Website concept · October 2026Private mock-up for review

In the Open

Work you can inspect.

Research, methods and public reasoning behind the practice. Each item shows its current stage. Registered means the plan is on public record; it does not mean results exist. In progress means the work is under way and nothing is yet reported.

Research

Reliability, understanding and use.

Scoping reviewRegistered

Empathy variance in large language models

A review of how expressed empathy is measured and reported across commercial language models, including variation within a model and consistency over time. Expressed empathy means what the text communicates, not an internal feeling.

https://doi.org/10.17605/OSF.IO/B8HJ6
ERI Phase 2aRegistered

Variability of expressed empathy across repeated responses

A prospectively registered descriptive study extending the Empathy Reliability Index, focused on variation across repeated large language model responses.

Registered 1 October 2026.

https://doi.org/10.17605/OSF.IO/YRGMJ
EducationIn progress

AI assessment benchmarks for formative feedback

Work on how to assess the consistency of AI-generated feedback across student populations. Formative feedback is intended to help students improve as they learn.

A research theme in development.

MethodsIn development

EvLander™

PEL’s approach to diagnosis and design. It asks how a piece of evidence or content will be received by each of the people who must act on it, and uses the answers to show where a project is likely to be used, where it is likely to matter, and where it is exposed. It draws on implementation science, a record of how evidence has failed to be used in earlier cases, and AI-assisted analysis whose consistency PEL examines in its own research.

EvLander is used in PEL engagements and is still being developed. The method has not been published or independently evaluated. It does not predict regulatory, reimbursement or clinical outcomes.

How we work in the open

Show the reasoning. State the limits.

We share findings, reasoning and outputs so readers can judge their basis. We credit the work we build on.

Client work and internal analytical methods remain confidential.

From research to your draft

Ask the next question.

The language model studies examine consistency in AI-generated text. They do not test whether an evidence plan will support later decisions. Evidence-plan diagnosis and content review are separate services with the limits described on their service pages.