Rebuild rl/: restore the proven shaped reward, lean campaign machinery - #36

Merged

aadarwal merged 24 commits into

mainfrom

worktree-rl-clean-runs

Jul 29, 2026

Commits on Jul 28, 2026

Design: RL clean rebuild — two Qwen runs, ported v3-b shaped reward
aadarwal
committed
7314334
Plan: RL clean rebuild implementation tasks
aadarwal
committed
2bdac8e
Add rectangle-baseline computation to rl/common
aadarwal
committed
ba1289b
Measure boundary chamfer in the evaluator
aadarwal
committed
5ff5f3c
Port the v3-b shaped reward (iou_n) as policy shaped_v3b
aadarwal
committed
6fc1bc5
Spec: extend to the 2x2 run matrix (user direction)
aadarwal
committed
e55a84a
Default Track A to the shaped reward at G=8
aadarwal
committed
2532a4d
Make the env shaped-reward test degenerate-target-robust
aadarwal
committed
c96a331
Remove the studies/ceremony apparatus from rl/
aadarwal
committed
028c219
Two-run-matrix orchestrator with fixed cross-run probe eval
aadarwal
committed
bdbb767
Prompt contract v3: anisotropy scale note
aadarwal
committed
0e1878b
Document the clean training path
aadarwal
committed
d7a760d
Stages resume instead of refusing an existing log dir
aadarwal
committed
97f81b1
Disable the SDK's pyqwest transport locally
aadarwal
committed
8749d53

Commits on Jul 29, 2026