velocitatem/PHANTOM
1
0
Fork 0
mirror of https://github.com/velocitatem/PHANTOM.git synced 2026-05-31 08:33:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

early emojification

Browse Source
This commit is contained in:
Daniel Rosel
2026-04-27 17:45:40 +02:00
parent acf5bb5409
commit b677e80b80
6 changed files with 338 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

214
paper/defense/defense.tex
View File

@@ -2,12 +2,15 @@
% 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}
@@ -29,6 +32,7 @@
\definecolor{PhantomSlate}{HTML}{24364C}
\definecolor{PhantomCyan}{HTML}{C97A3D}
\definecolor{PhantomIndigo}{HTML}{2F8F8A}
\definecolor{PhantomPeach}{HTML}{EEC39C}
\setbeamercolor{normal text}{fg=PhantomSlate,bg=PhantomPaper}
\setbeamercolor{alerted text}{fg=PhantomCyan!95!black}
@@ -78,6 +82,32 @@
\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}
\begin{document}
{
@@ -129,6 +159,71 @@
\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}[
@@ -138,44 +233,101 @@
]
\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}\\samples multiple quotes};
{\textbf{Recon session \roboticon}\\samples multiple quotes};
\node<2->[flow,fill=PhantomIndigo!16] (buy) at (3.1,1.1)
{\textbf{Clean execution session}\\buys using the best found quote};
{\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 $\hat q$, while true demand response $d(p\mid\theta)$ stays latent.};
{The platform sees behavior proxy $\hat q$ (\humanagentpair), while true demand response $d(p\mid\theta)$ stays latent.};
\end{tikzpicture}
\vspace{0.25em}
\begin{tikzpicture}[font=\scriptsize\sffamily,
card/.style={draw=PhantomInk,rounded corners=4pt,minimum width=3.85cm,text width=3.55cm,minimum height=1.4cm,align=center}]
\path[use as bounding box] (-6.05,-0.95) rectangle (6.05,0.95);
\node<4->[card,fill=PhantomInk,text=white] at (-4.05,0)
{\large$\mathrm{COI}(\pi)=\mathbb{E}[P]-\underline p$\\[-0.05em]\footnotesize pricing power KPI};
\node<5->[card,fill=PhantomCyan!16,text=PhantomSlate] at (0,0)
{\Large\bfseries$-9{,}014$\\[-0.05em]\footnotesize revenue units\\per +0.1 contamination};
\node<6->[card,fill=PhantomIndigo!12,text=PhantomSlate] at (4.05,0)
{\large$\lim_{N\to\infty}\mathrm{COI}=0$\\[-0.05em]\footnotesize theorem as intuition guide};
\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}%
}
\vspace{0.15em}
\uncover<6->{\scriptsize\textit{The theorem is used to guide intuition: independent reconnaissance pressure can drive realizable prices toward the floor.}}\\[0.2em]
\uncover<7->{\footnotesize\textbf{Implication:} most production dynamic pricing stacks are not yet ready for AI-agent traffic; when quote discovery and purchase split, session-based pricing overestimates willingness to pay.}
\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{enumerate}[<+->]\setlength{\itemsep}{0.45em}
\item Can agent and human sessions be reliably distinguished from behavioral interaction signals alone?
\item How do agents affect prices and revenue that can be generated?
\item Can we make new policies that maintain margin in agent contaminated systems?
\end{enumerate}
\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?}
\vspace{0.35em}
\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}
@@ -188,12 +340,12 @@
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) {Humans + Agents};
\node[box,fill=white] (web) at (2.9,1.45) {Next.js\\storefront};
\node[box,fill=white] (provider) at (5.8,1.45) {Pricing\\provider};
\node[box,fill=white] (redis) at (8.7,1.45) {Redis\\serve layer};
\node[box,fill=PhantomIndigo!10,minimum width=3.1cm] (kafka) at (4.35,-0.15) {Kafka topics\\behavior + price logs};
\node[box,fill=PhantomCyan!10,minimum width=2.8cm] (airflow) at (8.0,-0.15) {Airflow + worker\\batch updates};
\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);
@@ -357,7 +509,7 @@
\vspace{0.25em}
\begin{itemize}
\item<3-> The signed gap $g(\tau')$ is positive when a session is closer to agent behavior.
\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.
@@ -466,7 +618,7 @@
\vspace{0.25em}
\begin{itemize}
\item<2-> Baseline experiments use a query-tax leakage surrogate for tractability.
\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}