Update handoff with on-policy error-targeted reasoning SFT v4

candidates/budgie-alignment-v2/README.md

@@ -366,3 +366,8 @@ A second information RL attempt added 96 decontaminated, symbolically verified O
 ## 23. Long reasoning SFT before RL v3
 
 A staged SFT-then-RL protocol tested the proposal to build reasoning before resuming RL. Stage A used 800 long, decontaminated, symbolically verified OpenR1 mathematics solutions plus 108 manual information-retention examples; 208 verified math rows were held out. The curriculum contained 908 encoded examples (mean 577 tokens, max 1,536), one full epoch / 227 updates, FP32 master weights, BF16 autocast, peak LR 3e-7, and sampled-action KL 0.12 to `verified-math-a025`. Alpha .005 improved information dev 3/21 -> 4/21 with no loops and tied random verified math 1/32, but regressed fixed GSM 5/30 -> 4/30 and MATH 3/15 -> 2/15; ARC and FOLIO stayed 11/30 and 13/30. Smaller alphas had no information gain. Stage A was rejected, so Stage B RL was correctly not run. See `reasoning-sft-then-rl-v3/recovery/V3_FINAL_REPORT.md`.
+
+
+## 24. On-policy error-targeted reasoning SFT v4
+
+V4 replaced generic long-solution imitation with Budgie's own frontier. It generated 600 trajectories on 300 stratified, verified problems (12 greedy correct, 15 sampled correct, 26 problems with any correct path). Correct Budgie trajectories became self-anchors. Of 274 failed problems, 120 were selected; Qwen3.8 produced 30 concise minimal corrections, 27 independently passed `math_verify`, 3 fell back to verified references, and 90 additional failures used verified references. The SFT set contained 308 encoded records (104 weighted self-anchors, 60 weighted Qwen corrections, 90 reference corrections, 54 information-retention rows), trained for 77 updates with LR 3e-7 and leader KL 0.20. Alpha .10 improved information dev 3/21 -> 4/21 with no loop increase, but fixed GSM fell 5/30 -> 3/30 and MATH 3/15 -> 2/15; ARC/FOLIO stayed 11/30 and 13/30. Smaller alphas did not improve information pass rate. Stage A was rejected and RL remained blocked. See `onpolicy-error-targeted-v4/recovery/V4_FINAL_REPORT.md`.