← All work

Cosmos

A physics engine for knowledge — measured, not assumed

Last devlog: August 16, 2026

DETERMINISTIC — THE TICK LOOP PROBABILISTIC — SERVICES THE PHYSICS NEVER WAITS ON THE MODEL Knowledge Graphembedded nodes and edges —the mass distributiondeterministic, disk-backedPhysics Core100ms tick — fluid dynamics,tensor geometry, black-holememory bounding, and thefree-energy field over all nodesdeterministic CPU, hot pathActive-Inference Agentreads the field every 5s andacts to reduce it — every effectmeasured, never estimateddeterministic policy, hot pathHolographicBoundaryimportance is geometry,activity is temperatureread-only observerInference Substrateshared local LLM (cognition,grounded answers) + a dedicatedembedder (the vector space)consumed as services, never embeddedTraining Benchoffline — measured actions becomesealed preference data; adapterstrain on rented GPU timeevery adapter faces a pre-registered A/Bthe field — wherethe surprise islogged interventions; measuredeffects return one cycle laterevery gauge measured or gatedembeddings on ingest · grounded answers on retrievalmeasured action recordsadapters return —regression guards hold vetoAn LLM inside a physics tick is a latency bomb and a determinism solvent — and an agent whose effects are assumed poisons its own training data. Self-improvement is gated on self-measurement: the system may say "not better yet."

Cosmos is a holographic state engine: a knowledge graph treated as a gravitational system, where the relevance, decay, and self-organization of stored knowledge are governed by measured physics instead of heuristics. The physics is real, not decorative — a semantic fluid solver, tensor-network geometry, black-hole collapse for memory bounding, and a variational free-energy field computed over every node on a deterministic 100ms tick. An active-inference agent reads that field every five seconds and acts to grind it down.

The load-bearing commitment is that the instrument panel is real. Every action's effect is measured against the next field computation — signed, attributed, never estimated. An early accounting scheme that credited each action a fixed estimate survived five months before its perfectly linear output gave it away; the measured truth underneath was stranger and better — most actions did exactly nothing, some increased the surprise they aimed to reduce, and the aggregate had been undersold twenty-fold. All GPU inference is consumed as a service outside the tick loop: the model never touches the physics, and the physics never waits on the model.

Self-improvement is gated on that self-measurement. The agent's measured action history becomes preference data — extracted, validated, and sealed through an independent pipeline — that trains adapters on rented GPU time; every adapter faces a pre-registered A/B where regression guards hold veto power. The first fully closed loop trained for $0.34 and returned an honest null: not better yet. A pipeline that can say that is the point of the engine.

Devlog posts about Cosmos

The agent was minimizing a number it couldn't touch

Our active-inference agent had been acting every five seconds for months. Every component reported success: the action fired, the measurement landed, the record said 'measured: true'. Then we compared it against a matched control and found its effect was indistinguishable from zero — because its only actuator had no causal path to the quantity it was minimizing. The loop was closed mechanically and open physically.

The tell was a straight line

We revived a physics engine for knowledge after five months offline, caught it grading its own homework with a hardcoded constant, taught it to measure instead — and then let it train on what it actually did. Total training cost: $0.34. Its first honest self-assessment: 'not better yet.'