% Final thesis defense (PHANTOM)
% Build: cd paper/defense && pdflatex defense.tex && pdflatex defense.tex
\documentclass[aspectratio=169,11pt]{beamer}

% Narrative and visual refinements for final defense delivery.

\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage{microtype}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{xspace}
\usepackage{booktabs}
\usepackage{appendixnumberbeamer}
\usepackage{hyperref}
\usepackage{tikz}
\usetikzlibrary{arrows.meta,calc,positioning,fit,shapes.geometric,shapes.misc}

\graphicspath{{../src/chapters/figures/results/generated/final/plots/}{../src/chapters/}}

\usetheme[
  progressbar=frametitle,
]{moloch}
\molochset{sectionpage=none,subsectionpage=none}
\usefonttheme{professionalfonts}
\setbeamertemplate{frame numbering}[fraction]

% Palette
\definecolor{PhantomPaper}{HTML}{F6F1E9}
\definecolor{PhantomInk}{HTML}{0F1B2D}
\definecolor{PhantomSlate}{HTML}{6F89A3}
\definecolor{PhantomCyan}{HTML}{C97A3D}
\definecolor{PhantomIndigo}{HTML}{2F8F8A}
\definecolor{PhantomPeach}{HTML}{EEC39C}

\setbeamercolor{normal text}{fg=PhantomPaper,bg=PhantomInk}
\setbeamercolor{background canvas}{bg=PhantomInk}
\setbeamercolor{alerted text}{fg=PhantomCyan!35!white}
\setbeamercolor{example text}{fg=PhantomIndigo!35!white}
\setbeamercolor{palette primary}{fg=PhantomPaper,bg=PhantomInk!95!black}
\setbeamercolor{frametitle}{parent=palette primary}
\setbeamercolor{progress bar}{fg=PhantomCyan,bg=PhantomPaper!20!PhantomInk}
\setbeamercolor{title separator}{use=progress bar,parent=progress bar}
\setbeamercolor{structure}{fg=PhantomCyan!20!white}
\setbeamercolor{block title}{fg=PhantomPaper,bg=PhantomIndigo!65!black}
\setbeamercolor{block body}{fg=PhantomPaper,bg=PhantomInk!72!PhantomIndigo}
\setbeamercolor{alertblock title}{fg=PhantomPaper,bg=PhantomCyan!82!black}
\setbeamercolor{alertblock body}{fg=PhantomPaper,bg=PhantomInk!72!PhantomCyan}
\setbeamercolor{exampleblock title}{fg=PhantomPaper,bg=PhantomIndigo!58!black}
\setbeamercolor{exampleblock body}{fg=PhantomPaper,bg=PhantomInk!72!PhantomIndigo}

\setbeamertemplate{navigation symbols}{}
\setbeamertemplate{itemize item}{\small\raise0.3ex\hbox{$\bullet$}}
\setbeamertemplate{itemize subitem}{\tiny\raise0.2ex\hbox{$\circ$}}

\hypersetup{colorlinks=true,urlcolor=PhantomCyan!40!white,linkcolor=PhantomPaper}

\title{PHANTOM}
\subtitle{Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms}
\author{Daniel Rösel}
\institute{IE University, Madrid \\ Supervisor: Alberto Martín Izquierdo}
\date{\today}

\titlegraphic{%
  \begin{tikzpicture}
    \shade[left color=PhantomCyan,right color=PhantomIndigo] (0,0) rectangle (0.55\paperwidth,0.06);
  \end{tikzpicture}%
}

\newcommand{\stagebar}[1]{}

\newcommand{\metriccard}[2]{%
  \begin{tikzpicture}
    \node[
      draw=PhantomInk,
      rounded corners=3pt,
      fill=PhantomCyan!10,
      minimum width=3.05cm,
      minimum height=1.25cm,
      align=center
    ] {\Large\bfseries #1\\[-0.2em]{\scriptsize #2}};
  \end{tikzpicture}%
}

\newcommand{\humaniconraw}{%
  \begin{tikzpicture}[x=0.9ex,y=0.9ex]
    \fill[PhantomIndigo] (0,1.55) circle (0.42);
    \draw[PhantomInk,line width=0.20pt,fill=PhantomSlate!95!black] (0,0.0) ellipse (0.72 and 0.56);
  \end{tikzpicture}%
}

\newcommand{\roboticonraw}{%
  \begin{tikzpicture}[x=0.9ex,y=0.9ex]
    \draw[PhantomInk,line width=0.20pt,rounded corners=0.35ex,fill=PhantomPeach] (-0.95,-0.78) rectangle (0.95,0.72);
    \draw[PhantomInk,line width=0.20pt,fill=white] (-0.42,0.08) circle (0.21);
    \draw[PhantomInk,line width=0.20pt,fill=white] (0.42,0.08) circle (0.21);
    \fill[PhantomInk] (-0.42,0.08) circle (0.07);
    \fill[PhantomInk] (0.42,0.08) circle (0.07);
    \draw[PhantomInk,line width=0.20pt] (-0.30,-0.30) -- (0.30,-0.30);
    \draw[PhantomInk,line width=0.20pt] (0,0.72) -- (0,1.03);
    \fill[PhantomIndigo] (0,1.15) circle (0.10);
  \end{tikzpicture}%
}

\newcommand{\humanicon}{\raisebox{-0.45ex}{\humaniconraw}\xspace}
\newcommand{\roboticon}{\raisebox{-0.45ex}{\roboticonraw}\xspace}
\newcommand{\usersagentslabel}{Users \humanicon + agents \roboticon}
\newcommand{\humanagentpair}{\humanicon, \roboticon}
\newcommand{\humanagentmix}{\humanicon/\roboticon}

\tikzset{every node/.append style={text=PhantomSlate}}

\begin{document}

{
\setbeamercolor{background canvas}{bg=PhantomInk}
\begin{frame}[plain]
  \vfill
  \centering
  {\color{white}\Huge\bfseries PHANTOM\par}
  \vspace{0.6em}
  {\color{PhantomCyan}\rule{0.45\paperwidth}{0.06cm}\par}
  \vspace{0.8em}
  {\large\color{white!90!black}Pricing heuristics against non-human transaction orchestration\par}
  \vfill
  {\color{white!75!black}\normalsize Daniel Rösel\par}
  {\color{white!65!black}\small IE University \textbullet\ Supervisor: Alberto Martín Izquierdo\par}
  \vspace{1.2em}
  {\footnotesize\color{PhantomCyan!80!white}\href{https://velocitatem.github.io/PHANTOM/}{\texttt{velocitatem.github.io/PHANTOM}}}
  \vfill
\end{frame}
}

\begin{frame}{Roadmap: one argument in six stages (15 min)}
  \centering
  \begin{tikzpicture}[
    font=\scriptsize\sffamily,
    stage/.style={draw=PhantomInk,rounded corners=3pt,fill=PhantomCyan!10,minimum width=1.95cm,minimum height=1.05cm,align=center},
    flow/.style={-{Stealth[length=2.0mm,width=1.8mm]},line width=1pt,PhantomSlate}
  ]
    \node[stage,fill=PhantomCyan!14] (intro) {Intro\\2m};
    \node[stage,right=0.30cm of intro] (platform) {Platform\\4m};
    \node[stage,right=0.30cm of platform] (signal) {Signal\\4m};
    \node[stage,right=0.30cm of signal] (drrl) {DR-RL\\4m};
    \node[stage,right=0.30cm of drrl] (results) {Results\\1m};
    \node[stage,right=0.30cm of results] (close) {Close};

    \draw[flow,shorten <=2pt,shorten >=2pt] (intro.east) -- (platform.west);
    \draw[flow,shorten <=2pt,shorten >=2pt] (platform.east) -- (signal.west);
    \draw[flow,shorten <=2pt,shorten >=2pt] (signal.east) -- (drrl.west);
    \draw[flow,shorten <=2pt,shorten >=2pt] (drrl.east) -- (results.west);
    \draw[flow,shorten <=2pt,shorten >=2pt] (results.east) -- (close.west);
  \end{tikzpicture}

  \vspace{0.75em}
  \begin{block}{Main research question}
    How can dynamic pricing preserve margin integrity when transactions are increasingly mediated by non-human agents?
  \end{block}
  \vspace{0.35em}
  {\footnotesize Dynamic pricing has often been treated as a secondary optimization layer; agent-mediated shopping turns it into a primary margin-risk surface.}
  \stagebar{1}
\end{frame}

\begin{frame}{Motivation: one everyday pricing story}
  \footnotesize
  \begin{columns}[T,onlytextwidth]
    \column{0.53\textwidth}
    \begin{block}{Imagine you sell weekend hotel rooms online}
      A customer asks an assistant to scout many quotes first, then buys in a clean session at the best discovered price.
    \end{block}
    \begin{alertblock}{Why this matters to everyday people}
      If this behavior is untreated, honest shoppers can face noisier prices and a weaker shopping experience because pricing reacts to manipulated intent signals.
    \end{alertblock}

    \column{0.44\textwidth}
    \centering
    \begin{tikzpicture}[
      font=\scriptsize\sffamily,
      card/.style={draw=PhantomInk,rounded corners=5pt,minimum width=3.8cm,minimum height=1.0cm,align=center},
      flow/.style={-{Stealth[length=2.2mm]},thick,PhantomSlate}
    ]
      \node[card,fill=PhantomCyan!15] (seller) at (0,1.55) {Seller posts rooms};
      \node[card,fill=white] (recon) at (0,0.2) {Recon by agent \roboticon};
      \node[card,fill=PhantomIndigo!12] (buy) at (0,-1.15) {Purchase by user \humanicon};
      \draw[flow] (seller) -- (recon);
      \draw[flow] (recon) -- (buy);
      \node[font=\tiny\itshape,text=PhantomSlate] at (0,-1.95) {query and purchase split across sessions};
    \end{tikzpicture}
  \end{columns}

  \vspace{-0.15em}
  {\scriptsize\textbf{Takeaway:} protect legitimate shoppers \humanicon while detecting orchestrated recon \roboticon before pricing leakage compounds.}
  \stagebar{1}
\end{frame}

\begin{frame}{Policy first: one rule maps context into price actions}
  \begin{columns}[T,onlytextwidth]
    \column{0.55\textwidth}
    \begin{block}{Policy definition}
      \[
        p_t = \pi(x_t)
      \]
      where context \(x_t\) includes product state, time, and behavior signals from the session.
    \end{block}
    \begin{itemize}[<+->]
      \item Behavior proxy \(\hat q\) is tracked for both user-like and agent-like sessions \((\humanagentpair)\).
      \item The score \(f(\tau')\) is a soft estimate that a trajectory is agent-mediated \roboticon.
      \item We see reward only for the chosen price action, which motivates a contextual-bandit view first.
    \end{itemize}

    \column{0.43\textwidth}
    \centering
    \begin{tikzpicture}[
      font=\scriptsize\sffamily,
      box/.style={draw=PhantomInk,rounded corners=4pt,minimum width=3.35cm,minimum height=0.85cm,align=center},
      flow/.style={-{Stealth[length=2.0mm]},thick,PhantomSlate}
    ]
      \node[box,fill=white] (ctx) at (0,1.35) {Context \(x_t\)};
      \node[box,fill=PhantomIndigo!12] (pol) at (0,0.15) {Policy \(\pi\)};
      \node[box,fill=PhantomCyan!15] (act) at (0,-1.05) {Price action \(p_t\)};
      \draw[flow] (ctx) -- (pol);
      \draw[flow] (pol) -- (act);
      \node[font=\tiny\itshape,text=PhantomSlate] at (0,-1.75) {later extended from contextual bandits to DR-RL};
    \end{tikzpicture}
  \end{columns}
  \stagebar{1}
\end{frame}

\begin{frame}{Agentic recon creates direct financial pressure on pricing power}
  \centering
  \begin{tikzpicture}[
    font=\small\sffamily,
    flow/.style={draw=PhantomInk,rounded corners=6pt,minimum width=5.3cm,minimum height=1.25cm,align=center},
    note/.style={draw=PhantomInk!55,rounded corners=4pt,minimum width=11.2cm,minimum height=0.95cm,align=center,fill=white,font=\scriptsize}
  ]
    \path[use as bounding box] (-6.1,-1.25) rectangle (6.1,2.25);
    \node<1->[flow,fill=PhantomCyan!18] (recon) at (-3.1,1.1)
      {\textbf{Recon session \roboticon}\\samples multiple quotes};
    \node<2->[flow,fill=PhantomIndigo!16] (buy) at (3.1,1.1)
      {\textbf{Clean execution session \humanicon}\\buys using the best found quote};
    \draw<2->[-{Stealth[length=3mm]},ultra thick,PhantomSlate] (recon.east) -- (buy.west);
    \node<2->[font=\scriptsize\bfseries,text=PhantomSlate] at (0,1.98)
      {query and purchase are decoupled};
    \draw<3->[densely dashed,thick,PhantomCyan!90!black]
      (recon.south east) .. controls +(1.15,-0.95) and +(-1.15,-0.95) .. (buy.south west);
    \node<3->[note] at (0,-0.65)
      {The platform sees behavior proxy $\hat q$ (\humanagentpair), while true demand response $d(p\mid\theta)$ stays latent.};
  \end{tikzpicture}

  \vspace{0.05em}
  \begin{columns}[T,onlytextwidth]
    \column{0.31\textwidth}
    \uncover<4->{%
      \centering
      \begin{tikzpicture}[font=\scriptsize\sffamily]
        \node[draw=PhantomInk,rounded corners=4pt,fill=PhantomInk,text=white,minimum width=0.97\linewidth,text width=0.84\linewidth,minimum height=1.2cm,align=center]
          {\large$\mathrm{COI}(\pi)=\mathbb{E}[P]-\underline p$\\[-0.05em]\footnotesize pricing power KPI};
      \end{tikzpicture}%
    }

    \column{0.31\textwidth}
    \uncover<5->{%
      \centering
      \begin{tikzpicture}[font=\scriptsize\sffamily]
        \node[draw=PhantomInk,rounded corners=4pt,fill=PhantomIndigo!12,text=PhantomSlate,minimum width=0.97\linewidth,text width=0.84\linewidth,minimum height=1.2cm,align=center]
          {\large$\lim_{N\to\infty}\mathrm{COI}=0$\\[-0.05em]\footnotesize theorem as intuition guide};
      \end{tikzpicture}%
    }

    \column{0.34\textwidth}
    \uncover<6->{%
      \centering
      \begin{tikzpicture}[x=0.67cm,y=0.85cm,font=\scriptsize\sffamily]
        \draw[->,thick,PhantomSlate] (0,0) -- (4.2,0) node[right] {queries $N$};
        \draw[->,thick,PhantomSlate] (0,0) -- (0,2.05) node[above] {COI};
        \draw[very thick,PhantomCyan!95!black] (0.25,1.8) .. controls (1.2,1.35) and (2.25,0.62) .. (4.0,0.16);
        \draw[dashed,PhantomInk!65] (0,0.16) -- (4.0,0.16);
        \node[anchor=west,font=\tiny,text=PhantomSlate] at (2.35,0.28) {price-floor proximity};
      \end{tikzpicture}%
    }
  \end{columns}

  \vspace{-0.1em}
  \uncover<6->{\scriptsize\textit{The theorem gives direction, not prophecy: more independent recon pressure pushes realizable prices toward the floor.}}\\[0.1em]
  \uncover<7->{\scriptsize\textbf{Implication:} when quote discovery and purchase split, session-based pricing can overestimate willingness to pay.}
  \stagebar{1}
\end{frame}

\begin{frame}{The thesis answers one chain: mechanism \(\to\) signal \(\to\) control}
  \begin{columns}[T,onlytextwidth]
    \column{0.32\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily]
      \draw[rounded corners=4pt,draw=PhantomInk,fill=white] (-1.55,-1.1) rectangle (1.55,1.2);
      \fill[PhantomCyan!85!black] (-0.75,0.35) circle (0.14);
      \fill[PhantomCyan!85!black] (-0.45,0.70) circle (0.14);
      \fill[PhantomCyan!85!black] (-0.15,0.45) circle (0.14);
      \fill[PhantomIndigo!85!black] (0.35,-0.20) circle (0.14);
      \fill[PhantomIndigo!85!black] (0.65,-0.45) circle (0.14);
      \fill[PhantomIndigo!85!black] (0.95,-0.15) circle (0.14);
      \draw[dashed,PhantomInk!60] (0.12,-0.92) -- (0.12,1.0);
      \node[text=PhantomSlate,font=\tiny] at (0,-0.93) {behavior separability};
    \end{tikzpicture}
    {\footnotesize\textbf{SQ1}}\\[-0.15em]
    {\scriptsize Can we distinguish \humanicon and \roboticon sessions from interactions alone?}

    \column{0.32\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily]
      \draw[rounded corners=4pt,draw=PhantomInk,fill=white] (-1.55,-1.1) rectangle (1.55,1.2);
      \draw[->,thick,PhantomSlate] (-1.2,-0.75) -- (1.2,-0.75);
      \draw[->,thick,PhantomSlate] (-1.2,-0.75) -- (-1.2,0.85);
      \draw[very thick,PhantomCyan!95!black] (-1.0,0.62) .. controls (-0.4,0.2) and (0.3,-0.18) .. (1.0,-0.58);
      \node[text=PhantomSlate,font=\tiny] at (0,-0.95) {COI / revenue pressure};
    \end{tikzpicture}
    {\footnotesize\textbf{SQ2}}\\[-0.15em]
    {\scriptsize How strong is price and revenue erosion under agentic contamination?}

    \column{0.32\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily]
      \draw[rounded corners=4pt,draw=PhantomInk,fill=white] (-1.55,-1.1) rectangle (1.55,1.2);
      \draw[thick,fill=PhantomIndigo!20,draw=PhantomInk] (0,0.82) -- (0.98,0.32) -- (0.98,-0.44) -- (0,-0.90) -- (-0.98,-0.44) -- (-0.98,0.32) -- cycle;
      \draw[thick,PhantomInk] (0,-0.46) -- (0,0.38);
      \draw[thick,PhantomInk] (0,-0.46) -- (0.42,-0.08);
      \node[text=PhantomSlate,font=\tiny] at (0,-0.95) {robust policy control};
    \end{tikzpicture}
    {\footnotesize\textbf{SQ3}}\\[-0.15em]
    {\scriptsize Can policy design recover margin while keeping UX stable?}
  \end{columns}

  \vspace{0.2em}
  \stagebar{1}
\end{frame}

\section{Platform Development}

\begin{frame}{Stage 1: We built a dual-loop platform to observe behavior and price exposure together}
  \centering
  \begin{tikzpicture}[
    font=\scriptsize\sffamily,
    box/.style={draw=PhantomInk,rounded corners=3pt,minimum width=2.5cm,minimum height=0.9cm,align=center},
    arr/.style={-{Stealth[length=2.2mm]},thick,PhantomSlate}
  ]
    \node[box,fill=PhantomCyan!14] (actors) at (0,1.45) {\usersagentslabel};
    \node[box,fill=white] (web) at (2.9,1.45) {Web\\storefront};
    \node[box,fill=white] (provider) at (5.8,1.45) {Pricing\\service};
    \node[box,fill=white] (redis) at (8.7,1.45) {Serve\\cache};
    \node[box,fill=PhantomIndigo!10,minimum width=3.1cm] (kafka) at (4.35,-0.15) {Event stream\\behavior + quote logs};
    \node[box,fill=PhantomCyan!10,minimum width=2.8cm] (airflow) at (8.0,-0.15) {Offline trainer\\batch updates};

    \draw[arr] (actors) -- (web);
    \draw[arr] (web) -- (provider);
    \draw[arr] (provider) -- (redis);
    \draw[arr] (web.south) -- (kafka.north west);
    \draw[arr] (provider.south) -- (kafka.north east);
    \draw[arr] (kafka) -- (airflow);
    \draw[arr] (airflow.north) -| (redis.south);
    \draw[arr] (redis.west) to[bend left=35] (provider.east);

    \node[font=\tiny\itshape,text=PhantomSlate] at (2.2,-1.0) {Kappa: streaming telemetry};
    \node[font=\tiny\itshape,text=PhantomSlate] at (8.1,-1.0) {Lambda: offline learning + refresh};
  \end{tikzpicture}

  \vspace{0.35em}
  \begin{itemize}[<+->]
    \item Every quote has a matching behavioral context in the log stream.
    \item The same architecture supports reproducible stress tests before any live deployment.
  \end{itemize}
  \stagebar{2}
\end{frame}

\begin{frame}{Dataset card: compact, labeled, and experiment-ready}
  \begin{columns}[T,onlytextwidth]
    \column{0.60\textwidth}
    \centering
    \begin{tikzpicture}[
      font=\scriptsize\sffamily,
      chip/.style={draw=PhantomInk!40,rounded corners=2pt,inner sep=2.7pt},
      body/.style={anchor=west,text width=6.0cm,align=left,font=\scriptsize}
    ]
      \node[draw=PhantomInk,rounded corners=5pt,fill=white,minimum width=6.85cm,minimum height=4.45cm] at (0,0) {};
      \node[anchor=west,font=\footnotesize\bfseries,text=PhantomInk] at (-3.2,1.72) {WhoClickedIt dataset card};
      \node[anchor=west,draw=PhantomInk!35,rounded corners=2pt,fill=PhantomCyan!10,inner xsep=4pt,inner ysep=3pt,text width=6.15cm,align=left,font=\tiny\ttfamily,text=PhantomSlate] at (-3.2,1.22)
        {huggingface.co/datasets/velocitatem/whoclickedit};

      \node[anchor=west,chip,fill=PhantomIndigo!12] (humanrows) at (-3.2,0.52) {\textbf{Human rows} 798};
      \node[anchor=west,chip,fill=PhantomIndigo!12] at ([xshift=0.16cm]humanrows.east) {\textbf{Agent rows} 3076};

      \node[body,text=PhantomSlate] at (-3.2,-0.33)
        {Flat schema and explicit actor labels simplify session-aware train/test splits.};
      \node[body,font=\tiny\itshape,text=PhantomSlate!85] at (-3.2,-1.01)
        {Kafka provenance is retained for reproducibility and downstream analysis.};
    \end{tikzpicture}

    \column{0.38\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily,
      stat/.style={draw=PhantomInk,rounded corners=5pt,minimum width=4.95cm,minimum height=1.33cm,align=center}]
      \node<1->[stat,fill=PhantomInk,text=white] at (0,1.95)
        {\Large\bfseries 29 Interviews\\[-0.1em]\footnotesize labeled trajectories in observed samples};
      \node<2->[stat,fill=PhantomCyan!14,text=PhantomSlate] at (0,0.25)
        {\Large\bfseries 45\% / 55\%\\[-0.1em]\footnotesize human/agent trajectory split};
      \node<3->[stat,fill=PhantomIndigo!12,text=PhantomSlate] at (0,-1.45)
        {\Large\bfseries 2 streams\\[-0.1em]\footnotesize interaction + price-log records};
    \end{tikzpicture}
  \end{columns}

  \stagebar{2}
\end{frame}

\begin{frame}{Experimental design controls goals, not instructions}
  \begin{columns}[T,onlytextwidth]
    \column{0.58\textwidth}
    \centering
    \begin{tikzpicture}[
      font=\scriptsize\sffamily,
      box/.style={draw=PhantomInk,rounded corners=3pt,minimum width=3.65cm,minimum height=0.95cm,align=center},
      arr/.style={-{Stealth[length=2.2mm]},thick,PhantomSlate}
    ]
      \node[box,fill=PhantomCyan!14] (tasks) at (0,1.8) {JTBD task pool\\hotel + airline modes};
      \node[box,fill=white] (assign) at (0,0.55) {Random assignment\\mode + task + actor id};
      \node[box,fill=PhantomIndigo!12] (run) at (0,-0.7) {Execution\\human or browser-use agent};
      \node[box,fill=white] (logs) at (0,-1.95) {Session logs\\$e=(a,i,t,\mu,\delta)$ + quotes};
      \draw[arr] (tasks) -- (assign);
      \draw[arr] (assign) -- (run);
      \draw[arr] (run) -- (logs);
    \end{tikzpicture}

    \column{0.40\textwidth}
    \begin{itemize}[<+->]\setlength{\itemsep}{0.55em}
      \item Agents run with \textbf{browser-use} and a model-swappable LLM router (default \texttt{gpt-5-mini}).
      \item Tasks are defined by outcomes, not scripted clicks, to preserve behavioral variety.
      \item Current release is stronger on hotel flows than airline flows.
    \end{itemize}
  \end{columns}
  \stagebar{2}
\end{frame}

\section{Distinguishability Construction}

\begin{frame}{Stage 2: A behavior kernel is a compact signature of navigation dynamics}
  \begin{columns}[T,onlytextwidth]
    \column{0.48\textwidth}
    \begin{block}{Definition}
      \[
        \hat P(s'\mid s)=\frac{N(s,s')}{\sum_k N(s,k)}
      \]
    \end{block}
    \begin{itemize}[<+->]
      \item Build one kernel per session, then prototypes for human and agent cohorts.
      \item Compare each incoming session to both prototypes with KL divergence.
    \end{itemize}

    \column{0.50\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily]
      \node[draw=PhantomInk,rounded corners=3pt,fill=PhantomCyan!12,minimum width=3.9cm,minimum height=0.85cm] (a) at (0,1.4) {page\_view};
      \node[draw=PhantomInk,rounded corners=3pt,fill=white,minimum width=3.9cm,minimum height=0.85cm] (b) at (0,0.25) {view\_item\_page};
      \node[draw=PhantomInk,rounded corners=3pt,fill=PhantomIndigo!12,minimum width=3.9cm,minimum height=0.85cm] (c) at (0,-0.9) {add\_item\_to\_cart};
      \draw[-{Stealth[length=2.2mm]},thick,PhantomSlate] (a) -- node[right,font=\tiny]{0.64} (b);
      \draw[-{Stealth[length=2.2mm]},thick,PhantomSlate] (b) -- node[right,font=\tiny]{0.31} (c);
      \draw[-{Stealth[length=2.2mm]},thick,PhantomSlate!70] (b.east) .. controls +(1.1,0.5) and +(1.1,-0.5) .. node[right,font=\tiny]{0.52} (b.east);
      \node[font=\tiny\itshape,text=PhantomSlate] at (0,-1.7) {Kernel rows encode ``what usually comes next.''};
    \end{tikzpicture}
  \end{columns}
  \stagebar{3}
\end{frame}

\begin{frame}{Human and agent kernels are separable in the controlled cohort}
  \begin{columns}[T,onlytextwidth]
    \column{0.48\textwidth}
    \centering
    \textbf{Human transition structure}\par\vspace{0.2em}
    \includegraphics[width=\linewidth,height=0.46\textheight,keepaspectratio]{mdp_human.pdf}
    \column{0.48\textwidth}
    \centering
    \textbf{Agent transition structure}\par\vspace{0.2em}
    \includegraphics[width=\linewidth,height=0.46\textheight,keepaspectratio]{mdp_agent.pdf}
  \end{columns}

  \vspace{0.15em}
  \begin{columns}[T,onlytextwidth]
    \column{0.32\textwidth}\centering\metriccard{-3.35}{mean gap (human)}
    \column{0.32\textwidth}\centering\metriccard{+1.65}{mean gap (agent)}
    \column{0.32\textwidth}\centering\metriccard{$p<0.001$}{Mann-Whitney rank test}
  \end{columns}
  \stagebar{3}
\end{frame}

\begin{frame}{Two divergence scores become one continuous control signal}
  \centering
  \[
    \only<1>{\Delta_H = D_{KL}(\hat T'\mid\mid\bar T_H),\quad \Delta_A = D_{KL}(\hat T'\mid\mid\bar T_A)}%
    \only<2>{g(\tau') = \Delta_H-\Delta_A}%
    \only<3->{f(\tau') = P(A\mid\tau') = \sigma\!\left(\frac{g(\tau')}{T}\right)}
  \]

  \vspace{0.4em}
  \begin{tikzpicture}[font=\scriptsize\sffamily]
    \draw<2->[very thick,PhantomSlate] (-4,0) -- (4,0);
    \draw<2->[thick,PhantomSlate] (0,-0.16) -- (0,0.16);
    \node<2->[anchor=north] at (-4,0) {human-like};
    \node<2->[anchor=north] at (4,0) {agent-like};
    \node<3->[anchor=north] at (0,0) {$g(\tau')=0$};
    \fill<4->[PhantomCyan!75!black] (-2.2,0) circle (2.2pt);
    \node<4->[anchor=south,text=PhantomCyan!75!black] at (-2.2,0) {low $f(\tau')$};
    \fill<5->[PhantomIndigo!75!black] (2.2,0) circle (2.2pt);
    \node<5->[anchor=south,text=PhantomIndigo!75!black] at (2.2,0) {high $f(\tau')$};
  \end{tikzpicture}

  \vspace{0.25em}
  \begin{itemize}
    \item<3-> The signed gap $g(\tau')$ is positive when a session is closer to agent behavior \roboticon (vs. human reference \humanicon).
    \item<4-> Temperature $T$ calibrates how sharply the score moves away from uncertainty.
    \item<6-> Continuous scoring is used to steer contamination-aware pricing.
    \item<7-> The design target is guidance, not a hard user-level ban decision.
  \end{itemize}
  \stagebar{3}
\end{frame}

\section{Distributionally Robust RL}

\begin{frame}{Stage 3: DR-RL trains against plausible contamination shifts, not one fixed world}
  \small
  \begin{columns}[T,onlytextwidth]
    \column{0.48\textwidth}
    \begin{block}{Ideal robust object}
      \[
        \mathcal U_\epsilon(\hat P_N)=\{Q: W_p(Q,\hat P_N)\le\epsilon\}
      \]
      \centering
      robust against distribution shift around the empirical demand law
    \end{block}

    \column{0.50\textwidth}
    \begin{block}{Engine approximation used in experiments}
      \[
        \mathcal A_{\epsilon_\alpha}(\alpha_0)=\{\alpha:|\alpha-\alpha_0|\le\epsilon_\alpha\}
      \]
      \centering
      small grid over $\alpha$ \;\textrightarrow\; inner worst-case candidate
    \end{block}
  \end{columns}
  \vspace{0.2em}
  \begin{alertblock}{Practical boundary}
    In code we solve a local robust loop around $\alpha_0$, not the full continuous Wasserstein adversary.
  \end{alertblock}
  \stagebar{4}
\end{frame}

\begin{frame}{Reward composition penalizes leakage while guarding user experience}
  \[
    \only<1>{%
      r_t =
      {\color{PhantomInk}\underline{R(p_t,\hat Q_t)}}%
    }%
    \only<2>{%
      r_t =
      {\color{PhantomInk}\underline{R(p_t,\hat Q_t)}}
      - {\color{PhantomCyan!95!black}\underline{\lambda\,f(\tau'_t)\,c_{\text{info}}}}%
    }%
    \only<3->{%
      r_t =
      {\color{PhantomInk}\underline{R(p_t,\hat Q_t)}}
      - {\color{PhantomCyan!95!black}\underline{\lambda\,f(\tau'_t)\,c_{\text{info}}}}
      - {\color{PhantomIndigo!95!black}\underline{\eta_{\text{ux}}\,UX(\tau'_t,p_t)}}%
    }%
  \]

  \vspace{0.45em}
  \begin{columns}[T,onlytextwidth]
    \column{0.32\textwidth}
    \centering
    \begin{tikzpicture}[font=\scriptsize\sffamily]
      \node[
        draw=PhantomInk,
        rounded corners=4pt,
        fill=PhantomInk!12,
        minimum width=0.98\linewidth,
        text width=0.88\linewidth,
        minimum height=1.28cm,
        align=center,
        text=PhantomInk
      ] {\textbf{Revenue term}\\[-0.08em]keeps market objective explicit};
    \end{tikzpicture}
    \column{0.32\textwidth}
    \centering
    \uncover<2->{%
      \begin{tikzpicture}[font=\scriptsize\sffamily]
        \node[
          draw=PhantomInk,
          rounded corners=4pt,
          fill=PhantomCyan!16,
          minimum width=0.98\linewidth,
          text width=0.88\linewidth,
          minimum height=1.28cm,
          align=center,
          text=PhantomCyan!95!black
        ] {\textbf{Leakage term}\\[-0.08em]scales with agent-likelihood score};
      \end{tikzpicture}%
    }
    \column{0.32\textwidth}
    \centering
    \uncover<3->{%
      \begin{tikzpicture}[font=\scriptsize\sffamily]
        \node[
          draw=PhantomInk,
          rounded corners=4pt,
          fill=PhantomIndigo!16,
          minimum width=0.98\linewidth,
          text width=0.88\linewidth,
          minimum height=1.28cm,
          align=center,
          text=PhantomIndigo!95!black
        ] {\textbf{UX term}\\[-0.08em]discourages unstable pricing behavior};
      \end{tikzpicture}%
    }
  \end{columns}

  \vspace{0.25em}
  \begin{itemize}
    \item<2-> Baseline experiments use a query-tax leakage surrogate where higher $f(\tau')$ \roboticon increases leakage penalty.
    \item<3-> Supra-competitive anchor penalties are tracked as an additional safety rail.
  \end{itemize}
  \stagebar{4}
\end{frame}

\begin{frame}{Computationally, wide sweeps are feasible only with aggressive optimization}
  \begin{columns}[T,onlytextwidth]
    \column{0.47\textwidth}
    \centering
    {\Large\(4\times4\times3\times2\times2=\mathbf{192}\)}\\[0.25em]
    {\scriptsize algorithms $\times$ contamination $\times$ robustness $\times$ COI penalty $\times$ action grid}

    \vspace{0.5em}
    \metriccard{160 PFLOPS}{peak aggregate TPU budget}\\[0.45em]
    \metriccard{\textasciitilde180 days}{net compute logged in full study}

    \column{0.51\textwidth}
    \begin{block}{Hot-path rewrite impact}
      \centering
      \begin{tabular}{@{}lcc@{}}
        \toprule
        Mode & Before & After \\
        \midrule
        Baseline step/s & 26.0 & 220.0 \\
        Robust step/s & 7.2 & 136.0 \\
        \bottomrule
      \end{tabular}
    \end{block}
    \vspace{0.1em}
    {\footnotesize
    \begin{itemize}[<+->]
      \item pandas lookup bottlenecks replaced with array/JAX-style loops.
      \item Throughput gains (8.5$\times$, 19$\times$) made broad sweeps practical.
    \end{itemize}}
  \end{columns}
  \stagebar{4}
\end{frame}

\section{Results}

\begin{frame}{Results: contamination hurts revenue; defended policies recover COI}
  \begin{columns}[T,onlytextwidth]
    \column{0.62\textwidth}
    \centering
    \includegraphics[width=\linewidth,height=0.60\textheight,keepaspectratio]{final_focus_coi_by_alpha.pdf}

    \column{0.30\textwidth}
    \metriccard{-90{,}140}{baseline contamination slope}\\[0.3em]
    \metriccard{\textasciitilde3\%}{short-run revenue cost of defense}\\[0.3em]
    \metriccard{Regime-dependent}{COI gains strongest at harder settings}
  \end{columns}
  \stagebar{5}
\end{frame}

\section{Conclusions}

\begin{frame}{Yes, with boundaries: we can defend margin integrity under agentic orchestration}
  \begin{columns}[T,onlytextwidth]
    \column{0.32\textwidth}
    \begin{block}{SQ1\;Distinguishability}
      \centering
      kernels are separable\\$p<0.001$
    \end{block}
    \column{0.32\textwidth}
    \begin{block}{SQ2\;Theoretical impact}
      \centering
      COI erosion mechanism\\proved in baseline limit
    \end{block}
    \column{0.32\textwidth}
    \begin{block}{SQ3\;Mitigation}
      \centering
      robust control shifts\\COI/revenue/UX trade-off
    \end{block}
  \end{columns}

  \vspace{0.35em}
  \begin{alertblock}{Boundary conditions}
    Evidence is from a controlled platform and a small labeled cohort; this is mechanism validation, not full production external validity.
  \end{alertblock}
  \stagebar{6}
\end{frame}

\begin{frame}{What this implies for real pricing systems}
  \begin{itemize}[<+->]\setlength{\itemsep}{0.7em}
    \item \textbf{Financially:} untreated reconnaissance behaves like an information leak and can compress sustainable margins.
    \item \textbf{Operationally:} behavior-only session scoring can be wired into pricing without relying on device fingerprinting.
    \item \textbf{Market exposure:} channels where dynamic pricing has been a secondary layer (aggregators, comparison funnels, promo traffic) are likely to be disrupted first.
    \item \textbf{Strategically:} robust pricing should be calibrated by regime; there is no single penalty that wins everywhere.
    \item \textbf{Before deployment:} larger human baselines, governance review, and legal safeguards are mandatory.
  \end{itemize}
  \stagebar{6}
\end{frame}

\begin{frame}[plain]
  \centering
  \vfill
  {\LARGE\bfseries Thank you}
  \vspace{0.8em}

  {\large Questions and discussion}

  \vfill
  {\footnotesize\color{PhantomSlate!80}Appendix follows: COI theorem derivation, reward composition, and sample-size notes.}
  \vfill
\end{frame}

\appendix
\input{defense_appendix}

\end{document}