vec_log per-agent population mean - average over agents instead of over completed episodes#459

Open

eugenevinitsky wants to merge 12 commits into

emerge/temp_trainingfrom

ev/per-agent-logging

Commits on May 30, 2026

WIP: vec_log per-agent population mean to remove completion-rate bias

Eugene Vinitsky
and
claude
committed
drive/binding: drop forward-decl wiring comments

Eugene Vinitsky
and
claude
committed
env_binding: focus vec_log gate comment on the divide-by-zero invariant

Eugene Vinitsky
and
claude
committed
env_binding: drop unused num_agents extraction in vec_log

Eugene Vinitsky
and
claude
committed
drive: switch per-agent log aggregation from window-mean to EMA

Eugene Vinitsky
and
claude
committed
tests: cover the per-agent EMA logging contract

Eugene Vinitsky
and
claude
committed
drive: allocate per-agent log buffers once at init, drop dead grow path

Eugene Vinitsky
and
claude
committed
env_binding: guard eval-branch divide against per-env n=0

Eugene Vinitsky
and
claude
committed
ci: include per-agent-logging and drive-training tests in utest

Eugene Vinitsky
and
claude
committed
env_binding: skip empty entries entirely in eval-branch list

Eugene Vinitsky
and
claude
committed
drive: shorten log_ema_alpha half-life to 2 episodes; fix ruff-format

Eugene Vinitsky
and
claude
committed

Commits on Jun 1, 2026

Merge branch 'emerge/temp_training' into ev/per-agent-logging
Copilot
authored