Examples and evidence
Labs
Small examples that make the behavior and limits of AI systems easier to inspect.
Banking study
Follow a customer request through prompt-only and session-bound access, then examine the saved findings and limits.
Evaluation demonstration
Inspect synthetic order records, compare answers and see why correct totals alone do not establish a good answer.
Separate mock permissions demonstration
Inspect simulated publication requests, permissions and inert effects. This is a separate example from the banking study.
Explore inside the articles
Why great AI scores can fail in production
Choose what a fixed delivery rule can know, then check every prediction against the outcome.
What an AI explanation can actually tell you
Move the reference and split duplicate-feature credit while the prediction stays consistent.
Foundation models meet the spreadsheet: where TabPFN fits
Write a fair comparison for TabPFN and familiar models before running an experiment.
What happens between your prompt and the next word?
Inspect visible token pieces, then turn pair counts into probabilities and a continuation.
Mapping meaning: what embeddings reveal and what they hide
Compare full-vector similarity with a lossy map that deliberately drops negation.
Evaluating AI assistants beyond an accuracy score
An answer can get the total right while inventing its explanation. Trace each claim to a record.
Prompt is not permission
Follow a synthetic banking request through the application boundary and inspect the corpus.