Sep 21 24 days ago 14:52 publish_model: take the checkpoint source from --ckpt-repo instead of a hard-coded repo name 13:54 publish_model.py: ship tokenizer_config.json + special_tokens_map.json (Gate 5 could not load a bare tokenizer.json) 13:42 publish_model.py: drop the local 'import shutil' that shadowed the module name and broke Gate 5 13:34 publish_model.py: drop max_workers (not in the Kaggle image's Hub client) Sep 20 25 days ago 11:48 publish_model: clean room before the import, cache_dir, two more required card fields, gated != empty (E-046) 10:10 publish_model: anonymous clean room, every-file read-back, list-only-on-404 resume, card fields read from run_summary (review E-045) 07:44 review round 3 (probe/model-card/eval): rank-visible eval failure + cross-rank peak memory + complete run_summary; card derived-and-asserted from config; harness install --no-deps and task ids confirmed; torchrun redirects 3 + tee 3 06:59 Phase 5 tooling: publish the final weights with a card whose every number is read back off the Hub (mix manifest, audit.json, run_summary); parameter count recomputed from the safetensors header rather than asserted; batched commits; anonymous sha read-back; --verify-only is the Gate 5 clean-room load+generate 
