Architecture notes (high level)
This repo compares a dense baseline vs a structured "dual-stream" model on synthetic compositional tasks.
Key stabilized regime (what produced the replicated ood_long gains):
- chiasm_mode = bias (no information starvation)
- curriculum mixing that includes long+gap samples during training
- Schedule-2 with ramp; braid rehearsal prevents late-phase regression
See docs/preliminary_results.md for the replication panel and ablations.