dungeon-escape medium/simple: describe new per-group softmax-temperature-range tiers (step-reward metric), replacing the old epsilon group-noise description

tier_policies/README.md

@@ -28,28 +28,35 @@ sampling plausible near-expert actions instead of epsilon's uniform-random swap)
 Collector: `causal-gpt-rl @ 28aca86` —
 `collect.py --onnx tier_policies/<Env>.onnx --temperature <T>`.
 
-PushBlock/Pyramids tiers are scored by **mean step reward** (episode return ÷
-length), normalized `expert = 1 / random = 0` — step reward (not episode return)
-because these goal games' return is bimodal (goal spike vs timeout) while step
-reward grades efficiency continuously. SoccerTwos is self-play (both teams sampled
-at the same T, so shipped match return ≈ 0 and cannot grade skill); its tier skill
-is measured by **side-swapped match score** vs the fixed stock team.
+PushBlock/Pyramids/DungeonEscape tiers are scored by **mean step reward** (episode
+return ÷ length), normalized `expert = 1 / random = 0` — step reward (not episode
+return) because these goal games' return is bimodal (goal spike vs timeout) while
+step reward grades efficiency continuously. PushBlock/Pyramids use a **fixed** T; the
+cooperative DungeonEscape samples T **per group per match** from a range (all three
+group agents share one drawn T — one coherent skill level, like an early-training
+checkpoint) via
+`collect.py --onnx tier_policies/DungeonEscape.onnx --group-temperature-range <lo,hi>`.
+SoccerTwos is self-play (both teams sampled at the same T, so shipped match return ≈ 0
+and cannot grade skill); its tier skill is measured by **side-swapped match score**
+vs the fixed stock team.
 
 ## Tier temperatures — **FINAL** (calibrated & validated at 1M)
 
 | env | medium (~0.80) | simple (~0.60) |
 |---|---|---|
-| PushBlock | **T = 2.8** | **T = 4.15** |
-| Pyramids  | **T = 3.7** | **T = 6.0**  |
+| PushBlock | **T = 2.8** (fixed) | **T = 4.15** (fixed) |
+| Pyramids  | **T = 3.7** (fixed) | **T = 6.0** (fixed)  |
+| DungeonEscape | **T ~ U(1.5, 3.0)** (per group) | **T ~ U(3.0, 4.0)** (per group) |
 
 Fixed normalization anchors are each env's shipped `expert-v0` mean step-reward
-(PushBlock **0.354242** / Pyramids **0.012437**; random ~0). 1M production
-validation (normalized mean step-reward):
+(PushBlock **0.354242** / Pyramids **0.012437** / DungeonEscape **0.019459**;
+random ~0). 1M production validation (normalized mean step-reward):
 
-| env | medium norm | simple norm | mid-band |
+| env | medium norm | simple norm | note |
 |---|---|---|---|
-| PushBlock | 0.78 | 0.58 | 51 % / 61 % |
-| Pyramids  | 0.81 | 0.60 | 47 % / 61 % |
+| PushBlock | 0.78 | 0.58 | mid-band 51 % / 61 % |
+| Pyramids  | 0.81 | 0.60 | mid-band 47 % / 61 % |
+| DungeonEscape | 0.80 | 0.60 | per-group T range; group-success 0.91 / 0.85 |
 
 (vs epsilon's 0 % mid-band). Shipped as the `medium-v0` / `simple-v0` tiers in
 `ccnets/causal-gpt-rl-unity-datasets`.
@@ -67,7 +74,8 @@ Reproducibility is **distributional** (float / physics differ across machines),
 not bit-exact.
 
 Continuous envs (Crawler, Worm, Walker, 3DBallHard) degrade via Gaussian noise on
-the root stock model, so they need no patched policy here. DungeonEscape has no
-stock ONNX that exposes per-branch logits (its published policy is a causal
-sequence model emitting a sampled action), so it cannot use this temperature dial
-without a logits-exposing policy export.
+the root stock model, so they need no patched policy here. DungeonEscape's shipped
+policy is a causal sequence model that emits only a sampled action, so a
+logits-exposing export was built from the stock `Discrete(7)` ONNX (branch Gemm
+surfaced as `logits`) and lives here as `tier_policies/DungeonEscape.onnx`; its tiers
+use the per-group temperature range above.