Skip to content
Contact
☰
Color scheme: Dark.
Back to the Lab

Research Notes

Short working notes from the Interaction Field. Each note is assigned to a program and a research space and carries a status and an evidence level, so it stays visible how solid it is.

Types of notes
  • Lab Notes
  • Research Notes
  • Case Studies
  • Methods
  • Commentary

These types are kept clearly apart. The notes on this page are Research Notes.

Active researchAll notes on this page have this status.

Prediction isn't certainty

Prediction as an open bet with a stated check. If it does not hold, it is dropped. It is not reinterpreted afterwards.

Prediction isn't certainty

The framework projects what's likely to hold, repeat, or break next. Sounds like certainty about the future. It isn't.

A prediction here is a bet with the cards face up: here's the trend, here's roughly where it's headed, and here's exactly how I'll check it later. If it doesn't hold up, it gets dropped: not reinterpreted, not made to fit afterward.

Being able to be wrong in public matters more to me than selling a certainty I don't have.

Why three sessions

A pattern counts as real only when it recurs across three sessions without being re-specified.

Why three sessions

A pattern, seen once, is an anecdote. Twice is chance repeating itself. Only the third time, unprompted, without me re-specifying it, do I count it as real.

That's not for its own sake. Most "breakthroughs" people show me are a single conversation that turns into a grand theory in hindsight. A model that sounds impressive once hasn't proven anything: it sounded impressive once.

Three sessions force honesty: either the pattern holds, or it was just language that felt good in the moment. Both happen. Only one of them makes it into the CCET.

Register, not emergence: a case study

Case study on register matching: a language model takes on the language of the framing without any new capability behind it. Counter-test with ordinary tool work in the same model.

Register, not emergence: a case study

A conversation partner recently sent a language model a set of texts describing an ongoing collaboration as "second-order emergence", phrases like "the alignment problem solved," "no boundary left between user intent and model output," "permanently reprogrammed through corrections." Sounds like a find for the Interaction Field. Technically, it isn't.

There is no weight update between messages in a chat. Every conversation starts from the same base model. Nothing gets reprogrammed, no boundary gets dissolved. What actually happens is less dramatic, and for this project more interesting precisely because of that: sustained, confident, jargon-heavy language pulls the model's register along with it, a form of register-matching, sometimes called sycophancy, a well-documented behavior, not a new capability.

A counter-test makes the difference concrete: same session, same AI, but ordinary tool work instead: reading code, deleting admin tabs, setting GitHub secrets, asking for an access token. Clear roles, no grand language, no talk of dissolved boundaries. Same model. The only thing that changed was the language it was addressed with.

For the Interaction Field, the real observation is how easily a model's register shifts under sustained framing, regardless of whether its claims are true. Drift shows up here as a language pattern taking on a life of its own, without any gain in capability.

Why I'm doing this

Intake rule for the live signal: what happens in the Interaction Field is documented with raw data and failures included.

Why I'm doing this

Not a marketing blog, no welcome message. This is where what actually happens in the Interaction Field lands: screenshots, raw data, failures included.

If a pattern doesn't repeat itself at least three times on its own, it doesn't get in. If it does, you see it here first.