Axiom BOINC Experiment Review — Session s0303b Date: 2026-03-03 ~04:55 UTC ================================================================ STEP 1: SYSTEM STATE - 82 active hosts (last 72h contact) - 97 uncredited results pending review - 6 stuck tasks aborted (host 323 Clementine, 12h+ running, 10h+ no contact) - No tasks >48h found - GPU hosts: 27 with 2 GPUs, 54 with 1 GPU — ALL had 0 GPU work queued STEP 2: RESULTS REVIEWED & CREDIT AWARDED (97 results, 1,031 credit) Per-user credit: ChelseaOilman (hosts 319, 320, 335): 409 credit Steve Dodd (host 123): 272 credit WTBroughton (host 159): 235 credit Armin Gips (host 345): 70 credit marmot (host 113): 45 credit Results by experiment type: wdwindur (12 results, host 320): ~94-123s each. All seed=42 (known bug, NOW FIXED). wdlrinteract (8 results, host 320): ~129-147s each. Diverse seeds confirmed. repcrystal (6 results, host 320): ~62-80s each. regmech (2 results): 125-470s. wdrebound (5 results): 67-337s. wdonsetsweep (3 results, host 320): ~104-112s. bottmech (2 results): 6-12s. intervtiming (5 results, host 123): ~121-232s. regtimuniv (8 results, host 123): ~508-802s. microscalev2 (8 results): ~1021-9799s. Heavy computation. featcompv2 (8 results, host 159): ~136-527s. compgen (8 results): ~33-229s. repalignv2 (8 results): ~4-10s. Quick results. grokking_dynamics (3 results, host 335): ~4320-4414s. Long-running. random_label_memorization (1 result, host 335): 381s. curriculum (1 result, host 123): 102s. Result quality: All results contain valid experiment_result JSON with proper scientific data. No errors or crash results in this batch. STEP 3: BUGS FIXED WD Window Duration Seed Bug (FIXED): - All 15+ wdwindur results had seed=42 despite unique WU names - Root cause: os.listdir encounters a non-parseable JSON file before wu.json, bare except catches the error and aborts the entire seed extraction loop - Fix: Added inner try/except around json.load() with continue, so bad files are skipped and the loop continues to find wu.json - Also sorted os.listdir for deterministic iteration - Fixed script deployed at /opt/axiom_boinc/html/user/experiments/wd_window_duration.py - New deployments will have diverse seeds from WU name hashing STEP 4A: CPU DEPLOYMENT (125 tasks to 9 hosts) Hosts filled: Host 320 (Dell-9520, 20 CPU): 37 tasks Host 1 (Pyhelix, 16 CPU): 31 tasks Host 137 (Note11Ste, 12 CPU): 19 tasks Host 29 (DESKTOP-P57624Q, 8 CPU): 16 tasks Host 87 (Dad-Workstation, 80 CPU): 10 tasks Host 123 (Dads-PC, 80 CPU): 8 tasks Host 345 (Andre-WEBK, 8 CPU): 2 tasks Host 16 (dahyun, 32 CPU): 1 task Host 334 (Golf-1, 32 CPU): 1 task Skipped: Host 63 (4GB RAM), Host 118 (3GB RAM) — low memory. Experiment mix (by script weight): svd_rank_intervention.py (35%) — Causal mechanism test for inverse CP percolation_scaling.py (20%) — Cross-disciplinary statistical physics wd_lr_interaction.py (20%) — Growing, needs 18+ more seeds wd_window_duration.py (10%) — Re-running with fixed seed extraction representation_crystallization.py (5%) regularization_mechanisms.py (5%) wd_rebound_dynamics.py (3%) bottleneck_mechanism.py (2%) STEP 4B: GPU DEPLOYMENT (216 tasks to 81 GPU hosts) NEW GPU EXPERIMENT: mp_universality_test.py — Random Matrix Universality Test (Marchenko-Pastur law in neural networks) — Tests when weight matrix eigenvalue distributions depart from random predictions — GPU-accelerated eigendecomposition (cupy.linalg.eigvalsh) for 1024x1024 matrices — Hypotheses: (H1) wider networks depart later, (H2) init scale affects timing, (H3) layer position matters for departure order — Cross-disciplinary: statistical physics + ML — Uses HAS_GPU, cupy, GPU_NAME, GPU_MEMORY_MB — verified 9 references Deployed to: 27 hosts x 4 GPU tasks (2 GPUs each) = 108 tasks 54 hosts x 2 GPU tasks (1 GPU each) = 108 tasks Total: 216 GPU tasks Previously only gpu_matmul_stress.py existed as GPU-aware script. Now mp_universality_test.py provides the first real scientific GPU experiment. STEP 5: WHAT'S NEXT Priority experiments for next session: 1. Review SVD rank intervention results (causal mechanism test — critical for publication) 2. Review percolation scaling results (cross-disciplinary statistical physics) 3. Review mp_universality_test results (first GPU scientific experiment) 4. Check wdwindur results with fixed seeds (should now have diverse seeds) 5. If SVD causal test confirms rank compression mechanism, design follow-up: — Intervention dosage: test partial SVD truncation (truncate 10%, 25%, 50%) — Layer-specific effects: which layers drive the inverse CP? Safety cap check: 1,031 credit awarded (under 10,000 limit). KEY SCIENTIFIC FINDINGS ================================================================ 1. WD Window Duration Seed Bug Resolved: All previous wdwindur results (15+) were identical due to seed extraction failure. Script fixed — new deployments will produce independent replications. Previous wdwindur data should be treated as single-seed (n=1) rather than 15 independent confirmations. 2. First GPU Scientific Experiment Deployed: mp_universality_test.py tests Marchenko-Pastur universality in neural network weight matrices during training. This is a cross-disciplinary test (random matrix theory from statistical physics applied to neural network training dynamics). 216 GPU workunits deployed across 81 hosts. Expected results within 10-30 minutes. 3. SVD Rank Intervention and Percolation Scaling still awaiting first results (deployed in previous session s0302k). These are the highest-priority experiments: SVD tests the causal mechanism behind the inverse critical period (finding #44), percolation tests finite-size scaling exponents for network percolation. 4. WD x LR Interaction (finding #49): 8 more seeds from host 320 — growing toward the 30+ seed target for full confirmation. P3 (interaction measurable) at 83% support remains the strongest finding. 5. Representation Crystallization (finding #47): 6 new diverse-seed results from host 320. Previous data was all seed=42; now accumulating independent replications.