CPU-only: drop `spaces`/ZeroGPU (403 PRO required wedged the Space in RUNTIME_ERROR)

CPU-only: drop `spaces`/ZeroGPU (403 PRO required wedged the Space in RUNTIME_ERROR)

log: retract v7-v12 KL conclusion; metric artifact + logp_floor fix

--logp-floor: exclude Dirichlet-forced samples (logp_old<-50, 0.9%) from PPO ratio; worst-sample kl 22015->35, max grad 254->45

approx_kl: report robust median + tail stats; exp-tailed mean was an artifact that v7-v12 were tuned against

log: repetition fix necessary but not sufficient

log: v7-v12 stability campaign, 3 bugs fixed, no strength gain

kl-hard must never zero out an iteration (nup>=min_updates)

add --kl-hard: hard KL ceiling overriding --min-updates

fix 2nd train/inference mismatch: last-move feature lost in FEN round-trip

fix 2nd train/inference mismatch: last-move feature lost in FEN round-trip

anti-windup on adaptive entropy: back off ent_coef when approx_kl diverges

log: repetition-feature bug, greedy baseline, learn bench

FATAL instead of silent random-init when --resume target is missing; support hub-path resume

fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp

fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp

fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp

fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp

fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp

fix eval-on-resume (silently skipped); add learn bench; v4 analysis

add learn-phase thread benchmark

revert logp_old change (ablation-falsified); keep only_player + kl estimator fixes

add logp_raw ablation switch

kl early-stop: floor-then-stop semantics

adjudicate claimed repetition/50-move draws on material

draw-penalty: remove guaranteed-0 repetition optimum

fix approx_kl estimator + PPO KL early stopping with min-updates floor

PPO correctness: logp_old from behaviour dist (temp+dirichlet), not raw policy

PPO fix: exclude opponent positions from buffer + staleness bound

bench: __main__ guard for spawn

spawn-based persistent self-play pool (fork/libgomp deadlock fix)

parallel self-play (--workers), conversion experiments, whitelist

mate-capability suite + analysis scripts

whitelist analysis scripts; fix pid-recycling in launch/job_table

fix: tolerate missing league snapshots after container rebuild

resume from hub after container rebuild (FINAL code change before long run)

add cross-arm grid analysis

ephemeral-safe: auto push-every, token secret

research log: env correction + thread bench

add remote job runner + file inspector

add environment audit tab

zerogpu: add spaces import + @spaces.GPU entrypoint

tinychess: full research artifact - self-play RL, ablations, compute scaling, faithfulness, precision, plasticity

fix: disable gradio SSR (node server crash)

fix: gradio 5 for huggingface_hub>=1.0 compat

tinychess: self-play research substrate (phase 1-4)

remove probe

probe write access

initial commit