RQZ-Golf v1: Depth recurrence for parameter efficiency

records/track_non_record_16mb/rqz_golf_v1/README.md

@@ -0,0 +1,31 @@
+# RQZ-Golf v1 — Depth Recurrence for Parameter Golf
+
+## Approach
+
+Replace some unique layers with a single shared recurrent layer applied K times.
+This saves parameters (shared weights) while increasing effective depth.
+
+### Architecture
+- 7 unique layers (3 encoder + 4 decoder with U-Net skip connections)
+- 1 recurrent layer applied K=3 times with iteration embeddings
+- Effective depth: 10 layers (7 unique + 3 recurrent) vs baseline 9
+
+### Key ideas
+1. **Depth recurrence**: last block shares weights across K passes, saving ~3M params
+2. **Iteration embeddings**: learned per-pass vector so the layer knows which pass it's on
+3. **Stability scaling**: residual scaled by 1/sqrt(K) to prevent amplitude explosion
+4. **Test-time compute**: can increase K at inference (K'=6, 8, ...) for better BPB
+
+### Theoretical basis
+Inspired by Universal Transformers (Dehghani 2019) and Deep Equilibrium Models (Bai 2019).
+Each recurrent pass reconstructs the residual of the previous pass in latent space.
+
+## Config
+```
+NUM_UNIQUE_LAYERS=7
+NUM_RECURRENT_PASSES=3
+# All other params same as baseline
+```
+
+## Author
+Regis Rigaud (@TheCause)

records/track_non_record_16mb/rqz_golf_v1/submission.json

@@ -0,0 +1,9 @@
+{
+  "name": "RQZ-Golf v1",
+  "github_id": "TheCause",
+  "val_bpb": null,
+  "summary": "Depth recurrence: 7 unique layers + 1 shared recurrent layer (K=3 passes) with iteration embeddings and 1/sqrt(K) scaling. Test-time compute: increase K at inference. Preliminary baseline: 1.5283 BPB (1 shard, 1xA100).",
+  "date": "2026-03-19",
+  "track": "non_record",
+  "status": "experimental"
+}