Sep 21 14:58 p7 probe: prove --ckpt-repo chooses the publisher's weight source 14:56 mirror kernels/ so the kept code is self-contained 14:56 mirror probes/ so the kept code is self-contained 14:56 mirror ops/ so the kept code is self-contained 14:56 mirror config/ so the kept code is self-contained 14:52 publish_model: take the checkpoint source from --ckpt-repo instead of a hard-coded repo name 14:50 final report: 55 error records 14:50 README: name the three kept repos; record what the pruning removed and why a fresh run is unaffected 14:42 Phase 5: keep the raw publish/Gate 5 stdout with the docs, so the private scratch repo is redundant 14:35 README: references point at in-repo docs/ and guide/, not one machine 14:34 publish guide/: the operating manual distilled from 54 failure records, for future agent sessions 14:34 publish docs/: plan, mix rationale, preflight report, run log, frozen eval protocol, final report 14:27 add p6 launcher with a Hub heartbeat: a kernel that cannot report is a kernel you cannot debug (E-056) 14:27 README: full layout + how to train a new model (flags, invariants, pinning) for reuse 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 13:10 run_benchmarks: publish the effective shot count with its provenance, so a 0-shot primary row can never be read as few-shot 13:08 run_benchmarks: the 5-row smoke also verifies PRIMARY_METRIC against real results keys, so a wrong metric name cannot survive to the full columns 13:02 p6_cli_probe v3: build metric keys from the YAML's own metric/aggregation fields and rank candidate config files -- the first pass guessed metric_uri and picked a mmlu subject file 12:56 p6_cli_probe: locate lm_eval via a child process -- a user-site dir created mid-run is not on the parent's sys.path 12:50 p6_cli_probe: check run_benchmarks' flag list and PRIMARY_METRIC against the installed 0.4.13 CLI and its shipped YAMLs 12:25 run_benchmarks: install with deps, ask the task registry via the API, refuse a row whose primary metric key is absent (E-047/E-048) 12:15 p6_harness_probe v2: measure what a resolving lm-eval install moves, instead of assuming --no-deps is safe (it is not: sacrebleu import failure) 12:08 p6_harness_probe: does lm-eval 0.4.13 resolve on transformers 5.0.0, and what do the 9 task configs declare (configs only, no eval data) 11:48 run_benchmarks: real sample row counts, no abort on legitimate null shots, verdict honours status (E-046) 11:48 publish_model: clean room before the import, cache_dir, two more required card fields, gated != empty (E-046) 11:30 run_benchmarks: drop the unused binding from the transformers assertion 11:12 bootstrap: QUOTA_LEFT_HOURS 24.9 from the 11:10Z live read -- session 1 launch card 11:00 phase4_bootstrap: REV must be at or after the launcher it sha-pins -- point at c237b478 11:00 phase4_prep: REV must be at or after the launcher it sha-pins -- point at c237b478 10:58 p4_plan_sweep: repin to the D-018 launcher at c237b478 10:58 phase4_bootstrap: repin to the D-018 launcher at c237b478 10:58 phase4_prep: repin to the D-018 launcher at c237b478 10:58 launcher: adopt D-018 -- no gradient checkpointing, push every 127 steps, plan at the sustained 12,300 tok/s 10:16 p4_plan_sweep: drive the shipped plan() at each candidate rate to settle the session shape before D-018 10:10 run_benchmarks: task-list hard fail, per-task shot/denominator provenance, primary metric pre-registered, failed cells not zeros, push verified anonymously (E-045) 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) 09:43 p4_probe2_soak_runner: the wrapper as submitted for probe chain v8 / soak / rehearsal v4 09:43 p4_prep_runner: the wrapper as submitted for probe chain v8 / soak / rehearsal v4 09:43 p4_probe_runner: the wrapper as submitted for probe chain v8 / soak / rehearsal v4 09:26 phase4_bootstrap: E-044 repin (rank-0 stop assertions; memory gate moved onto the training-only leg) 09:26 phase4_prep: E-044 repin (rank-0 stop assertions; memory gate moved onto the training-only leg) 09:26 p4_stop_probe: E-044 repin (rank-0 stop assertions; memory gate moved onto the training-only leg) 09:25 launcher: repin the E-044 trainer 09:25 trainer: the post-stop assertions are rank-0 facts (E-044: hub_cb.steps is empty on rank 1 by construction, so the guard killed every correctly-checkpointed segment) 09:09 p4_prep_runner: re-pin to the body actually submitted 09:09 p4_probe_runner: re-pin to the body actually submitted 09:07 phase4_bootstrap: repin to b6210a29 (per-rank post-train marks + readable probe failures) 09:07 phase4_prep: repin to b6210a29 (per-rank post-train marks + readable probe failures) 
