rephrasing some things and updating language

docs/static/images/banner.svg

@@ -49,7 +49,7 @@
         <!-- COI Annotation -->
         <line x1="150" y1="150" x2="260" y2="150" stroke="#E37862" stroke-width="2" marker-start="url(#arrow)" marker-end="url(#arrow)"/>
         <text x="310" y="138" font-size="16" fill="#E37862" text-anchor="middle">average information rent</text>
-        <text x="310" y="118" font-family="Georgia" font-style="italic" font-size="22" fill="#E37862" font-weight="bold" text-anchor="middle">COI := E[P] - <tspan text-decoration="underline">p</tspan></text>
+        <text x="310" y="118" font-family="Georgia" font-style="italic" font-size="22" fill="#E37862" font-weight="bold" text-anchor="middle">COI = E[P] - <tspan text-decoration="underline">p</tspan></text>
     </g>
 
     <!-- Bottom: Agent Saturation -->

paper/src/auto/main.el

@@ -16,6 +16,7 @@
     "chapters/04-results"
     "chapters/05-discussion"
     "chapters/06-conclusion"
+    "chapters/acknowledgements"
     "article"
     "art12")
    (LaTeX-add-labels

paper/src/chapters/01-intro.tex

@@ -18,7 +18,7 @@ The current innovation boom in generative artificial intelligence and its applic
 
 The key stakeholders affected by the threat of increasing agent-driven traffic include online businesses and platform operators (especially in bot-heavy sectors like retail, travel, and financial services), their security, fraud, and engineering teams, end users whose accounts and data are exposed and whose experience degrades, regulators and legal stakeholders responding to breaches and fraud, and the attackers or bot operators driving the automation \parencite{imperva_rapid_2025}.
 
-The industry has already seen legal action in cases like Amazon against Perplexity \parencite{ghaffary_amazon_2025}, stemming from the difficulty of identifying traffic from hybrid systems like the Commet browser. This paper explores such systems to better understand what the interaction data looks like and what it means for dynamic pricing and recommendation systems downstream. This observed impact indicates a need for prevention of secondary negative effects on the ``legacy'' systems which power modern revenue sources for many companies. Dynamic pricing algorithms rely on directly translating demand features $q$ to new price assignments $\hat{p}$ across a catalogue of products of size $N$. This opens opportunities to design a \textit{tabula rasa} of digital market mechanisms that will shape the future of commerce in the age of artificial intelligence.
+The industry has already seen legal action in cases like Amazon against Perplexity \parencite{ghaffary_amazon_2025}, stemming from the difficulty of identifying traffic from hybrid systems like the Comet browser. This paper explores such systems to better understand what the interaction data looks like and what it means for dynamic pricing and recommendation systems downstream. This observed impact indicates a need for prevention of secondary negative effects on the ``legacy'' systems which power modern revenue sources for many companies. Dynamic pricing algorithms rely on directly translating demand features $q$ to new price assignments $\hat{p}$ across a catalogue of products of size $N$. This opens opportunities to design a \textit{tabula rasa} of digital market mechanisms that will shape the future of commerce in the age of artificial intelligence.
 
 \subsection{Solution Space Overview}
 Dynamic pricing systems, as presented by \textcite{mueller_low-rank_2019}, often deal with sparse low-rank data of demand signals which, combined with contamination from agents, creates complex interactions that impact pricing. To further complicate the problem, certain commercial settings such as the one presented by \textcite{amjad_censored_2017} must address the true demand of products under censored observations. This provides a formulation for handling demand in our case with multiple kinds of commercial mediators: $\hat{q} \gets q_A + q_H$ where $q_A$ represents the distribution of demand generated by agentic mediators and $q_H$ represents that of true human demand, these are two distinct populations with divergent objective functions.
@@ -64,4 +64,4 @@ Extract final result $r$ from terminal state\;
 \end{algorithm}
 
 
-The previously described goal of distinguishability allows us to formulate a task which entails taking raw interaction data for either actor and creating a composite demand estimate $\hat{q}$. We propose a robust optimization objective defined in our methodology, transforming the pricing problem into a form of Distributionally Robust Optimization \parencite{kuhn_distributionally_2025} where the learner must guard against adversarial contamination in observed demand distributors. In this setting we must learn to make decision that perform under the assumption of not having a single estimated probability distribution but under an ambiguity set of any distribution, of which we have limited information. In our case as stated is a mixture of distributions with a parameter which is unknown and non-stationary.
+The previously described goal of distinguishability allows us to formulate a task which entails taking raw interaction data for either actor and creating a composite demand estimate $\hat{q}$. We propose a robust optimization objective defined in our methodology, transforming the pricing problem into a form of distributionally robust optimization \parencite{kuhn_distributionally_2025} in which the learner guards against adversarial contamination in observed demand \emph{distributions}. The decision rule must perform when the data-generating law is not a single known distribution but any member of an ambiguity set described only partially. Here that law is a mixture whose weight and components need not be stationary.

paper/src/chapters/02-literature-review.tex

@@ -1,15 +1,15 @@
 \section{Literature Review}
 
-To better understand all wedges of the current works, we must start by exploring the nature of agents, agentic computer use and web automation, complementing that with economic reasoning and strategic interaction. The final surface to cover, leads us to data-driven dynamic pricing under uncertainty. The key technical risk is not ``agents buying things'' per se, but agents shaping the behavioral and demand signals that downstream pricing systems consume and depend on. This latter case of agents shopping is currently pending legal action in the case of \textcite{noauthor_amazoncom_2026} which is currently being treated as a violation of the Computer Fraud and Abuse Act. The introduction of these mediating actor entities into economic systems, is further creating a threat of false-name bidding \parencite{yokoo_effect_2004}, which prior research has explored in a trading context. Other research on pseudonyms in dynamic systems, demonstrate whitewashing in AI agents which can ignore defensive mechanisms by re-entry with different identities \parencite{feldman_free-riding_2004}. Dynamic pricing assumes demand proxies are behaviorally meaningful, while bot detection aims at security and access control. The missing bridge is a principled framework for distinguishing non-human reconnaissance from genuine human demand expression and integrating that distinguishability into pricing heuristics without degrading legitimate user experience (in our research tracked by the user-experience index). This gap, is what our contribution aims to address, particularly for the aforementioned stakeholder groups.
+To situate the work we review agents and agentic computer use, web automation, economic reasoning, and strategic interaction, then turn to data-driven dynamic pricing under uncertainty. The main technical risk is not ``agents buying things'' in isolation but agents reshaping the behavioral and demand signals on which downstream pricing depends. Related litigation is already underway---for example \textcite{noauthor_amazoncom_2026} under the Computer Fraud and Abuse Act. Mediating actors also revive classic concerns such as false-name bidding \parencite{yokoo_effect_2004}; pseudonymous re-entry can whitewash reputation and weaken defenses \parencite{feldman_free-riding_2004}. Dynamic pricing assumes demand proxies are behaviorally meaningful, whereas classical bot detection targets security and access control. The gap we target is a principled way to separate non-human reconnaissance from genuine human demand expression and to fold that signal into pricing without degrading legitimate users (we track harm with a user-experience index), for the stakeholders named in the introduction.
 
 \subsection{Agent Taxonomy and Definitions}
 
 An agent in the context of artificial intelligence is generally defined by anything that can reason and act upon observations of its environments (collected through some sensory inputs) and carry out actions through effectors. Moreover, a rational agent is an entity that is capable of perceiving the world around them and taking actions to advance specified goals. This definition by \textcite{russell_artificial_2021} is further developed in an economic context by \textcite{parkes_economic_2015}, suggesting AI research attempts to construct a synthetic \textit{homo economicus}, which may also be termed \textit{machina economicus}.
 A specific class or taxon of this \textit{machina economicus}, the Large Language Model (LLM) agent, is defined as an autonomous system capable of achieving goals and adapting post-training, often without needing explicit code or fundamental model changes \parencite{xia_evaluation-driven_2025}.
 
-We must however acknowledge the current SOTA as presented by OSWORLD simulations by \textcite{xie_osworld_2024} have demonstrated that multi-modal tasks across desktop and web interaction modes, have a top-performing score of only 12.24\% success, whereas humans have a higher 72\% success rate; this is linked to the lack of grounding of these agents and their inability of handling unexpected errors. This weakness matters for this research because it clarifies the near-term threat model: practical exploitation does not require a fully competent ``computer assistant'', only enough automation to perform high-volume reconnaissance actions (search/filter/open product pages, probe availability/price boundaries) that can contaminate behavioral signals. With the expected growth of these capabilities, this threat only becomes more perilous to revenue management systems.
+We must however acknowledge that OSWORLD simulations by \textcite{xie_osworld_2024} report a top success rate of only 12.24\% on multi-modal desktop and web tasks, versus about 72\% for humans, reflecting limited grounding and brittle recovery from unexpected errors. This weakness matters for this research because it clarifies the near-term threat model: practical exploitation does not require a fully competent ``computer assistant'', only enough automation to perform high-volume reconnaissance actions (search/filter/open product pages, probe availability/price boundaries) that can contaminate behavioral signals. With the expected growth of these capabilities, this threat only becomes more perilous to revenue management systems.
 
-We model an agent session as producing some events with lower in-session conversion levels relative to humans, this we state in our assumption that $P(\text{purchase} \vert A) < P(\text{purchase} \vert H)$ but with a potentially higher volatility in $\hat{q}$, which we observe through the look-to-book metrics in our simulation.
+We model agent sessions as producing lower in-session conversion than humans, i.e.\ $P(\text{purchase} \vert A) < P(\text{purchase} \vert H)$, with potentially higher volatility in $\hat{q}$, which we proxy with look-to-book metrics in the simulator.
 
 \subsection{Economic Agents: From Homo Economicus to Machina Economicus}
 
@@ -21,9 +21,9 @@ A HAP (HTTP Agent Profile) protocol has been developed as an internet draft by \
 
 \subsection{Problem Evidence and Market Impact}
 
-The statistical issue of contamination in dynamic pricing systems that observe demand features as a means to update prices has been documented in various previous contexts. The airline industry (which has accounted for 24\% of observed disruptions) has seen malicious activity with a measureable impact on skewing key performance indicators by behavior visible in the look-to-book metrics. Excessive reconnaissance traffic inflates search volume without corresponding completed bookings, thereby skewing demand forecasts and disrupting dynamic pricing models. Demand proxies have also been observed to cause significant threat to inventory management by creating artificial scarcity that distorts the demand-supply relationships in the enterprise model. Censored demand as shown by \textcite{amjad_censored_2017} can also be observed in low-bias demand under-estimation caused by a distortion effect coming from non-human traffic data \parencite{imperva_rapid_2025}.
+Contamination in dynamic pricing systems that observe demand features to update prices appears across several industries. Aviation (about 24\% of observed disruptions in one industry survey) illustrates how malicious or scripted traffic can skew KPIs visible in look-to-book metrics. Excessive reconnaissance traffic inflates search volume without corresponding completed bookings, thereby skewing demand forecasts and disrupting dynamic pricing models. Demand proxies have also been observed to cause significant threat to inventory management by creating artificial scarcity that distorts the demand-supply relationships in the enterprise model. Censored demand as shown by \textcite{amjad_censored_2017} can also be observed in low-bias demand under-estimation caused by a distortion effect coming from non-human traffic data \parencite{imperva_rapid_2025}.
 
-When dynamic pricing algorithms operate on highly contaminated or noisy data, the risk grows significantly in creating inaccurate price inferences. The emergent mitigation driven by un-informed reward and regret signals might lead to price suppression for sales continuity which results in harming margins and resulting in a revenue loss. System that poorly fit undesired behavior might result in price gouging, which calls for strong guardrails while preserving targeted business strategy \parencite{mullapudi_reinforcement_2025}.
+When dynamic pricing algorithms train on highly contaminated or noisy data, mis-inference risk rises. Mis-specified reward and regret signals can push prices down to preserve volume, eroding margins, while misfit to legitimate demand can produce the opposite failure mode; both call for guardrails that preserve commercial intent \parencite{mullapudi_reinforcement_2025}.
 
 
 %Documented instances of agent-driven market disruptions - Quantitative evidence of pricing manipulation - Case studies from affected industries
@@ -31,11 +31,11 @@ When dynamic pricing algorithms operate on highly contaminated or noisy data, th
 \subsection{Theoretical Foundations: Economic Parallels}
 
 
-Early hints of exploration of prices in a standard English auction explored by \textcite{varian_economic_1995} which hints at exploration of prices in a sequential manner, which leads to a marginally different cost to the bidder than the reservation price of the seller. This is a setting in which there is no cost incured by the buyer for their actions or exploring prices in the market. They propose that any agent responsable for the pricing of a good must be imune to dynamic strategies which might extract private information from a market. A key take-away which relates to the Vickery auction mechanism (also called a \textit{direct mechanism}) suggests that not only would defenses against such exploitation be necessary, but the construction of a mechanism in which revelation of the true willingness to pay is the dominant strategy for commerce.
+\textcite{varian_economic_1995} studies sequential exploration of prices in an English auction: the bidder's cost can differ slightly from the seller's reservation price. In that setting the buyer incurs no separate cost for searching or exploring prices. The authors argue that any party \emph{responsible} for pricing must be immune to dynamic strategies that extract private information. The link to the Vickrey (second-price) auction, a \textit{direct mechanism}, is that defenses against exploitation may need to pair with mechanisms in which truthful revelation of willingness to pay is incentive-compatible.
 
 Like in classical revenue-maximizing auctions \parencite{roughgarden_cs364a_2013} we assume that the human actor in our system has a private valuation $v$ which we formally draw from intrinsically defined distributions. The important note here is that the agent proxy does not have a mechanism to convey this private information into the demand data which directly impacts the pricing systems.
 
-The key component of this mediation between agents and commercial platforms lays in the transaction costs related to information gathering and negotiation. As proposed by \textcite{shahidi_coasean_2025} these costs are bound to collapse towards zero (which we demonstrate mathematically), calling for a re-evaluation of the boundaries between firms and markets. As argued by \textcite{coase_nature_1937}, the market participation and time associated with that participation, is critical part of the Coasean transaction cost logic which includes the discovery or relevant pricing within a given market. This process of price discovery without the presence of AI Agents can be time consuming and resource intensive. To build on top of this work we provide a proof of optimal conditions theorised by Coaes as an extension to AI-mediated markets.
+The mediation between agents and commercial platforms turns on transaction costs of information gathering and negotiation. \textcite{shahidi_coasean_2025} argue these costs tend toward zero (we give a complementary formal result in Section~3). \textcite{coase_nature_1937} treats search and participation time as central to Coasean transaction costs, including discovery of relevant prices. Price discovery without AI intermediaries is already costly; we extend the classical Coasean logic to AI-mediated markets.
 
 % Economic foundations: relating the problem to options pricing theory. Cost of Information (COI) concept and its relevance
 
@@ -43,13 +43,13 @@ The key component of this mediation between agents and commercial platforms lays
 
 \subsection{Landscape of Existing Work}
 
-Explorations of the algorithmic collusion by LLMs \parencite{fish_algorithmic_2025} has demonstrated a cross-model tendency of market division with a strong sensitivity to instructions provided in the ``system prompt''. If a dynamic pricing algorithm which is trained to respond to market signals learns to coordinate with competitor agents (or become manipulated by those agents), the market equilibrium is under threat of destabilization. This is particularly true for Q-learning pricing learners as demonstrated by \textcite{calvano_artificial_2018}.
+Work on algorithmic collusion by LLMs \parencite{fish_algorithmic_2025} reports cross-model sensitivity to instructions in the ``system prompt,'' including tendencies toward market division. If a dynamic pricing algorithm which is trained to respond to market signals learns to coordinate with competitor agents (or become manipulated by those agents), the market equilibrium is under threat of destabilization. This is particularly true for Q-learning pricing learners as demonstrated by \textcite{calvano_artificial_2018}.
 
 Our effort to combat contamination stems from research by \textcite{hardt_strategic_2015} on strategic classification, in conjunction with \textcite{liu_contextual_2024} who demonstrate a linear regret if contamination is ignored. The strategic classification adversarial effect comes from an effort to manipulate some representative features used in a learning pipeline, which can result in lower prices on loans or lower prices from dynamic pricing algorithms.
 
 To bridge the gap between detection and robust pricing, we look at work in Distributionally Robust Optimization (DRO). As defined by \textcite{kuhn_wasserstein_2024}, DRO provides a framework for decision-making under ambiguity, where the true data distribution is unknown but lies within a ``Wasserstein ball'' of a target distribution. In our context, the ``ambiguity set'' represents the uncertainty introduced by agentic reconnaissance. By optimizing for the worst-case distribution within this set, pricing mechanisms can become resilient to the distributional shifts such as the ones caused by non-human actors, effectively robustifying the revenue function against the contamination described in our problem statement.
 
-In order to create an environment in which prices can be tested against a demand estimate generated by some behavioral model, we take inspiration from the architecture proposed by \textcite{ie_recsim_2019} in the RecSim platform built for recommendation systems. By modeling the distinct user behavior as partially observable Markov decision processes, we can generate faithful interactions which allow us to generalize, past the constraint which is also present in recommendation systems, of rarely having enough experience with individual actor's interactions for good recommendations without generalization. The key inspiration comes from the user choice modeling which we translate to a user transition model for each distinct actor type (agent or human). We further consider the possibility of modeling our quantitative research platform using dynamic Bayesian networks for the sake of tractability within the system. The contribution or RecSim enables researchers to better understand learning algorithms in fixed environments, a gap we identify as needing to be bridged within the space of dynamic pricing.
+To build an environment where prices face a demand estimate from a behavioral model, we draw on RecSim \parencite{ie_recsim_2019}. Modeling user behavior as partially observable Markov decision processes yields synthetic interaction that generalizes past the usual cold-start limit of per-user data. We translate RecSim-style user choice modeling into per-class transition models (human versus agent). Dynamic Bayesian networks remain a tractability option for the full platform. RecSim's main contribution is a sandbox for recommender learners; we adapt that idea to dynamic pricing under contamination.
 % TODO: mention https://github.com/meta-pytorch/OpenEnv/tree/main/envs/browsergym_env
 
 We also acknowledge the difficulty in similarly affected fields such as authorship, where \textcite{ganie_uncertainty_2025} demonstrate the theoretical limits of the distributional divergence between text authored by a human or large language model. Their approach of computing the divergence between two distributions demonstrates purely theoretically that no classifier can outperform random guessing on their particular task. This is yet another factor to take into consideration when exploring the potential mitigation strategies.

paper/src/chapters/03-methodology.tex

@@ -128,7 +128,7 @@ Since the integrand vanishes as $N \to \infty$ for all $t > \underline{p}$, the
 \end{proof}
 
 
-This result naively proves that standard pricing policies $\pi$ fail to extract surplus in the presence of large-scale agentic search, necessitating a robust counter-mechanism.
+This result implies that standard pricing policies $\pi$ cannot extract the same surplus under large-scale agentic search without additional structure, which motivates the robust control layer below.
 
 % The DRO objective creates a lower bound on COI extraction, effectively guaranteeing a minimum margin even in the presence of adversarial agents. we need to prove this and demonstrate that in a theorem.
 
@@ -137,22 +137,22 @@ This result naively proves that standard pricing policies $\pi$ fail to extract
 
 \subsection{System Architecture: Hybrid Kappa-Lambda Architecture}
 
-In order for our research to have grounding in interactions we built a robust e-commerce web-platform. We initially conducted a survey of the leading platforms of airlines and hotel booking sites to identify the specific interface patterns that effectively manage complex travel data. Our analysis revealed a clear industry standard: while both sectors rely on tabbed service selection and left-sidebar filtering to streamline navigation, they diverge in result presentation: airlines utilize visual date-price bars and multi-step wizards to optimize for logistical transparency, whereas hotel platforms leverage image-led cards and scarcity triggers to drive emotional engagement and urgency. Our web framework defines a highly agnostic boilerplate which can be seeded with any data-modality with an easy-to-tailor pattern, which we leverage to define a \texttt{hotel} and \texttt{airline} mode. Both modes are then individually deployed via an environment level argument which adjusts the proxy routing with a custom middleware inside next.js to render only the desired mode. The purpose of this was to create a baseline adaptable to any use-case or desired commercial application.
+In order for our research to have grounding in interactions we built a robust e-commerce web-platform. We initially conducted a survey of the leading platforms of airlines and hotel booking sites to identify the specific interface patterns that effectively manage complex travel data. Our analysis revealed a clear industry standard: while both sectors rely on tabbed service selection and left-sidebar filtering to streamline navigation, they diverge in result presentation: airlines utilize visual date-price bars and multi-step wizards to optimize for logistical transparency, whereas hotel platforms leverage image-led cards and scarcity triggers to drive emotional engagement and urgency. Our web framework defines a highly agnostic boilerplate which can be seeded with any data-modality with an easy-to-tailor pattern, which we leverage to define a \texttt{hotel} and \texttt{airline} mode. Both modes are then individually deployed via an environment-level argument which adjusts the proxy routing with custom middleware in Next.js to render only the desired mode. The purpose of this was to create a baseline adaptable to any use-case or desired commercial application.
 
-The architecture of this platform begins with the deployed web-apps posting interaction data to our backend which processes them and stores each ingested interaction into a kafka cluster. This serves as our data reservoir tracking and associating each interaction with its session and importantly with which experiment it belongs to. Not only do we track the behavioral interactions, but our pricing provider micro-service, once called by the frontend reports the observed/queried price-product into kafka. This kafka cluster is subscribed to by our pipeline which is configured on a schedule in Airflow, with the possibility of manual trigger. The final stage of the pricing pipeline, submits computed dynamic pricing results into a redis database for quick updates which is then read by the pricing provider and displayed on the webapp. This is a very generic end-to-end mechanism which is applicable to a variety of different e-commerce tasks. We intentionally put emphasis on the development of this infrastructure to establish a reproducible framework for interaction and to minimize any noise.
+The architecture begins with deployed web applications posting interaction data to a backend that stores each record in Apache Kafka. Kafka acts as the reservoir linking sessions to experiments. Behavioral events and, separately, price observations from the pricing-provider microservice (invoked by the frontend) land in Kafka topics. A scheduled Airflow pipeline (with manual triggers) consumes the stream; the final pricing stage writes vectors to Redis for low-latency reads by the provider and display in the client. The pattern is deliberately standard---Kafka for durability and replay, Redis for serving---so the same skeleton applies across e-commerce settings. We invested in this stack to keep runs reproducible and to limit extraneous variance.
 
-\paragraph{Public Web Artifact} We transition the Kappa like architecture of the data collection to a Lambda architecture for actual learning in a surrogate environment. This allows us to move faster on data which is provided and helps us create a feedback loop for production deployment. To support further research in this intersection of fields we release P4P \footnote{\url{https://github.com/velocitatem/p4p}} as a public repository providing the interaction layer of the PHANTOM framework. This provides a configurable storefront which can be tailored to any commercial setting with a standardized session-level event tracking. We document the API adapters or what the framework expects in terms of schemas for pricing providers and log ingestion servicse. The repository is intended for controlled experimentation and method replication rather than production commerce deployment.
+\paragraph{Public Web Artifact} We transition the Kappa like architecture of the data collection to a Lambda architecture for actual learning in a surrogate environment. This allows us to move faster on data which is provided and helps us create a feedback loop for production deployment. To support further research in this intersection of fields we release P4P \footnote{\url{https://github.com/velocitatem/p4p}} as a public repository providing the interaction layer of the PHANTOM framework. This provides a configurable storefront which can be tailored to any commercial setting with a standardized session-level event tracking. We document the API adapters and expected schemas for pricing providers and log ingestion services. The repository is intended for controlled experimentation and method replication rather than production commerce deployment.
 
 \paragraph{Public Dataset} For reproducibility of the behavioral analysis and distinguishability experiments, we also release the interaction dataset used in this thesis as \textit{WhoClickedIt}. The dataset is hosted on Hugging Face \footnote{\url{https://huggingface.co/datasets/velocitatem/whoclickedit}} and is distributed as one flattened event sheet (\texttt{whoclicked.csv}) with explicit labels (\texttt{actor\_type}, \texttt{is\_agent}, and \texttt{record\_type}). The dataset card on that page documents the schema, collection process, and known limitations.
 
 
 \subsubsection{DevOps Principles}
 
-Reproducible results are key to quality research platforms, this is taken into mind when deploying and working with our research platform. From a deployment standpoint the platform can be deployed across a large variety of providers and can be run locally. When developing a new interaction modality apart from the ones that come out of the box, a simple template pattern can be followed. The middleware of the framework is designed to properly render the chosen modality from environmental variables, thus deployment of different or parallel version of the software can be easily parametrized.
+Reproducibility guided deployment choices: the stack runs locally or on common cloud providers. New interaction modalities follow a small template; middleware reads environment variables so parallel deployments (e.g.\ staging versus production-like experiments) differ only in configuration, not in forked codebases.
 
 \subsubsection{Online Dynamic Pricing}
 
-In order to collect data from actors under correct conditions we replicate a naive and simple dynamic pricing algorithm which runs in the background during the experiments.
+To expose participants to state-dependent prices without over-constraining the study, we run a transparent surge--discount heuristic in the background during data collection.
 The dynamic pricing done is handled by a pipeline which computes a demand estimate on a per-product basis of a specific window of the data, defined by the period $T$ which by default is 5 minutes. This dynamic pricing pipeline computes a demand estimate vector $\hat{q} \in \mathbb{R}^N$ by a weighted sum of interactions for each product, it additionally computes a price elasticity vector $\hat{\epsilon}$ in the same dimensions as our demand. The final features matrix is of the size $N \times 2$ which we translate to a new price vector $\hat{p} \in \mathbb{R}^N$.
 
 
@@ -181,7 +181,7 @@ where $p_0 \in \mathbb{R}^N$ is the base price vector (which is seeded into our
 
 We start from a practical constraint: we do not have access to proprietary production data. Because of that, we design our own fictional platform that still represents how commercial platforms work in the real world. The design comes from a survey of hotel and airline websites, where we extracted common interface components and used them as a high-level template for dynamic pricing environments.
 
-The interface is organized as a product catalog where each product belongs to a time-bounded price vector (for example, a daily pricing period). During each period we collect interaction data by instrumenting UI components and predefined action templates that are still customizable. This gives us control without losing realism.
+The interface is organized as a product catalog where each product belongs to a time-bounded price vector (for example, a daily pricing period). During each period we collect interaction data by instrumenting UI components and predefined action templates that are still customizable. That yields controlled variation while keeping the interface credible.
 
 Since users act with motivations, we define a pool of tasks (jobs to be done) and assign tasks randomly to participants.
 We discuss limitations and choices made in this experimental design in Section~\ref{sec:limitations_risks}.
@@ -218,7 +218,7 @@ Our web platform (developed in similar spirit to RecSim \parencite{ie_recsim_201
 
 To speak to realism, user interviews reported that the platform architecture mirrored standard booking interfaces and reduced the cognitive load required to learn the system. One participant described the flow as ``intuitive'' and close to a ``normal'' transaction, suggesting observed behavior was primarily driven by pricing treatment rather than interface novelty.
 
-The dynamic pricing mechanism elicited immediate behavioral adjustments. Participants were sensitive to price volatility: sudden boosts triggered urgency and faster booking attempts, while large listing-to-final discrepancies triggered deeper comparison behavior. This is comforting because the controlled setup still produces commercially relevant interaction data.
+The dynamic pricing mechanism elicited immediate behavioral adjustments. Participants were sensitive to price volatility: sudden boosts triggered urgency and faster booking attempts, while large listing-to-final discrepancies triggered deeper comparison behavior. The responses match what one expects from live commerce: sharp reactions to volatility and to list--checkout gaps, which supports external validity despite the lab setting.
 
 
 \subsubsection{Design of Training Factorial Study}
@@ -264,9 +264,9 @@ v4 & 64 (32 + 32) & us-central2-b & 32 Spot + 32 On-demand \\
 For connections from Madrid, we prioritize the europe-west4 allocation for latency-sensitive runs with the benefit of having the most grouped chips within a single region. This regional grouping is important for the deployment of our Kubernetes cluster which cannot span multiple regions. All sweep metadata, model checkpoints, and reward traces are logged in Weights \& Biases. % TODO: cite this (from bib)
 Hardware specifications are from the official Google Cloud TPU documentation \parencite{noauthor_tpu_2026,noauthor_tpu_2025-1,noauthor_tpu_2025}.
 
-Design of training processes: we build docker image with the fact in mind of different caching over layers in order to most speed up docker re-building and such we place the most volatile steps towards the end of the image building. What is means in practice is that any dependency installations are isolated so edits to source code do no trigger rebuilds. Only if we update our entry point of training a sweep, Docker will also rebuild the source-code copy stage. % TODO: cite Docker best practices on cache-efficient Dockerfile layering.
+Training images follow Docker layer caching: dependency layers are separate from the copy of application source so routine code edits do not invalidate the entire build; only changes to the training entrypoint or dependencies force a full rebuild.
 
-Due to the preemptive nature of the current demand of TPU chips we sttle for running our on demeaned as the primary source of compute. The on demand TPU pod of 32 chips spread across 4 virtual hosts creates a relatively unique parallelization setup. Despite our desire to use a traditional approach of clustering and perhaps deploying SLURM jobs of our sweep agent, the lack of predictability in provisioning each instance of a compute resource makes this an high friction layer we do not want to add.
+TPU capacity is scarce and often preemptible, so we rely primarily on on-demand pods for workloads that must finish without interruption. A typical reservation is a 32-chip pod across four worker VMs; that layout already gives enough parallelism for our sweep driver without adding a separate cluster scheduler. We considered SLURM-style job arrays, but fluctuating provisioning times would have added operational overhead with little benefit for our workload, so orchestration stays in the container and Ray layer described below.
 
 \subsubsection{Interaction Schema}
 
@@ -301,9 +301,7 @@ $\mathcal{A}_{\text{filter}}$ & \texttt{search}, \texttt{filter\_date}, \texttt{
 \end{table}
 
 This partition enables the weight function $\omega$ from Eq.~\ref{eq:qhat} to assign category-specific signal strengths, with $\omega(\mathcal{A}_{\text{cart}}) > \omega(\mathcal{A}_{\text{dwell}}) > \omega(\mathcal{A}_{\text{nav}}) > \omega(\mathcal{A}_{\text{filter}})$ reflecting decreasing commitment.
-It's important to acknowledge that this creates a very blatant assumption in the weighting, and we motivate the scale of each weight by the per-category observed divergence between each behavioral profile.
+The ordering cart $>$ dwell $>$ nav $>$ filter is a deliberate simplification: we set it from early data by ranking categories by KL divergence between human and agent transition rows and then spacing weights in powers of two. The simulator encodes cart $=4.0$, dwell $=2.0$, nav $=1.0$, filter $=0.5$; unknown actions map by prefix to the nearest category.
-In the simulator baseline this order is encoded with a compact fixed scale: cart $=4.0$, dwell $=2.0$, nav $=1.0$, filter $=0.5$. Unknown actions are mapped by prefix heuristics to the nearest category.
-We back this up by saying that each weight was assigned by observing an initial small dataset and computing KL divergence between each interaction type; the ones with the highest divergence receive a proportionately high weight in our demand estimation. From the order which we observe in divergences, we assign a multiple of 2 increase in weight ascending form the lowest weight of $0.5$ in rare filtering operations.
 
 The metadata record $\mu$ varies by action type. For product views, $\mu$ contains the observed price $p_{\text{obs}}$ and product attributes. For dwell events, $\mu$ includes the element text and accumulated hover duration. This heterogeneous structure is captured via a schema-on-read approach in our Kafka ingestion pipeline, where events are validated against type-specific schemas before storage.
 
@@ -320,9 +318,9 @@ To train a robust pricing learner, we need a simulator that can generate realist
 \subsubsection{Ground-Truth Distinguishability}
 Because sessions are collected under controlled experimental conditions where each actor is assigned a known type at the start of the trial, labels $Y_s \in \{H, A\}$ are available as ground truth rather than as the output of a heuristic classifier. We therefore estimate separate transition kernels directly from each labeled partition $\mathcal{D}_H$ and $\mathcal{D}_A$, treating the resulting $\hat{\mathcal{T}}_H$ and $\hat{\mathcal{T}}_A$ as the ground-truth behavioral profiles for each class. We then ask a direct methodological question: are the kernels distinguishable enough to justify downstream pricing control that depends on that distinguishability?
 
-To answer this, we compute per-session KL divergence scores against both class-level centroids. For each session $s$ in either partition, we fit a session-level event transition kernel $\hat{\mathcal{T}}_s$ from that session's trajectory alone, then compute its average KL divergence to the human centroid ($\Delta_{H,s}$) and to the agent centroid ($\Delta_{A,s}$). The per-session distinguishability score is the gap $\Delta_{H,s} - \Delta_{A,s}$: a negative value indicates proximity to human behavior, a positive value indicates proximity to agent behavior. The reason behind KL divergence for profile analysis is grounded in its nature and tailored characteristics for probability distributions.
+For each session $s$ we fit a session-level transition kernel $\hat{\mathcal{T}}_s$, then average KL divergence to the human centroid ($\Delta_{H,s}$) and to the agent centroid ($\Delta_{A,s}$). The distinguishability score is the gap $\Delta_{H,s} - \Delta_{A,s}$ (negative $\approx$ human-like, positive $\approx$ agent-like). KL is used because it compares full categorical rows, not single features.
 
-The normality assumption cannot be made for KL divergence distributions, which are right-skewed and bounded below by zero, so we do not use a Student's $t$-test. Instead we apply a Mann-Whitney $U$ test \parencite{mann_test_1947} on the per-session gap scores between the two groups. The Mann-Whitney test is a rank-based nonparametric test that compares the stochastic ordering of two independent samples without distributional assumptions, making it appropriate for small samples drawn from skewed populations. We report $U$, the exact two-sided $p$-value, and group-level descriptive statistics for the gap scores.
+Gap scores are skewed and nonnegative, so we test cohort differences with a Mann--Whitney $U$ test \parencite{mann_test_1947} rather than a $t$-test. We report $U$, the two-sided $p$-value, and descriptive statistics for each group.
 
 \begin{definition}[Kullback-Leibler Divergence for Transition Distributions]
 Let $P_e$ and $Q_e$ be categorical distributions over destination states following event $e$, derived from human and agent trajectories respectively. The KL divergence between these distributions is:
@@ -331,7 +329,7 @@ Let $P_e$ and $Q_e$ be categorical distributions over destination states followi
 \end{equation}
 where $\mathcal{S}_e$ denotes the set of destination events that follow $e$ in the human trajectories.
 \end{definition}
-The asymmetry of KL divergence is a point we leverage to natively create divergence from human behavior, to gather signal of the dissimilarity from human-like interactions.
+We exploit KL asymmetry so that ``distance from human-like'' is explicit in the score, not only distance from agents.
 
 To obtain this statistic, we aggregate transitions by triggering event $e$ and treat normalized outgoing probabilities as categorical distributions $P_e$ (human) and $Q_e$ (agent). We intersect shared event labels, then accumulate log-ratio contributions over shared destinations. Large contributions, including near-zero $Q_e(k)$ cases, identify transitions where one actor class is difficult to mimic.
 
@@ -382,27 +380,27 @@ Because contamination level $\alpha$ and demand shift are non-stationary online,
 
 From these two divergences we define the gap score:
 \begin{equation}
-g(\tau') := \Delta_H(\tau') - \Delta_A(\tau').
+g(\tau') = \Delta_H(\tau') - \Delta_A(\tau').
 \end{equation}
 Positive values indicate trajectories farther from the human centroid and closer to the agent centroid.
 
 We map this gap to a weak agent probability using a temperature-controlled logistic map:
 \begin{equation}
-f(\tau') := P(Y=A\mid\tau') = \operatorname{softmax}(-\Delta_A,-\Delta_H)_A = \sigma\left(\frac{\Delta_H-\Delta_A}{T}\right), \quad T>0.
+f(\tau') = P(Y=A\mid\tau') = \operatorname{softmax}(-\Delta_A,-\Delta_H)_A = \sigma\left(\frac{\Delta_H-\Delta_A}{T}\right), \quad T>0.
 \end{equation}
 The session-level control signal injected into pricing is then
 \begin{equation}
-\hat{\alpha}(\tau') := f(\tau').
+\hat{\alpha}(\tau') = f(\tau').
 \end{equation}
 
 This turns distinguishability into an operational control input in the engine. On a per-customer or use-case basis, a similar data collection and fitting process should be repeated to obtain domain-specific behavior kernels.
 
-In implementation, we maintain an alternating game-history stack (our \textit{Limbo} stack) and execute it explicitly every epoch with exactly two transitions: first the platform publishes a price vector (leader move), then the market responds with trajectory-derived demand (follower move).
+In implementation we keep an alternating game-history buffer and advance it each epoch with two transitions: the platform publishes a price vector (leader move), then the environment returns trajectory-derived demand (follower move). The codebase names this structure \textit{Limbo}; the appendix lists it under the same label for readers who inspect the repository.
 
 To avoid notation drift, we separate two COI objects used for different purposes:
 \begin{align}
-\text{COI}_{\text{level}}(\pi) &:= \mathbb{E}[P]-\underline{p} \quad \text{(global reporting KPI)} \\
+\text{COI}_{\text{level}}(\pi) &= \mathbb{E}[P]-\underline{p} \quad \text{(global reporting KPI)} \\
-\text{COI}_{\text{leak}}(p,\tau') &:= f(\tau')\cdot \text{InfoValue}(p,\tau') \quad \text{(local control penalty)}
+\text{COI}_{\text{leak}}(p,\tau') &= f(\tau')\cdot \text{InfoValue}(p,\tau') \quad \text{(local control penalty)}
 \end{align}
 where $\text{COI}_{\text{level}}$ is evaluated at policy level and $\text{COI}_{\text{leak}}$ is evaluated per observed quote during training. We connect local leakage to expected global erosion with the operational assumption
 \begin{equation}
@@ -485,7 +483,7 @@ In practice, we parameterize this with a session-level leakage term:
 \end{equation}
 where $f(\tau')$ is the weak agent probability and $\text{InfoValue}$ is implemented either as a constant query-tax surrogate or as a revelation surrogate $-\log\pi(p\mid\tau')$.
 
-To make the intuition of our $\max \min$ easier in connection to the COI term which we are subtracting, we introduce the strongest possible penalization and try to maximize only for the worst case scenario in which the leakage is extremely high and that negation sends a signal to pick the candidate of the hardest problem.
+The inner minimization selects the contamination candidate that makes the penalized reward smallest, so the outer policy update faces the worst plausible leakage scenario inside the ambiguity set rather than an average case.
 
 For the baseline engine reported here, we intentionally use the constant query-tax surrogate to keep the mechanism minimal:
 \begin{equation}
@@ -547,13 +545,13 @@ We now present the complete pricing mechanism that integrates the behavioral dis
 \end{algorithm}
 
 
-The algorithm operates in discrete epochs indexed by $t$. At each epoch, the platform applies one discrete multiplicative price action, the environment samples a batch of sessions, and demand is recomputed from weighted events. Robustness is implemented as an inner minimization over a small local grid of contamination candidates around nominal $\alpha_0$, matching the current engine implementation. The history buffer $\mathcal{L}$ (what we are calling the ``Limbo'' stack in our implementation) enforces the alternating Stackelberg structure by preserving the temporal sequence of price publications and demand observations.
+The algorithm operates in discrete epochs indexed by $t$. At each epoch, the platform applies one discrete multiplicative price action, the environment samples a batch of sessions, and demand is recomputed from weighted events. Robustness is implemented as an inner minimization over a small local grid of contamination candidates around nominal $\alpha_0$, matching the current engine implementation. The history buffer $\mathcal{L}$ enforces the alternating Stackelberg structure by preserving the temporal sequence of price publications and demand observations.
 
 %The defensive price update in Line 24 implements contamination-aware margin shrinkage: as estimated contamination $\hat{\alpha}_t$ rises, the margin $(p^{\mathrm{ref}} - c)$ is reduced by factor $\kappa\in[0,1]$, with projection $\Pi_{\mathcal{P}}$ ensuring feasibility. In subsequent experiments this heuristic rule is replaced by DR-RL policy $\pi^*$ from Eq.~\ref{eq:robust_policy}.
 
 \subsection{Parallelization Strategy}
 
-To avoid preemption of compute mid-training we settle on using a v4 generation, 40 chip compute node with 5 parallel workers. The login node creates an orchestration node with Ray \parencite{moritz_ray_2018} and we distribute ray compute nodes per each other worker.
+To reduce mid-job preemption we standardize on a TPU v4 allocation with 40 chips and five workers. A head process launches Ray \parencite{moritz_ray_2018} and attaches workers across the remaining hosts.
 
 \subsubsection{Computational Cost Analysis of the Simulation Step}
 The per-step cost of Algorithm~\ref{alg:phantom_loop_clean} is not uniform across its components. To inform hardware provisioning and to identify where algorithmic improvements are most impactful, we profile the hot path of the engine using Python's \texttt{cProfile} instrumentation over 20 environment steps under two configurations: a baseline with the robustness inner loop disabled ($K=1$, $\epsilon_\alpha=0$) and a standard robust setting ($K=5$, $\epsilon_\alpha=0.2$). Both runs use $M=10$ sessions per market call and $N=3$ products.

paper/src/chapters/05-discussion.tex

@@ -4,18 +4,16 @@
 
 \subsection{Transition to Agentic Market Microstructure}
 
-Our analysis of the interaction dynamics between the platform and non-human actors suggests that the current static pricing models are insufficient for an agent-mediated economy. If we assume a transition toward a direct revelation mechanism, where actors must reveal their true valuation of a good through bidding dynamics, we inevitably introduce significant stochasticity into the pricing system. Unlike traditional e-commerce where prices are relatively sticky, such a mechanism implies a high volatility characteristic of financial equity markets (without the fungability however).
+Our analysis of interaction dynamics between the platform and non-human actors suggests that static posted-price models are a weak match for an economy in which software agents mediate search and purchase. If one pushes toward direct-revelation or auction-like pricing, volatility rises: prices behave more like traded claims than like sticky retail quotes, though without the fungibility of securities.
-
-However, ecommerce commodities differ fundamentally from financial securities: they possess a hard floor defined by unit economics and reservation prices. The market might react enthusiastically to an iPhone priced at \$1, such a transaction is not permissible. The platform must establish an initial valuation anchor ($P_{0}$) defined by the marginal cost plus a target margin, around which the market price is permitted to fluctuate. We float the introduction of GenAI Agents as Institutional Market Makers. As the arms race for greater autonomy of agnetic systems grows, the commercial viability of AI agents has the potential to disseminate into every-day users directly interacting with them rather than e-commerce platforms. This is also under the assumption of expected transactional capabilities being given to AI Agents.
-
 
+E-commerce goods differ from financial assets in a hard way: unit economics and reservation values set a floor. The market might ``want'' an iPhone at \$1; the platform cannot honor that. Pricing therefore needs an anchor $P_{0}$ (cost plus target margin) around which offers may move. In that setting, large language model (LLM) agents resemble institutional liquidity providers: they quote, probe, and clear subsets of flow. As autonomy of agentic systems increases, end users may delegate browsing and checkout to assistants rather than to retailer sites directly, which shifts where demand signals originate. The scenario presumes agents eventually hold delegated payment authority; until then, our results bound a near-term reconnaissance-heavy regime.
 
 \subsection{Risk Assessment and Limitations}
 \label{sec:limitations_risks}
 
-This technology does not come without a more bitter side, ethical concerns do arise from the idea of deploying black-box like solutions to set prices based on a behavioral attributes. Approaches like universal behavioral profile modeling (UBPM) used in recommendation systems is very broadly utilized.
+Behavior-based pricing raises predictable ethics questions when models are opaque: a behavioral profile can become a basis for price discrimination or exclusion if deployed without governance. Universal behavioral profile modeling (UBPM) in recommendation already shows how fine-grained traces enable strong personalization; the same machinery applied to prices needs guardrails.
 
-In our experimental setup we randomly assign each user to a platform and, within that platform, assign them to a task. Figure~\ref{fig:exp_design_tree} summarizes this design decision tree.
+In our experiments participants are randomized to platform mode and task. Figure~\ref{fig:exp_design_tree} summarizes the assignment tree.
 
 \begin{figure}[ht]
   \centering
@@ -26,8 +24,8 @@ In our experimental setup we randomly assign each user to a platform and, within
   \label{fig:exp_design_tree}
 \end{figure}
 
-Although our participant sample size is somewhat low for humans, we do a one-to-one balance of human-to-agent experimental sessions. This way we are observing a uniform distribution of participation from each participating side. Our sample size of participants might look scarce, but each participant generates a rich amount of data, with a totality of 3,874 rows of data.
+The human sample is small but each session is long-form; we balance human and agent sessions one-to-one so cohorts are comparable despite different population sizes. The row-level dataset still contains thousands of events.
 
-With a system like this there is potential for strong drift given the rapid advance of agentic systems and user preference. Our intent behind adding the UX term into the reward shaping process was to further address the risk of degraded user experience. Looking deeper at the underlying methodology, reinforcement learning does not come without it's complications such as reward hacking and often the lack of intepretability which is quite critical in systems that have a strong impact on the revenue of a company.
+Rapid change in agent capabilities and user expectations induces model drift; the UX term in reward shaping was included partly to penalize policies that sacrifice legitimate users for short-run revenue. Reinforcement learning adds its own risks---reward hacking and limited interpretability---which matter when policies touch live revenue; deployment would require monitoring and constraints beyond what we exercised here.
 
 % \subsection{Implications of Findings} Interpretation of results and altenrative scenarios with broader market implications.

paper/src/chapters/06-conclusion.tex

@@ -1,26 +1,22 @@
 \section{Conclusion}
 
-Our research has explored how reinforcement learning works within pricing systems and environments which are substantially disrupted by an adversarial participant. Our findings include the optimization for our newly introduced metrics.
+This thesis examined reinforcement-learning policies for dynamic pricing when a fraction of traffic is orchestrated by non-human agents intent on extracting information before purchase. We introduced COI-oriented metrics, a behavioral distinguishability layer, and a distributionally robust training loop; empirical runs show where robustness helps and where it must be tuned.
 
 \subsection{Summary of contributions}
-The contribution was not without the advice of many experienced experts in the field. We thank Marco Casalaina VP Products, Core AI and AI Futurist at Microsoft for the initial critical discussion on the topic of dynamic pricing systems and the spark which has lead to this work. Eugene Bykovets, PhD pointing out the parallels in blockchain systems and the complexity of anonymous interaction and understanding of intent. Importantly, the contributions of Alberto Martín Izquierdo, my academic advisor for the support over and for taking on the challenge of this ambitious work. Many breakthroughs were thanks to numerous discussions with my peers on the topics covered here.
-A thanks to the head of innovation at Amadeus for insight into the industry split on the topic of collapsing margins. Finally we acknowledge the power and use of generative AI technologies for in depth research, rapid prototyping and surfacing of key topics and niches.
-
-Now we very explicitly mention what we contribute in this paper:
 \begin{itemize}
-    \item TPU-accelerated parallelization of the behavioral simulation and reinforcement learning pipeline, making large-scale factorial sweeps tractable.
+    \item TPU-accelerated parallelization of the behavioral simulation and reinforcement learning pipeline, making large factorial sweeps tractable.
     \item Formalization of non-human transaction orchestration in e-commerce as a distinct source of contamination in dynamic pricing systems.
-    \item Definition of the Cost of Information (COI) as a mechanism-level quantity for pricing power, together with a theorem showing its erosion under increasing agent saturation.
+    \item Definition of the Cost of Information (COI) as a mechanism-level quantity for pricing power, together with a theorem on its erosion under increasing agent saturation.
-    \item Design and implementation of a controlled e-commerce research platform, built on a hybrid Kappa-Lambda architecture, for collecting and replaying high-fidelity interaction trajectories.
+    \item Design and implementation of a controlled e-commerce research platform on a hybrid Kappa--Lambda architecture for collecting and replaying high-fidelity interaction trajectories.
-    \item Construction and empirical validation of a behavioral distinguishability framework that distinguishes human and agent sessions from interaction signals alone using transition kernels and KL-based divergence.
+    \item Construction and empirical validation of a behavioral distinguishability framework that separates human and agent sessions from interaction signals alone using transition kernels and KL-based divergence.
-    \item Development of a generative contamination mechanism that injects learned agent behavior into the pricing environment for controlled robustness experiments.
+    \item A generative contamination mechanism that injects learned agent behavior into the pricing environment for controlled robustness experiments.
-    \item Translation of behavioral distinguishability into a defensive pricing mechanism through a distributionally robust reinforcement learning formulation of pricing under non-stationary contamination.
+    \item Translation of distinguishability scores into defensive pricing via distributionally robust reinforcement learning under non-stationary contamination.
-    \item Empirical evidence that agent contamination reduces revenue and that robustness is condition-dependent, requiring explicit calibration rather than a one-size-fits-all penalty.
+    \item Evidence that contamination depresses revenue and that robustness gains are regime-dependent, so penalties and radii need calibration rather than a single default.
-    \item Release of a reusable public experimental artifact for reproducing and extending research on dynamic pricing under agent-mediated traffic.
+    \item Release of a public experimental artifact (code and dataset) for reproducing and extending work on agent-mediated traffic.
 \end{itemize}
 
-\subsection{Future Works and Next Steps}
+\subsection{Limitations and future work}
 
-In our effort to tackle this work we initiated a set of constraints which we hope to relax in future iterations and hope that some of these will be addressed in industry. First of these constraints is the weighting of different actions within the demand estimation, which we would ideally find through learned methodology. Next, assumption of perfect alternating turns between the platform and the market calls for a fixed length non-strictly alternating state definition with a history of actions to possibly allow for the development of multi agentic or multi platform simulation. In our simulation we also make assumptions of non-perishable supply of items, which creates the biggest sim-to-real gap in our system. We also would like to further remove intra-session stationary nature of the contamination parameter to further create high-fidelity non-stationarity within a single evaluation window.
+Several constraints are intentional and could be relaxed later. Action weights in the demand proxy are hand-set; learning them from data is an obvious next step. The Stackelberg interface assumes a clean alternation between platform move and market response; richer histories (multi-agent, multi-platform) would need a less rigid state definition. Non-perishable catalog supply in the simulator widens the sim-to-real gap for inventory-constrained domains. Within-session contamination is modeled as stable; time-varying $\alpha$ inside a session would better match some attack patterns.
 
-For deployment of this it is advised to collect a higher sample size of human baselines and to complement this with the simulated agentic sessions and to mind the matrix scaling for very large catalog sizes.
+Before any deployment, human baselines should grow beyond the convenience sample used here, and catalog scaling laws should be re-checked when transition matrices grow with SKU count. For the deployment of this methodology presented in our work.

paper/src/chapters/acknowledgements.tex

@@ -1,3 +1,7 @@
-\section{Acknowledgements}
+\section*{Acknowledgements}
 
-Eugene Bykovets, PhD - ETH
+This research was supported by the TPU Research Cloud program, which provided access to Google Cloud Tensor Processing Unit (TPU) accelerators, including TPU v4, v5e, and v6e.
+
+I am grateful to Marco Casalaina (VP of Product, Core AI, Microsoft) for an early conversation on dynamic pricing that helped frame the problem. Eugene Bykovets (Ph.D.) pointed out useful parallels with blockchain systems and the difficulty of inferring intent under pseudonymity. Alberto Mart\'{i}n Izquierdo supervised this work and accepted an unusually wide brief. Several peers contributed through discussion of the topics covered here. The head of innovation at Amadeus offered industry perspective on margin compression under automation.
+
+Generative tools were used for literature search, prototyping, and drafting support; all claims, experiments, and final wording remain the author's responsibility.

paper/src/main.tex

@@ -18,14 +18,18 @@
 \end{titlepage}
 
 \begin{abstract}
-With accelerated growth of Large Language Model agents in e-commerce, a novel adversarial dynamic to digital markets emerges. This paper addresses the vulnerability of dynamic pricing systems to AI intermediaries that decouple the information gather stages from the transaction execution. By conducting reconnaissance in isolated sessions, agents circumvent the ``Cost of Information'' (COI) defined as the accumulated price premium typically via demand expression estimators. We formally define this phenomenon and derive the Cost of Information Theorem, proving that as the saturation of independent, utility-maximizing agents increases, the platform's ability to sustain a COI converges to zero, rendering standard dynamic pricing mechanisms incentive-incompatible. To respond to this threat, we propose a defensive framework which integrates behavioral economics with Adversarially Distributionally Robust Optimization (DRO). We introduce a custom e-commerce research platform built on a hybrid Kappa-Lambda architecture, designed to capture and simulate high-fidelity controlled interaction trajectories. We further demonstrate through modeling that human and agent behaviors exhibit distinct transition probability kernels, enabling the construction of discriminative models based on Kullback-Leibler divergence. These behavioral signals serve as inputs for a Distributionally Robust Reinforcement Learning (DR-RL) agent. We formulate the pricing problem as a Stackelberg game where the learner optimizes against an ambiguity set of demand distributions defined by the Wasserstein distance. This approach allows the pricing policy to remain robust against non-stationary contamination without overfitting to deterministic demand curves. Extensive TPU-accelerated factorial training demonstrates that while agent contamination causally reduces short-term revenue, our robust mechanism successfully preserves COI margin integrity and market equilibrium, particularly under higher contamination ratios and larger catalog sizes. Additionally, we show that integrating a balanced UX penalty drastically reduces supra-competitive pricing tendencies, minimizing degradation to the legitimate human user experience. Finally, we release our custom interaction framework and dataset as public artifacts to support future research in agent-mediated traffic.
+\noindent
+Large language model (LLM) agents are spreading in e-commerce; one consequence is intermediaries that can separate information gathering from transaction execution. This thesis studies dynamic pricing when agents reconnoitre in isolated sessions and thereby weaken the \emph{Cost of Information} (COI), the premium platforms typically extract once demand signals are expressed.
+
+We formalize the phenomenon and prove a Cost of Information theorem: as independent, utility-maximizing agents saturate price queries, the platform's sustainable COI goes to zero, so ordinary dynamic pricing is incentive-incompatible in the limit.
+
+The defensive design combines behavioral signals with distributionally robust optimization (DRO). We implement a controlled storefront on a hybrid Kappa--Lambda architecture and show that human and agent sessions induce different transition kernels. Kullback--Leibler divergence to class prototypes yields session scores that feed a distributionally robust reinforcement learning (DR-RL) policy, posed as a Stackelberg game with a Wasserstein ambiguity set over demand so the learner does not collapse to a single empirical demand curve under shifting contamination.
+
+Factorial training on TPUs shows the expected short-run revenue hit from contamination and that the robust objective recovers COI and equilibrium structure in harder regimes (higher contamination, larger catalogs), accounting for UX to prevent supra-competitive pricing. Code and an interaction dataset are released for work on agent-mediated traffic.
 \end{abstract}
 
 \noindent\textbf{Keywords:} Dynamic Pricing, LLM Agents, Adversarial Machine Learning, E-commerce, Behavioral Detection, Reinforcement Learning
 
-\vspace{1em}
-\noindent\textbf{Acknowledgments:} This research was supported by the TPU Research Cloud program, which provided access to Google Cloud TPU accelerators (including TPU v4, v5e, and v6e).
-
 \vspace{0.5em}
 \noindent\textbf{Project page:} \url{https://velocitatem.github.io/PHANTOM/}
 
@@ -37,6 +41,8 @@ With accelerated growth of Large Language Model agents in e-commerce, a novel ad
 \input{chapters/05-discussion}
 \input{chapters/06-conclusion}
 
+\input{chapters/acknowledgements}
+
 \printbibliography
 
 \clearpage