Presentations and walkthroughs on research design, computational methods, measurement, and inference. These talks examine how studies are constructed, how research instruments are tested, and how methodological choices affect what we can conclude.
Several demonstrate AI-enabled research systems, including simulated experiments, automated study audits, and multi-agent testing environments.
Research talk · Sandia National Laboratories · Center for Computing Research · AI & Data Science Research · September 2026
0:00 / 0:00
Full spoken text
A fifteen-minute talk for the Center for Computing Research at Sandia National Laboratories. I treat credibility as an evidentiary question rather than a fixed property of a model: what evidence supports relying on a system, for what, and under what conditions? I use an experimental system I built for studying interacting AI agents to examine how network structure, communication, failure, model capability, and emergent organization change what those systems do and what we can infer from their behavior.
An illustrated talk on collective behavior in networks of communicating AI agents. I begin with individual agents and build toward networks in which they hold different information, exchange messages, and reach collective conclusions. From there, I examine how communication can improve judgment or reinforce errors, how information changes as messages are relayed and rewritten, and how network structure shapes coordination, performance, and failure. Drawing on network science, research on human groups, and experiments with language-model agents, I work through what these systems can accomplish, what the evidence establishes, and what remains uncertain. The accompanying text develops the underlying concepts, mechanisms, and research alongside the animations.
Research talk · Collective inference & AI reliability · Proposed to DARPA's Defense Sciences Office · November 2026
An animated research talk examining how networks of language-model agents can reach confidently incorrect judgments even when the evidence distributed across their members supports the correct answer. Drawing on controlled experiments, I examine how communication content and network structure affect collective accuracy, why outcomes diverge across trials, how agents respond to others’ judgments, and how a single planted verdict can shift a group's final belief. These findings motivate Collective Evidence Assurance, a research program proposed for DARPA’s Defense Sciences Office, which would develop measures of how agents weigh communicated evidence and test whether early signs of failure can predict collective error across previously untested networks, models, and tasks. The presentation distinguishes established experimental findings from proposed mechanisms and future research, with the central question being whether collective failures can be anticipated and prevented beyond the settings in which they were first observed.
Auditing a field-experiment proposal with an AI agent
Research walkthrough · Independent work · September 2026
A walkthrough of me reasoning through a proposed field experiment with Claude Code before launch. I use the session to reconstruct what the proposal is claiming, test what the design can identify, work through alternative explanations, and decide what would need to change before I would fund it. My commentary on the right tracks the reasoning as it develops.
Building and simulating a behavioral experiment with an AI agent
Research walkthrough · Independent work · September 2026
A walkthrough of me designing a field experiment with Claude Code, starting from a client brief, and rehearsing it before launch. I use the agent to review the evidence, build the experiment as protocol and code, and run it hundreds of times in synthetic worlds where the true effect is known, to find where it could mislead a decision, repair it, and retest the repair. My commentary on the right tracks the reasoning, what I decide, and how the agent's work is checked.
Building and validating a multi-agent research system with an AI agent
Research walkthrough · Independent work · September 2026
A walkthrough of me building the instrument for a computational experiment with Claude Code, starting from a short research brief. The system sends the same information through differently structured networks of language-model agents and records how it changes at each step. I use the agent to turn the theory into testable graphs, routing, provenance and measures, and to check each piece before trusting it, stopping at a completed, verified build rather than a study run. My commentary on the right keeps my research decisions, Claude's development work, and the experimental agents' behavior distinct.
Building and rehearsing an employee survey and psychometric instrument with an AI agent
Research walkthrough · Independent work · September 2026
A walkthrough of me turning a leadership request for an employee survey into a measurement instrument with Claude Code, and rehearsing it before any employee answers it. I use the agent to define the constructs, research prior measures, build the item bank and survey, and test the analysis against synthetic employees whose true experiences and response styles are known. The rehearsal ends at a pilot-ready instrument, with cognitive pretesting and a real pilot as the next steps. My commentary on the right tracks the measurement reasoning and what each result changes for the decision.
Reproducing and extending Schelling's segregation model with an AI agent
Research walkthrough · Agent-based modeling · Independent work · October 2026
A recorded Claude Code session reconstructs and extends Thomas Schelling's agent-based model of segregation, examining how an AI development agent specifies behavioral rules, tests an implementation, and evaluates the results. The protocol was prepared in a separate AI-assisted planning session. The first reproduction fails its predetermined test. A revised measure is validated before larger-board experiments reproduce the general relationship between individual preferences and aggregate segregation. Further tests reveal how the outcome depends on implementation choices, imperfect perception, and the structure of relevant neighborhoods. The commentary follows the decisions, failed predictions, corrections, and remaining uncertainties, distinguishing what the model establishes from what it cannot explain about residential segregation.
An illustrated talk on organizational network analysis, worked through Northline Systems, a fictional company built from synthetic data. I start with what a network is and what can move through its ties. From there, I work through measurement, formal and working structure, cohesion, and centrality. I then turn to knowledge, vulnerability, overload, message content, and access, before looking at change over time and a proposed reorganization. The final sections cover interventions, evidence and privacy, and networks that include AI systems. The text sits alongside the deck, so you can read through it at your own pace.
An interactive walkthrough using Org Signal, a browser-based environment I built for social network analysis
Teaching walkthrough · October 2026
A walkthrough of social network analysis in Org Signal, the browser-based environment I built for teaching and conducting network analysis. The software runs on the left, carrying out each step of the analysis, while the narration explains what is being done, why, and what the resulting evidence can and cannot support.
The walkthrough moves from networks drawn by hand, collected through an interview, and surveyed across a group to synthetic organizations with known structure, message records, classic published networks, and perceived networks. Along the way, it treats tie construction as an analytic choice, compares observed patterns with random-network baselines, examines sensitivity through resampling and alternative constructions, and ends with what is required to make a defensible claim from network data.
Building a multi-agent application system with an AI agent
Walkthrough · Multi-agent workflow · Independent work · October 2026
A recorded Claude Code session builds a multi-agent system for job applications around a fictional candidate and fictional employers. Scout agents screen a board of postings, Claude writes an editorial brief for the selected role, an author agent drafts a cover letter and résumé, and three reviewers with different questions and different source files examine the drafts. After the first revision, I ask which questions those reviews left unexamined, and three more reviewers assess fit to the role, the selection of evidence, and the documents' layout. The session ends with a handoff to the candidate, who decides what is accurate, what may be disclosed, and whether to apply. A Materials page holds the prompts, reviews, triage decisions, and every version of the documents.
Walkthrough · Qualitative text analysis · Independent work · October 2026
A recorded Claude Code session carries out twenty analyses of 13,509 public Hacker News comments in which developers describe their own use of AI coding tools between 2021 and 2026. The session samples 2,534 comments, pilots and revises a coding prompt, checks every returned quotation against its source, and builds categories from the data. The analyses range from individual concept maps and metaphor families to paired severity judgments, negative cases, and a comparison of three theoretical lenses. The commentary distinguishes results that rest on a script, on the session's reading of a sample, and on the model's judgment alone. No person coded a sample, so agreement between the model's coding and a human coder's is unmeasured.