The Calibration Program.
Research program and forthcoming book.

Why we're running this

In the last twenty-four months, every F500 organization has put generative AI into the hands of its leaders. Vendors claim it makes decisions faster, sharper, more data-informed. Critics claim it erodes judgment, inflates confidence, and floods portfolios with low-quality work. Both camps are arguing from anecdote.

The Calibration Program is a focused attempt to look at the actual effect, at two levels of analysis, with public methodology, hand-validated findings, and honest scope. We are not running it because we know the answer. We are running it because nobody does, and because the answer changes how every operator we advise should think about AI rollouts.

Both studies are open. We publish the methodology before the fieldwork, the coding scheme before the data, and the data before the opinions. We publish whatever we find, the flattering patterns and the unflattering ones. This is the work that funds the opinions.

This research becomes a book. The Calibration Program is the foundation for a forthcoming book: The Calibration Debt: How AI Changes Executive Judgment in Enterprise (target publication Q4 2027). The book combines the two core studies with governance frameworks, case studies from participating companies, and practical tools any F500 leader can use to audit their own calibration debt.

How this work actually gets done.

The Calibration Program is led by a single principal researcher , the founder of NS Studio, with frontier LLMs (Claude, GPT-4-class) used as a structured research assistant under direct human validation. We are open about this because it is the methodology, not a workaround.

What the human does. Designs the questions and the coding scheme. Conducts every executive interview personally. Hand-codes the validation subsets that anchor every batch of LLM-assisted analysis. Decides what the patterns mean. Writes every word of every published finding and every book chapter.

What the LLM does. Transcribes interview audio. Reads transcripts and extracts coded statements at scale, prompted with the published coding scheme. Drafts data tables for human review. Helps with the volume work that would otherwise require a research team.

What we publish so you can check the work. Every prompt used. Every coding scheme. Inter-rater agreement statistics between the human and the LLM on the validation subsets. The validation samples themselves. If the LLM and the human disagreed too much on a batch, that batch gets re-done. You can replicate the work, extend it, or disconfirm it.

This is what AI-native research looks like in 2026. The alternative, secretly using LLMs while pretending the work was done by hand, is what most consulting firms are doing right now. We'd rather just say what we're doing.