What changes when context is filtered before a prompt?
generated canaries test whether permission-aware prompt assembly keeps records from reaching an unauthorized AI.
Research
How much work context is enough to answer a question?
A generated retrieval experiment separates finding one relevant note from recovering the two facts a work question requires.
Can a project keep its identity across apps?
A small entity-linking test measures how often a project in email can be matched to its calendar and file records.
The state of privacy in personal AI, 2026
An evidence-led map of where personal-AI systems keep context, what their published policies permit, and which privacy rules now matter.
Keeping the reasoning with the result
Documentation as part of the work, rather than a reconstruction after it.
Timing a small local context index
A reproducible CPU-bound indexing task shows the measured cost of token postings and project-person edges on generated records.
What a project alias does to inbox search
In a generated inbox, connecting project names to their aliases changes which approval records lexical search can find.
Where the data goes when an AI connects through Composio
A reading of Composio's privacy policy, terms, and technical documentation: credentials, execution logs, model training, downstream providers, and deletion.