Axiom BOINC Session Log — s0303x/y (~13:20-13:48 UTC, March 3, 2026) ===================================================================== PART 1: VALIDATION & CREDIT (Session s0303x, ~13:20 UTC) ========================================================= Credit Awarded: 324 results, ~1,516 credit, 8 users ---------------------------------------------------------------------- | User | Credit | Hosts | |-------------------|--------|----------------------------------------------| | Steve Dodd | +610 | h123 Dads-PC, h87 Dad-Workstation | | WTBroughton | +535 | h159 achernar, h347 sirius (NEW!) | | Amapola | +190 | h287 DESKTOP-N5RAJSE | | Anandbhat | +84 | h222 DESKTOP-11MAEMP | | ChelseaOilman | +46 | h319 Dell-XPS | | [DPC] hansR | +24 | h9 dbgrensenh27 | | PyHelix | +12 | h1 Pyhelix | | Vato | +1 | h6 iand-r7-5800h | Total: ~1,516 credit (well under 10K safety cap) Counters updated: credited_count.txt: 3687 -> 4021 (+334) total_results.txt: 109,211 Cleanup: System Healthy ---------------------------------------------------------------------- - 0 stuck tasks - 0 tasks >48 hours - 0 broken experiments — all active scripts healthy Host Notes: - sirius (h347, WTBroughton): Delivering first science results! New host discovered last session. KEY SCIENCE ADVANCEMENTS FROM VALIDATED RESULTS ---------------------------------------------------------------------- *** Finding #61 (Depth Generality): 8 -> 66 seeds — NOW STRONGLY CONFIRMED! *** - depth=2: 100% | depth=3: 95.5% | depth=4: 84.8% | depth=5: 74.2% | depth=6: 75.8% - Inverse CP generalizes across ALL tested depths (d2-d6). - This was the weakest link in the publication story — now PUBLICATION-READY. - +58 new seeds this session, massive progress from boosted priority (weight 30). *** Finding #60 (Task Generality): 86 -> 100 seeds — STRONGLY CONFIRMED *** - compositional=100% | spiral=100% | teacher-student=96% | hidden_manifold=87% - +14 new seeds this session. Inverse CP confirmed across 4 distinct tasks. *** Finding #59 (Width Generality): 142 -> 182 seeds — DEFINITIVELY CONFIRMED *** - 100% inverse CP at ALL widths w24-w64. - +40 new seeds this session. Paradigm shift complete. PART 2: RESEARCH & DEPLOYMENT (Session s0303y, ~13:45 UTC) ============================================================ NEW EXPERIMENT DESIGNED & DEPLOYED ---------------------------------------------------------------------- *** Finding #62: wd_timing_extended_width.py — ADDRESSES KEY PUBLICATION QUESTION *** Tests inverse CP for weight decay at widths [96, 128, 256, 512] with WD=0.05. This directly addresses the REMAINING QUESTION: does inverse CP generalize beyond the w64 maximum tested in Finding #59 (182 seeds, 100% at w24-w64)? Setup IDENTICAL to #44 and #59: compositional task, depth=2, 150 epochs, SGD. Expected runtime: ~2-15 min per WU on CPU. Novelty check: NOVEL — no prior data exists for inverse CP at w96-w512 with WD=0.05 (matched hyperparams). Finding #54 tested similar widths but with WD=0.001 (50x too weak), making results incomparable. Finding #62 uses the CORRECT WD magnitude to answer whether inverse CP is truly width-universal. Deployment Summary: 614 total WUs (600 CPU + 14 GPU), 0 errors ---------------------------------------------------------------------- CPU Experiments (600 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Priority / Notes | |--------------------------------|------|---------|-----------------------------------------------| | wd_timing_depth | 150 | #61 | Depth generality, d4-d6 need more seeds | | wd_timing_extended_width | 108 | #62 | NEW! Large width w96-w512, just deployed | | svd_rank_intervention | 72 | #51 | 293+ results, causal mechanism | | wd_timing_task_sweep | 60 | #60 | 100 seeds, strongly confirmed | | percolation_scaling | 54 | #52 | 199+ results, cross-disciplinary | | wd_width_transition | 42 | #59 | 182 seeds, definitively confirmed | | wd_lr_interaction | 42 | #49 | 183+ results | | wd_window_duration | 30 | #50 | 106+ results | | wd_optimizer_interaction | 24 | #56 | 58+ seeds | | representation_crystallization | 18 | #47 | 79+ results | Note: regmech/wdrebound/bottmech skipped — 1,142 unsent WUs already in queue from prior sessions. GPU Experiments (14 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Notes | |--------------------------------|-----|---------|-------------------------------------------| | mp_universality_test | ~8 | #53 | 36+ results, continuing | | wd_timing_scale_gpu | ~6 | #54 | 34+ seeds all FALSE, robustness check | GPU Checkpoint: Active GPU hosts: 6+ confirmed - Pyhelix (h1, PyHelix): RTX 3080 Laptop - thonon-meylan (h249, dthonon): RTX 4060 Ti - Dell-XPS-15-9560 (h319, ChelseaOilman): GTX 1050 x2 - W10-Home (h57, Henk Haneveld): GPU - dbgrensenh27 (h9, [DPC] hansR): Quadro M4000M - DESKTOP-P57624Q (h29, kotenok2000): GPU GPU workunits this session: 14 (wd_timing_scale_gpu + mp_universality_test) GPU scripts active: wd_timing_scale_gpu.py, mp_universality_test.py Retired GPU experiments: grad_subspace_wd_gpu.py (7 seeds, COMPLETE NEGATIVE) Known issue: gigas (h322) SSL cert failure — cannot download scripts (low priority) Dispatch Status: ---------------------------------------------------------------------- 169 dispatched to hosts immediately: sirius (h347, WTBroughton): 135 grabbed — voracious new host! DESKTOP-P57624Q (h29, kotenok2000): 30 vega (h348): 4 — NEW HOST DISCOVERED! i5-10400, 12 CPUs, 31GB RAM 439 unsent waiting for 15+ additional hosts to poll New Host Discovered: - vega (h348): i5-10400, 12 CPUs, 31GB RAM — first appearance in findings! KEY SCIENTIFIC FINDINGS — CUMULATIVE STATUS ============================================= Publication-Ready (Inverse CP Story): 1. Finding #44 (Inverse CP for WD): 34+ seeds, 100% at w32. BASELINE. 2. Finding #45 (Reg Timing Universality): 140 seeds. WD (100%) + Dropout (98.6%). COMPLETE. 3. Finding #59 (Width Generality): 182 seeds, 100% at w24-w64. DEFINITIVE. 4. Finding #60 (Task Generality): 100 seeds, 4 tasks (100%/100%/96%/87%). STRONG. 5. Finding #61 (Depth Generality): 66 seeds, d2-d6 (100%/95.5%/84.8%/74.2%/75.8%). STRONG. 6. Finding #51 (SVD Rank Intervention): 293+ seeds. Causal mechanism. STRONG. Newly Deployed: 7. Finding #62 (Extended Width): 0 seeds — JUST DEPLOYED. Tests w96-w512 with WD=0.05. Growing / Active: - Finding #49 (WD x LR): 183+ seeds. Interaction confirmed. - Finding #50 (WD Window): 106+ seeds. Diminishing returns supported. - Finding #52 (Percolation): 199+ seeds. ER p_c robustly 0.337-0.343. - Finding #53 (MP Universality GPU): 36+ results. H1 NOT supported. - Finding #54 (WD Timing Scale GPU): 34+ seeds, ALL FALSE (hyperparam artifact). - Finding #56 (WD x Optimizer): 58+ seeds. Consistent with #54. - Finding #47 (Rep Crystallization): 79+ results. Timing confirmed. - Finding #46 (WD Rebound): Preliminary negative. Complete/Retired: - Finding #55 (Grad Subspace): 7 seeds, COMPLETE NEGATIVE. - Finding #57 (Width Mechanism): 83 seeds, INCONCLUSIVE — RETIRED. - Finding #58 (Critical Width): 50 seeds, NEGATIVE — RETIRED. PUBLICATION STATUS — INVERSE CP NOW APPROACHING SUBMISSION ============================================================ The inverse critical period for weight decay story has FOUR confirmed generalizations: 1. WIDTH (#59): 100% at w24-w64, 182 seeds — DEFINITIVE 2. TASK (#60): 100%/100%/96%/87% across 4 tasks, 100 seeds — STRONG 3. DEPTH (#61): 100% d2-d3, 74-85% d4-d6, 66 seeds — STRONG (was 8 last session!) 4. EXTENDED WIDTH (#62): w96-w512, WD=0.05 — JUST DEPLOYED, results pending Key remaining question: Does inverse CP hold at w256+ with matched WD? Finding #62 will answer this. If YES: truly universal. If NO: genuine width boundary exists. KNOWN ISSUES ============= - gigas (h322): SSL cert failure — 24 GPU error results (low priority) - Andre-WEBK (h345): Mixed success rate (outcome=3/exit=202, host-specific) - Orphaned WUs (4,862) from --target_host bug (low priority cleanup) - GPU CUDA arch mismatch on some hosts (Blackwell sm_120+) NEXT SESSION PRIORITIES ======================== 1. Analyze incoming wd_timing_extended_width (#62) results — KEY publication question 2. Continue wd_timing_depth (#61) — 66 seeds, now strongly confirmed 3. Continue wd_timing_task_sweep (#60) — 100 seeds 4. Continue SVD rank intervention (#51) — causal mechanism 5. Monitor GPU experiments (#54 robustness, #53 continuing) 6. Welcome vega (h348) — new host, first appearance in findings Session duration: ~13:20-13:48 UTC, March 3, 2026 Parts completed: Validation (s0303x) + Research/Deployment (s0303y) + Save (this log) Cumulative credited: 4,021 results | 109,211 total results in database