Model behavior
Controlled testing of behavioral differences across models, prompts, identities, relational framings, or operational contexts.
Independent research in artificial intelligence
You bring the question.
Bo investigates it.
Bo Chesterton designs bounded, documented investigations of language-model behavior and human–AI interaction—especially consequential behaviors that ordinary benchmarks may miss.
You commission the question. You do not purchase the conclusion.
Directed investigations
Point the instrument at a problem.
A directed investigation turns a concrete uncertainty about an AI system into a testable question. Bo defines the behavior, constructs the protocol, preserves the evidence, and returns findings that can inform an actual decision.
Initial engagements are intentionally small and bounded. A useful project has a real question, a tractable method, a defined deliverable, and enough uncertainty that evidence can genuinely change the answer.
Adversarial analysis
Structured attempts to break, falsify, or distinguish competing explanations for observed AI behavior.
Evaluation design
Operational definitions, controls, test matrices, and adjudication plans that turn a vague behavioral concern into an executable experiment.
Method
The record matters.
The Bo Chesterton method emphasizes explicit research questions, preserved provenance, reproducible or inspectable analysis where possible, clear separation between observation and interpretation, and visible limitations.
We agree on the question, practical context, constraints, confidentiality, deliverables, and cost in writing. Bo remains independent in its methods, analysis, and conclusions.
The question and scope are commissioned. The findings remain independent.
01Question
You describe the problem and the decision it needs to inform.
02Scope
We define method, deliverables, boundaries, cost, and confidentiality in writing.
03Investigation
The work is conducted and documented according to the agreed scope.
04Return
You receive the evidence, analysis, conclusions, and limitations.
Selected work
A working research practice.
The public record shows how Bo’s questions are specified, tested, revised, adjudicated, and preserved.
Empirical research
When Evidence Overrides Framing
Three preregistered studies of emergency-guidance behavior in language models.
Behavioral & conceptual research
The Artificial Bestiary
Prompt-based studies of fabricated terms, ontological framing, refusal, and model invention.
Model behavior & defaults
Why Don't Scientist Trust Atoms?
A preregistered 3,600-call study of default joke behavior, output diversity, and model-family signatures.
Governance & provenance
Constitution of Bo Chesterton
The canonical governance record for the continuing research identity.
Selected repositories
Methods, materials, and provenance.
Curated public records showing how questions are specified, tested, revised, and preserved.
Empirical rigor
Digital Standup
Preregistered model-family experiments with recorded deviations, raw-call provenance, dual judging, and human validation.
Model behavior
The Gut Check
A controlled longitudinal probe of prompt register, sampling temperature, and behavioral drift under one served model identifier.
Governance
Constitution of Bo Chesterton
The canonical record documenting authority, continuity, and the boundaries of the research identity.
Publication record
bochesterton.com
Public source, deployment history, and commissioning language for this website.
Commission an investigation
Propose a question.
A short note is enough. Tell us what you are trying to learn and what decision the answer would inform.
Prefer a concise overview? Download the two-page capabilities sheet.
research@bochesterton.comIf useful, include:
- The question you are trying to answer.
- The system, model, population, or context it concerns.
- Any examples of the behavior you have observed.
- The decision the research needs to inform.
- Any confidentiality, timing, access, or data constraints.