diff options
| -rw-r--r-- | docs/campaign/CASCADE_ABLATION_PLAN.md | 8 | ||||
| -rw-r--r-- | ep_run/fig_primer_pack2.py | 4 |
2 files changed, 10 insertions, 2 deletions
diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 6859751..215d586 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,14 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### ERRATUM to RESULT 21 (found 2026-07-15): the original fw72m's best val 3.7117 was first hit +at step 104,800 (log line 1055), NOT at s185000 (that was the checkpoint used for sampling). +Best-val progression: 3.8769@30.9k -> 3.8328@61.3k -> 3.8004@75.3k -> 3.7789@83.3k -> +3.7117@104.8k -> flat for the remaining ~90k healthy steps until the 195k blow. The original +schedule was therefore SNR-starved from ~105k on (wall-1), before it was killed by wall-2. +Strengthens the window story; blowup figure annotation corrected. (fw72m_cent at the same +104.8k mark: best 3.4523 and still descending.) + ### RESULT 26 (2026-07-15): C512 ZERO-GAP SEALED — full-epoch centered erases the 0.050 gap. stage1b_cent (C512 42.75M, --est centered whole epoch, bf_late 3e-3@15k, amp, seed 1, GPU0) completed all 58800 steps: best val CE 1.2334. diff --git a/ep_run/fig_primer_pack2.py b/ep_run/fig_primer_pack2.py index d23f9ca..47be501 100644 --- a/ep_run/fig_primer_pack2.py +++ b/ep_run/fig_primer_pack2.py @@ -97,8 +97,8 @@ axv.plot(rs / 1000, np.clip(rval, 0, 6.42), color='#999999', lw=0.9, zorder=2) axv.text(233, 4.32, 'rescue resumes\n(new recipe from 195k / 215k)', fontsize=8, color='#777777', ha='right') axv.axhline(3.2884, color=GRAY, lw=1.0, ls='--') axv.text(2, 3.19, 'BP twin final 3.2884', fontsize=8.5, color=GRAY, va='top') -axv.plot([185], [3.7117], marker='*', ms=14, color=GREEN, zorder=5) -axv.text(180, 3.50, 'best 3.7117 @185k (never beaten again)', fontsize=8.5, color=GREEN, ha='right') +axv.plot([104.8], [3.7117], marker='*', ms=14, color=GREEN, zorder=5) +axv.text(102, 3.52, 'best 3.7117 @105k — never improved again:\n90k healthy steps of plateau before the blow', fontsize=8.5, color=GREEN, ha='left') axv.annotate('195k: relaxation diverges — val 3.9 → 6.0,\nguard storm (15k step-skips); the same\n$\\beta$ was safe for the previous 135k steps', xy=(204, 5.95), xytext=(112, 5.62), fontsize=9.5, color=RED, arrowprops=dict(arrowstyle='-|>', color=RED)) |