original_id branch_at_run run_commit preserved_branch_tip integrated_commit status val_bpb memory_gb description evidence notes 96dfb80 codex/rtx5060-target 96dfb80 7a87b6d keep 1.049962 2.9 baseline 5060 consumer path batch8 eval8 results.tsv baseline and first experiment share the same branch session 7a87b6d codex/rtx5060-target 7a87b6d 7a87b6d discard 1.122834 2.9 human: hard salience gate on VE using BOS goal-fit minus mean-state redundancy results.tsv f3c6755 codex/salience-chew f3c6755 f3c6755 discard 1.193117 2.9 human: contextual VE gate inputs using BOS context minus sequence-mean redundancy results.tsv e7620ce codex/salience-seq-controller e7620ce e7620ce discard 1.125752 2.9 human: sequence-level contextual controller over attention and MLP residual branches results.tsv 9a52d53 codex/salience-head-diversity 9a52d53 9a52d53 discard 1.063856 2.9 human: q-head redundancy regularizer coeff 0.01 results.tsv fa2d7b6 codex/salience-head-diversity-small fa2d7b6 fa2d7b6 discard 1.062289 2.9 human: q-head redundancy regularizer coeff 0.001 results.tsv 72c3e7d codex/salience-novelty-curriculum 72c3e7d 72c3e7d discard 1.125527 2.9 human: inverted-U novelty curriculum over per-sequence training losses results.tsv 11c07fd codex/window-full 11c07fd 11c07fd discard 1.067392 2.9 full attention window pattern L on all layers results.tsv cebc334 codex/no-checkpoint-5060 cebc334 cebc334 keep 0.970973 6.2 disable activation checkpointing on RTX 5060 profile results.tsv 0b1c731 codex/no-checkpoint-5060 0b1c731 0b1c731 keep 0.946895 6.2 no checkpointing plus WARMDOWN_RATIO 0.3 results.tsv 8a91ebd codex/no-checkpoint-5060 8a91ebd 8a91ebd keep 0.924877 6.2 no checkpointing plus WARMDOWN_RATIO 0.2 results.tsv e68fceb codex/no-checkpoint-5060 e68fceb e68fceb keep 0.921273 6.2 no checkpointing plus WARMDOWN_RATIO 0.1 results.tsv f69c531 codex/constant-lr-test f69c531 f69c531 discard 0.924117 6.2 no checkpointing plus constant LR warmdown 0.0 results.tsv branch session is short because the commit was made right before switching back f272afd codex/adaptive-lr-meta f272afd f272afd discard 0.948483 6.2 human: adaptive LR multiplier from AGI meta layer results.tsv f279649 codex/anti-salience-plain-transformer f279649 f279649 discard 1.053944 6.0 human: anti-salience ablation disabling explicit value-residual pathway results.tsv 30c79cf codex/bipolar-lr-controller 30c79cf 30c79cf discard 0.932085 6.2 human: bipolar promotive/aversive LR controller results.tsv run happened on codex/bipolar-lr-controller; codex/bipolar-lr-controller-final is just the archive pointer ae4cd00 codex/matrix-lr-0p05 ae4cd00 b72a6b3 d9a5dec keep 0.898970 6.2 no checkpointing plus WARMDOWN_RATIO 0.1 plus MATRIX_LR 0.05 recovered from b72a6b3:current_best_run.json missing from results.tsv; tip b72a6b3 only updates current_best_run.json 29d1f4f codex/matrix-lr-0p06 29d1f4f 29d1f4f discard 0.937846 6.2 trunk: MATRIX_LR 0.06 results.tsv 87f0423 codex/salience-adamw-v1 87f0423 87f0423 discard 0.903525 6.2 human: saliencew optimizer v1 on AdamW parameter groups results.tsv f196724 codex/salience-adamw-v2 f196724 59a96bc 57977df keep 0.867978 6.2 human: saliencew optimizer v2 on AdamW parameter groups results.tsv + 59a96bc:current_best_run.json 59a96bc only updates current_best_run.json; 57977df is the equivalent trunk cherry-pick b8d903b codex/salience-muon-gate-v1 b8d903b b8d903b discard 0.871641 6.2 human: first salience gate on Muon matrix path results.tsv ledger row points at b8d903b, but the commit diff is instrumentation-only; likely commit/description mismatch eff3bbf codex/salience-muon-gate-v2 eff3bbf ade5764 c0ba7b2 keep 0.861187 6.2 human: tuned salience gate on Muon matrix path results.tsv + ade5764:current_best_run.json ade5764 only updates current_best_run.json; c0ba7b2 is the equivalent trunk cherry-pick adam2m01 codex/adamw-beta2-salience-v1 274e194 274e194 discard 0.876374 6.2 human: salience-aware AdamW second-moment rule results.tsv synthetic synthetic row id; tip commit matches the experiment valmuon1 codex/value-embeds-muon 3a5d4a6 baa4bb2 discard 0.983434 6.1 human: move value_embeds from AdamW path to Muon results.tsv synthetic + reflog the run state survives as 3a5d4a6 on codex/no-checkpoint-5060; codex/value-embeds-muon never advanced past baa4bb2 valeadam1 codex/value-embeds-plain-adamw 9a5e0a2 9a5e0a2 discard 0.871933 6.2 human: value_embeds on plain AdamW with lower LR results.tsv synthetic run state is preserved as the revert commit 9a5e0a2 scaladam1 codex/per-group-adamw-variants 4bc58c2 ace33e7 keep 0.857426 6.2 human: scalar optimizer back to plain AdamW results.tsv synthetic + current_best_run run commit is 4bc58c2; tip ace33e7 only updates current_best_run.json wteadam1 codex/wte-plain-adamw ba60543 4066d9d keep 0.852087 6.2 human: token embeddings back to plain AdamW results.tsv synthetic + b95a22a:current_best_run branch run is ba60543; 4066d9d only updates current_best_run.json; b95a22a current_best_run.json points at 5e84d0d, but that SHA is a misleading cherry-pick target rather than the exact run commit decaysw1 codex/salience-decay-v1 b95a22a discard 0.877244 6.2 human: salience-controlled decay on saliencew groups results.tsv synthetic + 11.13 minute branch session no commit survives for the exact run state; the branch tip was reused later as an alias to the wte winner lmheps1 ambiguous (late sessions on codex/no-checkpoint-5060 then codex/per-group-adamw-variants) ace33e7 discard 0.876828 6.2 human: lm_head epsilon 1e-8 results.tsv synthetic + reflog no committed run state survives; codex/lmhead-eps-1e8 and codex/lmhead-eps-1e8-v2 are placeholder aliases created after the run window