Layer 03 · REASON · four-layer stack
OpenMythos is our open reasoning architecture: a looped transformer with sparse Mixture-of-Experts routing and switchable attention backends. Instead of scaling parameter count to scale reasoning depth, it re-enters its own latent space — same weights, more iterations, convergence-stopped.
The reasoning layer consumes layer-02 correlations: entities, campaigns, IOCs with provenance. Garbage-in is prevented upstream, not cleaned up downstream.
The case is encoded as a continuous latent state — not a one-shot token stream. State persists across loops.
The same weights iterate N times over the latent state, deepening inference without growing the model. Convergence-based early stop halts when the verdict stops changing — computation spent only where judgement is uncertain.
Output is a structured judgement: finding, confidence, and citations back to the specific events that earned it. A judgement without provenance is not published here.
Scaling parameter count to buy reasoning is the expensive road. Recurrent loops buy depth with iteration time instead — a different point on the cost curve, and one that fits on our own hardware.
The layer where judgement is formed runs on our own machines (≤4 GB VRAM target). No third-party dependency between our evidence and our conclusions.
OpenMythos is published — architecture, weights approach, and engineering choices. Intelligence infrastructure should be inspectable by its customers.
Because loops stop on convergence, every verdict carries the depth actually needed. More loops = more contested case, and that number is logged too.