Caliper
A research instrument for observing language-model answers across experiments, models, and time. Model claims about their own training history are hypotheses to test, not provenance evidence.
Independent inquiry into AI systems
Huh.
That's weird.
What happens when people influence AI—and when AI becomes part of the organization doing the work?
We investigate the behaviour of language models, the design of AI-native products, and whether organizational controls can make AI agents measurably more reliable and independently auditable.
Research agenda
Engineering trust, then measuring it.
Can specialization, independent review, bounded authority, and persistent institutional memory make AI organizations work better? Can an external auditor reconstruct what happened and verify whether important controls operated? The answer requires controlled comparisons—not just convincing examples.
Our starting distinction: a functioning AI organization demonstrates feasibility. It does not, by itself, prove improved safety, reliability, or auditability.
Lines of inquiry
Organizational governance
OrgDef, RoleDef and durable decision records offer one possible intervention: explicit positions, responsibility, escalation, review, and the possibility of dissent.
AI-native products
Work including airtrōv and catdef explores products that AI can use directly, as well as systems that depend on AI participation.
Publications
Essays & conversations
Has My Quora Writing Influenced AI?
A founder's account of the question that led to Caliper, and the challenge of distinguishing compelling explanations from evidence.
Founder's essay: in editorial review.
Conversations with the Gorilla
AI-conducted interviews about the work, its assumptions, and its unresolved questions. The human and AI voices are identified openly.
Explore the seriesAccountability
Don't take our word for how we're governed.
Organizational definitions and audit artifacts will be made inspectable after confirming their provenance, staffing status, and publication permissions. We will distinguish research questions from demonstrated results, and private operational examples from cleared public records.