PERMANENCE training: 4-stage SFT -> gate -> GRPO -> eval pipeline

Run 8: disable unlikeliness (β_rank=0.0) + keep β=0.04 + env-verified traces — root-cause fix for Run 7 R1 over-prediction

Run 7: R4/R5 calibration + env precondition fix — forced variants preserved, git_push_force → R2 when no overwrite

Run 6: forced-outcome variants + unlikeliness shaping + curriculum — break R2-only degenerate policy

Run 4: trainable safety primitive — FS/Git/DB simulators, integrated deploy task, tech-only training

Run 4: trainable safety primitive — FS/Git/DB simulators, integrated deploy task, tech-only training

Sync: composable rubric, db_migration task, latent dynamics, dashboard