Index the two unhealed TT-3 topology controls

README.md

@@ -53,6 +53,23 @@ and at equal size and equal rate the Qwen2.5 architecture is harder to compress
 (2.0204 vs 0.6915, both ~7B at ~40%). Scale mitigates the architecture effect but does not
 remove it. Each subdirectory's README carries the budget sweep behind its own rate.
 
+## Matched-budget TT controls — **UNHEALED**
+
+The same 3-site parameterisation with the boundary bond pinned to `r_0 = 1`, i.e. an open
+tensor TRAIN, at the *same per-layer* budget as the ring directories above. They isolate one
+variable — whether closing the ring helps at equal parameters — and nothing else.
+
+| Directory | Base model | Params | vs the ring | Licence |
+|---|---|---:|---:|---|
+| [`tt3-llama2-7b/`](./tt3-llama2-7b) | NousResearch/Llama-2-7b-chat-hf | 4.0207B | 4.0231B (−0.06%) | `llama2` |
+| [`tt3-qwen2.5-3b/`](./tt3-qwen2.5-3b) | Qwen/Qwen2.5-3B-Instruct | 2.0766B | 2.0783B (−0.08%) | `qwen-research` — **non-commercial** |
+
+> **These two received NO knowledge distillation.** Every ring directory above was healed on
+> 60M tokens; these were not. Do not place their scores in the same table as the healed rows,
+> or next to a dense teacher — that compares two different things. They are control arms,
+> published so the ring-versus-train comparison can be reproduced, not models to deploy. Each
+> README states which layers could not afford ALS and what that costs.
+
 ## ⚠️ These are NOT standard `transformers` checkpoints
 
 `state_dict_fp16.pt` holds **ring cores**, not dense `nn.Linear` weights. Loading requires