Update handoff with math-augmented constrained information RL v2

candidates/budgie-alignment-v2/README.md

@@ -356,3 +356,8 @@ After explicit user authorization, the strongest full-benchmark checkpoint `veri
 ## 21. Information-management RL v1
 
 A manual long-dialogue RL attempt targeted state updates, stale-step avoidance, selective retrieval, delta summaries, conflict clarification, source boundaries, exact constraints, and repetition. It used 105 disjoint multi-turn cases, 252 stochastic Budgie rollouts, 63 manual ideal demonstrations, Qwen3.8 group rewards, clipped token-level GRPO/PPO with leader KL, a longer GRPO+BC run, and manual ideal-response consolidation. The full specialist source improved held-out information pass rates (dev 3/21 -> 7/21; test 1/21 -> 5/21) but increased loops. A low-drift alpha .025 improved dev to 4/21 and test to 2/21 without extra loops, then failed the protected reasoning gate: GSM 5/30 -> 3/30 and MATH 3/15 -> 1/15, with ARC 11/30 and FOLIO 13/30 unchanged. The branch is rejected; `verified-math-a025` remains active and the public root is unchanged. Read `information-rl-v1/recovery/FINAL_REPORT.md`. One reconstructable hybrid source is retained; superseded/rejected full checkpoints were removed from the current tree.
+
+
+## 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`.