Small software studio
Studio note · 30 August 2026
We make small tools for curiosity, reasoning, learning, and following an idea past the first answer.
Most of them begin as a question in Noesis. A few survive long enough that we become willing to maintain them.
In the workshop
Four things have most of our attention right now.
Live learning experiment
A symbolic-logic tutor that makes the steps visible. The current build teaches propositional logic with short lessons, graded practice, spaced review, and progress derived from actual work rather than points.
The question now is not how to make it more addictive. It is how to make returning feel good without making the learner better at the interface than at logic.
Try the live build ↗ Public source ↗Content research
Questions that ask you to commit before the explanation arrives — then follow what your answers imply without pretending disagreement is a bug.
We are currently doing the unglamorous part: hostile reviews of the questions themselves, looking for hidden correct answers, cheap tells, and ways a sophisticated participant could game the exercise instead of answering honestly. The same work feeds Doxograph, our attempt to map beliefs without flattening uncertainty into a score.
Playable prototype
A social roleplaying experiment where characters have their own memories, beliefs, motives, and blind spots — while the engine, not the model, remains responsible for what actually happened.
The current prototype is deliberately small: a handful of people, asymmetric information, ordinary language, and enough memory to let trust, suspicion, and manipulation persist between scenes.
Infrastructure · migrating
The machinery behind the machinery: a coordinator for AI research, review, and implementation work built around explicit human authority, independent challenge, and evidence that survives the chat that produced it.
We are extracting it from Noesis into a standalone multi-repository system without letting “autonomous” quietly become “has the keys to everything.”
What the research is changing
We use independent research lanes to disagree usefully. Agreement between models is not evidence; surviving the objections is more interesting.
Three independent engagement-and-learning reports, followed by two independent UI/UX research lanes, kept pointing away from reward economies and toward retrieval, spacing, explanatory feedback, fading scaffolds, honest competence signals, and transfer away from the training interface.
Externalize is where we test whether that can still feel satisfying.
Our orchestration work keeps landing on the same boring but useful rule: let models reason broadly, but keep identity, permissions, policy, and consequential actions outside the model and mechanically bounded.
A persuasive answer should not acquire root privileges by being persuasive.
In the RPG work, a dramatic local-model failure initially looked like a reason to add more epistemic machinery. It reproduced under one configuration, collapsed under small semantic-preserving changes, and did not recur across six hosted model families.
So we are resisting the machinery and going back to the game.
How the pieces fit
Noesis asks “could this work?”
Whyward contains what we are prepared to ship.
In between: prototypes, adversarial reviews, failed ideas, awkward evidence, and the occasional decision to stop. We would rather kill an attractive idea than manufacture confidence in it.