Upload ablation/ABLATION_REPORT.md with huggingface_hub

ablation/ABLATION_REPORT.md

@@ -0,0 +1,64 @@
+# Routing Ablation Experiment
+
+**Date:** 2026-01-22T15:22:36.045048
+
+## Purpose
+
+Test if the routing breakthrough is genuine or an artifact of:
+1. Readout alignment (τ-weighted write + τ-weighted read)
+2. Noise-only interference (Gaussian vs structured)
+
+## Part A: Readout Neutralization
+
+### Question
+Does C/D dominate B under UNIFORM readout (not just τ-weighted)?
+
+### Results
+
+**Basin Width Table (50% threshold)**
+
+| Readout | A | B | C | D |
+|---------|---|---|---|---|
+| uniform | 0 | 2048 | 4096 | 4096 |
+| slow_only | 0 | 1024 | 4096 | 4096 |
+| tau_weighted | 0 | 1024 | 4096 | 4096 |
+
+### Verdict: ROUTING_GENUINE
+
+C/D dominate B even under UNIFORM readout:
+  B (uniform readout): 2048
+  C (uniform readout): 4096
+  D (uniform readout): 4096
+→ Routing genuinely improves global retention, not just aligned readout.
+
+## Part B: Structured Interference
+
+### Question
+Does routing hold against low-rank correlated and repeating interference?
+
+### Results
+
+**Basin Width Table (uniform readout, 50% threshold)**
+
+| Interference | B | C | Delta |
+|--------------|---|---|-------|
+| gaussian | 2048 | 4096 | +2048 |
+| low_rank | 512 | 1024 | +512 |
+| repeating | 256 | 512 | +256 |
+
+### Verdict: ROUTING_ROBUST
+
+Routing advantage holds across ALL interference types.
+
+## Overall Verdict
+
+**ROUTING_CONFIRMED**
+
+Routing is GENUINE and ROBUST:
+  - Holds under uniform readout (not alignment artifact)
+  - Holds under structured interference (not noise-specific)
+→ Ready to integrate into FDRA training.
+
+---
+
+*Report generated by routing_ablation_experiment.py*