The model is
only one part
of the machine.
We study the systems around intelligence: memory that holds across time, state that can be inspected, rules that survive the prompt, and improvements that have to pass a test.
What should the system know that the model cannot hold alone?
Scaffolding over scale
We investigate how memory, tools, workflow state, verification, and authority can turn a capable model into a more dependable working system.
Absence blindness
A model cannot reliably notice a missing check or artifact unless the system represents what should exist and compares it with what actually does.
Continuous temporal memory
Neocortex explores memory as a searchable history of events, evidence, beliefs, goals, and open loops rather than a single summary squeezed into the next prompt.
Safe harness improvement
Keith applies the release-engineering pattern to the agent itself: candidate harness changes are isolated, tested against the current version, and promoted only when they prove better within safety limits.
Move the critical structure out of the prompt.
Compiled World is an architectural hypothesis: keep the language model as the reasoning engine while moving the world definition and valid changes into explicit machinery around it.
Give the world a version
The objects, state, actions, invariants, and transition rules are compiled into a content-addressed package. A run can name the exact world definition that governed it.
Keep facts somewhere explicit
Work state lives in typed structures instead of being reconstructed from prose after every turn. The model reasons over the world; it is not the only place that world exists.
Make every change pass through rules
Proposed actions move through defined transitions and closed outcomes, keeping success, rejection, conflict, retry, and uncertainty visible.
Treat improvement like a release
Changes to the world or the agent's working machinery can be versioned, tested, compared, and promoted deliberately instead of appearing inside a live process.
Compiled World presents a design and an evaluation plan. It does not yet claim a measured performance result.
Read the sourceResearch codenames for ideas still being tested.
Paperclip and Arthur are internal systems, not products or public agent choices. They give the research somewhere concrete to fail, change, and become clearer.
Paperclip
An internal agent for exploring how typed state, tools, explicit world rules, and reviewable changes can make long-running work more dependable.
Arthur
An internal agent in the same research family, used to test how much operational structure can move out of prose and into systems that can be inspected.
A strong idea still has to meet the evidence.
Architecture is not a benchmark
A clear design can explain intended behavior and failure boundaries. It does not by itself prove better task performance.
A prototype is not a release
Internal agents and in-development runtimes are labeled as such until their user path and release gates have been qualified.
A claim keeps its source
Research, implementation evidence, and observed product behavior remain separate so one cannot quietly stand in for another.
Our research is published as architecture and hypotheses first. Product and performance claims have to earn a different kind of evidence.