Update handoff with long reasoning SFT before RL v3

candidates/budgie-alignment-v2/README.md

@@ -361,3 +361,8 @@ A manual long-dialogue RL attempt targeted state updates, stale-step avoidance,
 ## 22. Math-augmented constrained information RL v2
 
 A second information RL attempt added 96 decontaminated, symbolically verified OpenR1 math problems, 192 stochastic Budgie math rollouts, 96 verified references, a 72/24 random-math train/dev split, and a 2:1 math-to-information update ratio. Training started from `verified-math-a025` and used leader KL, PPO clipping, reference BC, and global gradient projection that removed information-gradient components conflicting with math gradients (37.3% conflict rate). Alpha .01 improved information dev 3/21 -> 4/21 with no loops and random verified math dev 1/24 -> 2/24. It preserved fixed GSM 5/30 and ARC 11/30, but MATH fell 3/15 -> 2/15 and FOLIO 13/30 -> 12/30. Alpha .005 did not improve information pass rate and fell to MATH 1/15, FOLIO 12/30. All candidates were rejected. The method reduced damage versus v1 but was not Pareto-safe. See `information-rl-v2-math-augmented/recovery/V2_FINAL_REPORT.md`.
+
+
+## 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`.