Upload README.md with huggingface_hub

README.md

@@ -1,41 +1,37 @@
 ---
-dataset_info:
-  features:
-  - name: prompt
-    dtype: string
-  - name: completion
-    dtype: string
-  - name: turn_id
-    dtype: string
-  - name: traj_id
-    dtype: string
-  - name: source
-    dtype: string
-  - name: language
-    dtype: string
-  - name: phase
-    dtype: string
-  - name: stratum
-    dtype: string
-  - name: n_prefix_chars
-    dtype: int64
-  - name: n_prefix_tokens
-    dtype: int64
-  - name: teacher_lift
-    dtype: float64
-  - name: modal_share
-    dtype: float64
-  - name: n_pool_refs
-    dtype: int64
-  splits:
-  - name: train
-    num_bytes: 545136361
-    num_examples: 18138
-  download_size: 150886734
-  dataset_size: 545136361
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: mit
+tags: [sft, affine, sn120, cue-ablation]
 ---
+
+# cartoon-6a1f44a7
+
+`iamPi/tessera-77d11909` with a fixed cue appended to the end of every **thought**.
+
+The SFT completion is `</think>\nTHOUGHT: {z}\n\n{y}`; only `z` enters
+Reason (`a_i = lpC(y_i|z_A) - lpC(y_i|empty)`), so the cue is placed at the end
+of `z`, before the blank line that introduces the action.
+
+Cue (102 chars):
+
+```
+
+
+The analysis is complete. Output the next bash command only, with no explanation. Next bash command:
+```
+
+Rows: 18138. Every row was verified to round-trip through the validator's
+own `split_rollout` both before and after the rewrite.
+
+## Provenance of the cue
+
+Measured on a 300-turn held-out slice against the contract teacher
+(`zai-org/GLM-4.5-Air-FP8`), appending this cue to an SFT checkpoint's thoughts
+moved mean Reason 0.00660 -> 0.01372 (+0.00713, SE 0.00178, z +3.99). It held
+under a fresh teacher-reference draw (+0.00051, against -0.00088 for a
+fixed-policy king control), so it is not an artifact of one reference sample.
+
+Caveats worth carrying: ~85% of the gain sits in ~7% of turns (the tau=0.03
+log-mean-exp tail), the median per-turn effect is near zero, and the B-gate
+pass rate fell 0.527 -> 0.493 (floor 0.30). Training a model to *emit* the cue
+is not the same intervention as appending it post-hoc, so +0.00713 is an upper
+bound on what transfers.