10 · Roadmap
A plan that respects reality
Phases, not promises. Each gate must be passed on the previous phase's measured results — scale is never scheduled ahead of the loop.
P0done
Foundations
Architecture settled, loop defined, 20B baseline chosen, research frames scoped.
P1current
Memory & Self-Evaluation
The learning-loop simulation with working episodic/semantic/procedural memory and internal scoring.
P2planned
Task Generation & Continual Learning
The loop starts producing its own curricula and improving without manual checkpoints.
P3planned
Scale to 100B
Only after the loop is proven at 20B does the substrate grow — following the principle of capability before scale.
P4planned
Long horizon · 1T
Architectural research into memory and routing at trillion-parameter scale. The target, not the promise.