| Age | Commit message (Expand) | Author |
|---|
| 2026-04-08 | Round 38: add --penalty_lam flag to cifar_resmlp.py for Mode 2 cross-method test | YurenHao0426 |
| 2026-04-08 | Round 36: upgrade (b) wording + add EP random-target neg control to §3 | YurenHao0426 |
| 2026-04-08 | Round 35: SB and CB also show data-agnostic Mode 1 growth on random targets | YurenHao0426 |
| 2026-04-08 | Round 34 random-target ablation: Mode 1 fires under random labels too | YurenHao0426 |
| 2026-04-08 | Round 32+33 H2 ablation: add no_residual_add flag; falsify residual-as-cause ... | YurenHao0426 |
| 2026-04-08 | Add perturbation correlation metric calibration | YurenHao0426 |
| 2026-04-08 | Extend perturbation audit to vanilla early-epoch checkpoints | YurenHao0426 |
| 2026-04-08 | Add perturbation correlation audit (round 19's recommended alt metric) | YurenHao0426 |
| 2026-04-08 | Add BP+penalty control (round 19's #4 critical experiment) | YurenHao0426 |
| 2026-04-08 | Add vanilla DFA early-epoch checkpoint training (round 19 disambiguation) | YurenHao0426 |
| 2026-04-08 | Add null calibration script: training-Bs vs fresh-Bs cos on penalized DFA | YurenHao0426 |
| 2026-04-08 | Add penalty lambda 3-seed summary script + checkpoint save in penalty test | YurenHao0426 |
| 2026-04-07 | Add fast direction-quality measurement on existing DFA checkpoints | YurenHao0426 |
| 2026-04-07 | Add ViT-Mini DFA training script that saves checkpoint + Bs | YurenHao0426 |
| 2026-04-07 | Add DFA direction-quality direct test (codex round 13 option c) | YurenHao0426 |
| 2026-04-03 | Fix CNN compute_bp_grads: remove inter-layer detach so gradients flow to all ... | YurenHao0426 |
| 2026-04-03 | Add checkpoint saving to ep_synthetic.py | YurenHao0426 |
| 2026-04-03 | Fix EP credit sign in cnn_baseline.py | YurenHao0426 |
| 2026-04-02 | Fix EP credit sign: negate (h_nudge - h_free)/β to align with BP grad direction | YurenHao0426 |
| 2026-04-02 | Add --d_hidden arg to ep_baseline.py for d=512 support | YurenHao0426 |
| 2026-04-02 | Fix ep_synthetic: bp dict needs L+1 entries for EP credit comparison | YurenHao0426 |
| 2026-04-02 | Add EP synthetic ladder script | YurenHao0426 |
| 2026-04-02 | Fix CNN state bridge: use custom CNNStateBridge for variable input dims | YurenHao0426 |
| 2026-04-02 | Add SB and CB methods to cnn_baseline.py | YurenHao0426 |
| 2026-04-02 | Fix gelu_ablation.py: compute method-specific Gamma instead of hardcoded 1.0 | YurenHao0426 |
| 2026-04-02 | Add GELU/ReLU ablation script for CIFAR MLP | YurenHao0426 |
| 2026-04-02 | Add CNN baseline: SmallCNN with BP/DFA/EP on CIFAR-10 | YurenHao0426 |
| 2026-04-02 | Add EP baseline implementation (Scellier & Bengio 2017) for CIFAR MLP | YurenHao0426 |
| 2026-04-01 | Add d512_sparsity.py: support sparsity for d=512 checkpoints | YurenHao0426 |
| 2026-04-01 | Add cifar_d512_confirmatory.py: L=4 d=512 with checkpoint saving | YurenHao0426 |
| 2026-04-01 | Add clean_sparsity_persample.py: per-sample gradient stats | YurenHao0426 |
| 2026-04-01 | Add clean_sparsity_full.py: independent-process full sparsity analysis | YurenHao0426 |
| 2026-04-01 | Add clean gradient check: independent Python process per method, GPU 1 | YurenHao0426 |
| 2026-04-01 | Add confirmatory supplement: T1-T4 from checkpoints (no retraining) | YurenHao0426 |
| 2026-04-01 | Add extended sparsity analysis: A4 per-layer, B1 snapshots, B2 active subset,... | YurenHao0426 |
| 2026-04-01 | Add BP support sparsity analysis: threshold sweep + gradient histograms | YurenHao0426 |
| 2026-03-31 | Update naive StateErr v3: L2 norm ratio formula, with checkpoints saved | YurenHao0426 |
| 2026-03-31 | Add BP supplement for A2 CIFAR: 10 seeds with acc, Gamma, rho, naive_StateErr | YurenHao0426 |
| 2026-03-31 | Add naive state prediction baseline for A1 and A2 | YurenHao0426 |
| 2026-03-30 | Add confirmatory paper experiments: A1-A4, all 10 seeds complete | YurenHao0426 |
| 2026-03-27 | Add Phase 10A.8C: 3-seed replication — scaffold gains are marginal | YurenHao0426 |
| 2026-03-27 | Add Phase 10A.8: freeze-with-decay confirms stale aux is main freeze failure ... | YurenHao0426 |
| 2026-03-27 | Add Phase 10A.7: minimal aux compression — continuous trainability is essen... | YurenHao0426 |
| 2026-03-26 | Add Phase 10A.6: gain requires trainable depth-aware aux, not semantic credit | YurenHao0426 |
| 2026-03-26 | Add Phase 10A.5: blend gain is implicit regularization, not learned credit | YurenHao0426 |
| 2026-03-26 | Add Phase 10A: no prefit threshold — even random Vec blend beats DFA by +1.3% | YurenHao0426 |
| 2026-03-26 | Add Phase 9B+9C: periodic refit fails, top-down curriculum neutral | YurenHao0426 |
| 2026-03-25 | Add Phase 9A: checkpointed handoff — blend(Vec+DFA) outperforms pure DFA | YurenHao0426 |
| 2026-03-25 | Add Phase 8: schedule timing test — online co-learning is the remaining bot... | YurenHao0426 |
| 2026-03-25 | Add Phase 7A: snapshot time sweep shows early snapshots have positive held-ou... | YurenHao0426 |
