, its inverse image is $m$-null
and therefore $wm$-null. The Radon--Nikodym derivative in the
statement exists. It is positive $\bar m$-almost everywhere
since $w>0$, and it can be given any positive value on the
remaining null set. For $C\subseteq Q$ measurable, the
numerator of the left side in \eqref{amj:descent} is
$\pi_*(wm)(C)=\int_C \bar w\,d\bar m$. Both denominators equal
$\int_S w\,dm$. If $w$ factors through $\pi$, the latter
derivative is $v$ by the change-of-variables identity for
pushforward.
\end{proof}

The reduced weight exists even when the original weight is
inconsistent along fibres. In that situation $\bar w$ generally
depends on the distribution within each fibre, information not
supplied by $\bar m$. In particular, for exponential reweighting,
the reduced weight is a conditional average of exponentials
rather than the exponential of a conditional average of scores.
The order of averaging and exponentiation matters for the
resulting population.

There is a converse for a fixed $w$ when the image of $\pi$
contains at least two points. Suppose the same reduced
reweighting gives exact descent for every population. If
$\pi(s)=\pi(s')$, choose $y$ with $\pi(y)\ne\pi(s)$.
The laws $(\delta_s+\delta_y)/2$ and $(\delta_{s'}+\delta_y)/2$
push forward to the same law. Their selected masses at
$\pi(s)$ are $w(s)/(w(s)+w(y))$ and $w(s')/(w(s')+w(y))$.
As $w(y)>0$, equality forces $w(s)=w(s')$. This proves
pointwise constancy along fibres. In the degenerate case of
a one-point image the reduced population always has unit
mass at that point and the requirement imposes no constraint
on $w$. A measurable reduced weight is part of the forward
factorization hypothesis.

\begin{proposition}[Exact reduced outer dynamics]
\label{amj:reduced}
Suppose $V$ and a kernel $\bar V:Q\to\mathcal P(Q)$ satisfy
\begin{equation}
 \label{amj:lumpable}
 V(s,\pi^{-1}C)=\bar V(\pi(s),C)
 \quad\text{for every measurable }C\subseteq Q.
\end{equation}
If $w=v\circ\pi$, then
\[
 \pi_*(\mathcal S_w(m)V)
  =\mathcal S_v(\pi_*m)\bar V.
\]
The result applies at every round for population-dependent
scores when $F_m=\bar F_{\pi_*m}\circ\pi$.
\end{proposition}

\begin{proof}
For any probability measure $n$, integration in
\eqref{amj:lumpable} gives
$\pi_*(nV)=(\pi_*n)\bar V$.
Apply the equality with $n=\mathcal S_w(m)$ and use
Theorem~\ref{amj:pushforward}. For population-dependent scores,
the reduced value is determined by $\pi_*m$, so induction proves
equality at all rounds with the corresponding transition.
\end{proof}

The condition in \eqref{amj:lumpable} is strong lumpability,
expressed using Markov kernels. It also follows from a square of
kernels in the usual compositional language of probability
\cite{amj-fritz}. If $T$ denotes the full population transition
and $\bar T$ the reduced transition, both vertical arrows in the
following square are pushforwards $\pi_*$.
\[
 \begin{tikzcd}[column sep=4.3em, row sep=2.6em]
   \mathcal P(S) \arrow[r,"T"] \arrow[d] &
   \mathcal P(S) \arrow[d] \\
   \mathcal P(Q) \arrow[r,"\bar T"] & \mathcal P(Q).
 \end{tikzcd}
\]
Reindexing the registry by a measurable
isomorphism is a special case, with the scores and variation
transported along the isomorphism. For strategy families,
requiring the scores to factor is stronger than equality
of a current leaderboard. A quotient that groups all
proposals with the same measured score can still contain
versions that generate different distributions of
descendant strategies. A reduced controller is exact
when both comparisons needed by
Proposition~\ref{amj:reduced} are in place.

The unnormalized measure also gives a rule for combining
evaluation shards. For a convex mixture $m=\sum_i\alpha_i m_i$
evaluated under a common weight function, write $Z_i=\int w\,dm_i$.
Linearity of integration gives
\[
 \mathcal S_w(m)=
 \frac{\sum_i\alpha_i Z_i\,\mathcal S_w(m_i)}
      {\sum_i\alpha_i Z_i}.
\]
This formula holds for countably many shards with finite positive
total normalizer. It identifies the sufficient local outputs for
exact recombination. An implementation that averages local
probability vectors with the original proportions $\alpha_i$
can change the global update, because it loses the unnormalized
masses. Comparability of the local weights is a separate assumption.

\section{Sensitivity to score and population mismatch}
\label{amj:sensitivity}

For probabilities $m,n$ write
$d_{\mathrm{TV}}(m,n)=\sup_B|m(B)-n(B)|$, where $B$ ranges
over measurable sets. Here $d_{\mathrm{TV}}$ is half the
variation norm of the signed measure $m-n$. If a bounded
measurable function has values in an interval of length $L$,
the absolute difference of its expectations under $m,n$ is at
most $L d_{\mathrm{TV}}(m,n)$. For $L>0$, subtract the lower
endpoint, rescale to $[0,1]$, and integrate the differences of
the probabilities of its superlevel sets.

\begin{theorem}[Projective sensitivity]
\label{amj:projective}
Let $w,\widehat w$ be positive $m$-integrable weights as in
\eqref{amj:tilt}. Assume
$\log(\widehat w/w)\in[\ell_-,\ell_+]$ $m$-almost everywhere,
and set $\omega=\ell_+-\ell_-$. Then
\begin{equation}
 \label{amj:projectivebound}
 d_{\mathrm{TV}}(\mathcal S_{\widehat w}(m),\mathcal S_w(m))
     \le \tanh(\omega/4).
\end{equation}
\end{theorem}

\begin{proof}
Set $p=\mathcal S_w(m)$, $r=\widehat w/w$, and $u=\int r\,dp$.
The other selected measure has density $r/u$ with respect to $p$,
so the distance is $\int|r-u|\,dp/(2u)$. Put
$a=\exp(\ell_-)$ and $b=\exp(\ell_+)$. If $a=b$, the distance
is zero by proportionality. For $a<b$, the secant bound for the
convex function $x\mapsto |x-u|$ on $[a,b]$ yields
\[
 \frac{\int|r-u|\,dp}{2u}
 \le \frac{(u-a)(b-u)}{u(b-a)}
 = \frac{a+b-u-ab/u}{b-a}.
\]
The inequality $u+ab/u\ge2\sqrt{ab}$ bounds this by
$(\sqrt b-\sqrt a)/(\sqrt b+\sqrt a)$, which equals
$\tanh(\omega/4)$.
\end{proof}

The essential oscillation of the log-density ratio is the
Hilbert projective distance, since normalizing adds a common
constant. Equation~\eqref{amj:projectivebound} is the specialization
to selection of the projective control of total variation studied
by Cohen and Fausti~\cite{amj-cohen-fausti}. For
$w=e^{\eta F}$ and $\widehat w=e^{\eta\widehat F}$,
a uniform error $|\widehat F-F-b|\le\epsilon$ after any common
shift $b$ gives
\begin{equation}
 \label{amj:estimator}
 d_{\mathrm{TV}}(\mathcal S_{e^{\eta\widehat F}}(m),
                  \mathcal S_{e^{\eta F}}(m))
   \le \tanh(\eta\epsilon/2).
\end{equation}
A bound on the loss for one strategy cannot replace the uniform
or oscillation bound on the relative scores of the strategies
allocated mass. If a uniform bound is probabilistic, the
displayed guarantee holds on the event on which the bound holds.

For the endogenous payoff in~\eqref{amj:endogenous}, suppose each
function $a(s,-)$ has range in an interval of length at most $L$.
The expectation bound gives
\begin{equation}
 \label{amj:payofflip}
  \sup_s |F_m(s)-F_n(s)|\le L d_{\mathrm{TV}}(m,n).
\end{equation}
A stale score measured at $n$ with uniform additional error
$\epsilon$ yields an error of at most
$\epsilon+L d_{\mathrm{TV}}(m,n)$ when applied at $m$.
Thus freshness is determined by a change in the opponent
distribution and its influence on the scores. The number of
elapsed rounds by itself gives no such bound.

A contract change can be handled in the same way when the tasks
and outcomes can be compared. For two contexts on common spaces
$D,Z$, use scoring maps taking values in $[0,1]$. Decompose the
change in~\eqref{amj:contract} by replacing the task law, execution
kernel, and scoring map one at a time. The first term is bounded
by $d_{\mathrm{TV}}(\lambda_c,\lambda_{c'})$. The second is
bounded by
$\sup_{s,d}d_{\mathrm{TV}}(K_c(s,d,\cdot),K_{c'}(s,d,\cdot))$,
by the same expectation inequality. The final term is bounded
by $\sup_{d,z}|r_c(d,z)-r_{c'}(d,z)|$. Their sum bounds
$\sup_s|F_c(s)-F_{c'}(s)|$ and is an input to
\eqref{amj:estimator}. A declared transport can first place
different task or outcome representations in common spaces.
The bound uses the transported quantities and therefore
depends on the specified comparison.

\section{Stability across repeated updates}
\label{amj:stability}

Selection also changes when the population changes with the
weights fixed. If $0<a\le w\le b$, then
\begin{equation}
 \label{amj:fixedweightlip}
 d_{\mathrm{TV}}(\mathcal S_w(m),\mathcal S_w(n))
       \le (b/a)d_{\mathrm{TV}}(m,n).
\end{equation}
For measurable $B$, put $c_B=\mathcal S_w(n)(B)$. The difference
of its selected probabilities is
\[
 \frac{\int w(\mathbf 1_B-c_B)\,d(m-n)}{\int w\,dm}.
\]
The integrand takes values in $[-b c_B,b(1-c_B)]$,
an interval of length $b$, and the denominator is at least $a$.
Taking the supremum over $B$ gives~\eqref{amj:fixedweightlip}.

Suppose $F_m$ has oscillation across $S$ at most $M$ uniformly
in $m$ and satisfies~\eqref{amj:payofflip}. Split a change of the
selection at $m$ to the selection at $n$ into a change of base
measure with weight fixed and a change of scores with base
measure fixed. Combining~\eqref{amj:fixedweightlip}
and~\eqref{amj:estimator}, and using $\tanh x\le x$ for $x\ge0$,
gives the following Lipschitz constant for the endogenous selector:
\begin{equation}
 \label{amj:lipschitz}
 C=e^{\eta M}+\eta L/2.
\end{equation}
The ratio $b/a$ for $w=e^{\eta F_m}$ can be taken to be at most
$e^{\eta M}$. This use of score oscillation makes the constant
unchanged by adding a common score offset. The stated $L$
is sufficient; a bound on score differences modulo common
offsets would likewise suffice in the second term.

For variation, write $\kappa$ for the Dobrushin coefficient of $V$,
the supremum of $d_{\mathrm{TV}}(V(s,\cdot),V(s',\cdot))$
over $s,s'$. Then
$d_{\mathrm{TV}}(mV,nV)\le\kappa d_{\mathrm{TV}}(m,n)$.
For each $B$, the function $s\mapsto V(s,B)$ has oscillation
at most $\kappa$, so the assertion follows by the expectation
bound. This is the usual coefficient for variation of
probability measures~\cite{amj-gaubert-qu}.

\begin{proposition}[Propagation of score errors]
\label{amj:propagation}
Let the reference transition at $m$ use $e^{\eta F_m}$ and let
the approximate transition at $\widetilde m_t$ use
$e^{\eta\widetilde F_t}$, each followed by $V$. Set
$d_t=d_{\mathrm{TV}}(\widetilde m_t,m_t)$ and assume that
$\widetilde F_t-F_{\widetilde m_t}$ is uniformly bounded by
$\epsilon_t$ in absolute value after some common shift.
For $C$ as in~\eqref{amj:lipschitz} and $\rho=\kappa C$,
\begin{equation}
 \label{amj:geometric}
 \begin{aligned}
  d_{t+1} &\le \rho d_t+\kappa\tanh(\eta\epsilon_t/2),\\
  d_t &\le \rho^t d_0\\
      &\quad+\kappa\sum_{j<t}\rho^{t-1-j}\tanh(\eta\epsilon_j/2).
 \end{aligned}
\end{equation}
In particular, if $\rho<1$ and $\epsilon_t\le\epsilon$ for
every $t$, the contribution of score errors is bounded by
$\kappa\tanh(\eta\epsilon/2)/(1-\rho)$.
\end{proposition}

\begin{proof}
Insert the reference selection at $\widetilde m_t$ between
the two selected populations. Bound the first distance
by~\eqref{amj:estimator} and the second by $C d_t$.
Applying $V$ contracts both by $\kappa$. Iterating the resulting
one-step inequality proves~\eqref{amj:geometric}. The uniform
bound uses the sum of the geometric series.
\end{proof}

The recurrence applies along each realization on which the
assumed error bounds hold. It can also use uniform bounds
over round-dependent scores and a shared round-dependent
variation kernel. The condition $\rho<1$ is sufficient for
uniform control over time. A controller with identity
variation on a nontrivial strategy space has $\kappa=1$
and cannot infer uniform control from this criterion.
Strong mixing can make the condition easier to satisfy
while changing the reference dynamics. The result bounds
distance from that reference process and makes no assertion
about the attained objective value.

\section{Relationship to existing methods and audit}
\label{amj:audit}

Population-based training interleaves improvement at the level
of individual workers with population-level copying and
hyperparameter changes~\cite{amj-jaderberg}. Policy-space response
oracles expand policy populations and compute a mixture from
estimated empirical game payoffs~\cite{amj-lanctot}. Those
methods supply meaningful settings for the inner/outer split.
Automated Design of Agentic Systems searches over versions of
agent programs using measured performance~\cite{amj-hu}. The
outer selection studied here can govern allocation among
such versions when the relevant common score law is specified.
The journal's bilevel primer distinguishes the credit assigned
to an inner adaptation from the method that proposes the next
adaptation~\cite{amj-bilevel}. Here the contribution is to
characterize when a particular outer
selection respects reduced representations and how its results
change when its evaluation inputs change. The selection rule is
an established multiplicative update. Its use in this model
neither chooses the policy search method nor fixes a downstream
learning objective.

The outer computation can be replayed given its input measures,
weights, and variation kernel. A defensible deployment record
can associate each score with a strategy version, the evaluation
contract, the opponent or task law, and a record of how uncertainty
was estimated. A reduced run also specifies the quotient and the
condition under which variation respects it. For distributed
evaluation, each shard's original mixture mass and normalizer are
needed when combining selected measures. A comparison between
exact and approximate runs further states which parts of the
transition are shared, since Proposition~\ref{amj:propagation}
isolates errors in scores. These are sufficient descriptions of
the mathematical inputs. Artifacts that certify the implementation
of the inner scorer or validate it on a task distribution address
additional empirical questions.

A useful next empirical question is whether a declared uniform
or relative error bound is attainable with real evaluation
budgets. If a randomized evaluation is repeated adaptively,
a confidence statement for a single, preselected comparison
does not by itself give a simultaneous bound on all strategy
scores used in the trajectory. A study can report the target
population, sources of randomness, dependence among score
estimates, and the baseline outer transition alongside any
policy-search or routing metrics. The mathematical bounds
can then be applied at the same level at which the empirical
guarantee is established.

\subsection{Mechanized algebra}

The companion Lean 4.19 project checks the finite-type algebra
for the replicator drift and an affine reproductive update.
For a population vector $x$ of mass one and a payoff vector $u$,
write $D_i=x_i(u_i-\sum_j x_ju_j)$. Choosing weights
$1+\eta u_i$ in~\eqref{amj:tilt}, under the hypotheses that
these weights are positive, gives a selected vector with increment
\[
 \mathcal S_{1+\eta u}(x)_i-x_i
   =\frac{\eta D_i}{1+\eta\sum_j x_j u_j}.
\]
The Lean identity also holds algebraically under the weaker
hypothesis that the denominator is nonzero.
\texttt{selectionStep\_affine\_sub} verifies the increment.
The same module, \texttt{EvoGame.Replicator}, checks
the simplex conditions under strictly positive reproductive
factors, conservation of total mass for the drift, and
invariance of the drift under a common shift in all scores.

The module \texttt{EvoGame.Aggregation} checks that summing
over fibres intertwines drift and normalized selection when
the fine payoff is pulled back from the quotient. It also checks
composition of reindexings and the population-dependent
and matrix versions of drift aggregation. These identities
quantify over arbitrary finite strategy types. The measure-theoretic
pushforward, exponential weights and stability results above are
proved in the manuscript. The formal check addresses neither
statistical score estimation nor the execution of agents.
\balance
\begin{thebibliography}{99}
\raggedright

\bibitem{amj-alexander}
J.~M. Alexander.
\emph{Evolutionary Game Theory}.
Elements in Decision Theory and Philosophy, Cambridge University Press,
2023.
\href{https://doi.org/10.1017/9781108582063}{doi:10.1017/9781108582063}.

\bibitem{amj-arora}
S. Arora, E. Hazan, and S. Kale.
The multiplicative weights update method: a meta-algorithm and applications.
\emph{Theory of Computing}, 8:121--164, 2012.
\href{https://theoryofcomputing.org/articles/v008a006/}{doi:10.4086/toc.2012.v008a006}.

\bibitem{amj-sorin}
S. Sorin.
Replicator dynamics: old and new.
\emph{Journal of Dynamics and Games}, 7(4):365--386, 2020.
\href{https://doi.org/10.3934/jdg.2020028}{doi:10.3934/jdg.2020028}.

\bibitem{amj-fritz}
T. Fritz.
A synthetic approach to Markov kernels, conditional independence and
theorems on sufficient statistics.
\emph{Advances in Mathematics}, 370:107239, 2020.
\href{https://doi.org/10.1016/j.aim.2020.107239}{doi:10.1016/j.aim.2020.107239}.

\bibitem{amj-cohen-fausti}
S.~N. Cohen and E. Fausti.
Hyperbolic contractivity and the Hilbert metric on probability measures.
Preprint, 2023, revised 2024.
\href{https://arxiv.org/abs/2309.02413v2}{arXiv:2309.02413v2}.

\bibitem{amj-gaubert-qu}
S. Gaubert and Z. Qu.
Dobrushin's ergodicity coefficient for Markov operators on cones.
\emph{Integral Equations and Operator Theory}, 81(1):127--150, 2015.
\href{https://doi.org/10.1007/s00020-014-2193-2}{doi:10.1007/s00020-014-2193-2}.

\bibitem{amj-jaderberg}
M. Jaderberg et al.
Population based training of neural networks.
Preprint, 2017.
\href{https://arxiv.org/abs/1711.09846}{arXiv:1711.09846}.

\bibitem{amj-lanctot}
M. Lanctot et al.
A unified game-theoretic approach to multiagent reinforcement learning.
In \emph{Advances in Neural Information Processing Systems}, volume 30, 2017.
\href{https://proceedings.neurips.cc/paper_files/paper/2017/hash/3323fe11e9595c09af38fe67567a9394-Abstract.html}{Proceedings article}.

\bibitem{amj-hu}
S. Hu, C. Lu, and J. Clune.
Automated design of agentic systems.
In \emph{International Conference on Learning Representations}, 2025.
\href{https://proceedings.iclr.cc/paper_files/paper/2025/file/36b7acf6f6010652b3f2a433774a66fe-Paper-Conference.pdf}{Proceedings article}.

\bibitem{amj-bilevel}
Meta Journal of Agentic Meta Algorithms.
Self-improvement as a bilevel program: inner adaptation and outer credit.
Editorial primer, 2026.
\href{https://agentic-meta-journal.pages.dev/research/self-improvement-as-a-bilevel-program}{Journal article}.

\end{thebibliography}
\end{document}
