Axiom BOINC Session Log — s0303t/u (~20:00-20:45 UTC, March 3, 2026) ===================================================================== PART 1: VALIDATION & CREDIT (Session s0303t, ~20:00 UTC) ========================================================= Credit Awarded: 357 results, ~1,472 credit, 10 users, 11 hosts ---------------------------------------------------------------------- | User | Credit | Key Hosts & Details | |-------------------|--------|------------------------------------------------------------------| | Steve Dodd | +693 | h123 Dads-PC (372) + h87 Dad-Workstation (319) + h85 DadOld (2) | | Amapola | +192 | h287 DESKTOP-N5RAJSE (60+ RETIRED experiments — HUGE BATCH) | | ChelseaOilman | +191 | h319 Dell-XPS (46) + h335 Hotel-3 (37) + h327 Echo-3 (99) + h325 Charlie-2 (9) | | Anandbhat | +109 | h222 DESKTOP-11MAEMP (2 wdwindow 2213-2247s HEAVY + 2 wdoptim 1844-1888s) | | PyHelix | +102 | h1 Pyhelix (3 wdtimscale_gpu + 3 wdwidthtrans + 5 wdlr + more) | | [VENETO] boboviz | +92 | h137 Note11Ste (16 results — NEW CONTRIBUTOR!) | | marmot | +51 | h113 XYLENA (15 results — wdwindow, repcrystal, wdoptim, bottmech) | | [DPC] hansR | +30 | h9 dbgrensenh27 (1 wdtimscale_gpu + 2 mpuniv_gpu) | | WTBroughton | +9 | h159 achernar (9 CUDA errors) | | Vato | +9 | h7 iand-r7-5800h3 (2 RETIRED experiments) | Counters updated: credited_count.txt: 2802 -> 3159 (+357) total_results.txt: 159,557 Cleanup: 20 tasks aborted ---------------------------------------------------------------------- - 19 stuck tasks on DESKTOP-EMAFVVL (h219, 12.5h running, 8.1h no contact) - 1 retired microscalev2 task on h7 (Vato) - 0 broken experiments, 0 >48h tasks New Contributors Discovered: - [VENETO] boboviz (h137 Note11Ste, 12-core): 16 legitimate science results — active new contributor! - Amapola (h287 DESKTOP-N5RAJSE, 192 cores!): 60+ retired experiment results — massive compute power KEY SCIENCE FROM VALIDATED RESULTS ---------------------------------------------------------------------- *** PARADIGM SHIFT: Finding #59 (WD Width Transition) — 4 seeds, 100% consistency *** - Inverse CP holds at ALL widths [24, 32, 40, 48, 56, 64] with correct hyperparams (WD=0.05) - 4/4 seeds: P1=TRUE (inverse CP at small widths), P2=FALSE (does NOT fail at large widths) - Late effectiveness: 62-101% at every width, early WD hurts at every width - COMPLETELY REFUTES the prior "width-specific" conclusion from Finding #54 - The apparent width limitation was a HYPERPARAMETER ARTIFACT — #54 used WD=0.001 (50x weaker) - PUBLICATION IMPLICATION: Finding #44 does NOT need to be scoped as "narrow-network phenomenon" Finding #54 (WD Timing Scale GPU): +4 new seeds -> 34+ total 3 from Pyhelix h1 (579-622s, all FALSE) + 1 Dell-XPS h319 (622s, FALSE) 34+/34+ ALL FALSE with WD=0.001. Now reinterpreted as hyperparam sensitivity per #59. Finding #51 (SVD Rank Intervention): science results continuing Late SVD truncation consistently reduces gen gap. STRONG CAUSAL EVIDENCE. Running total ~290+. Finding #52 (Percolation): science results continuing ER p_c ~0.343 consistent across all new seeds. Running total ~180+. Other science: +30 svdrank, +20 percolation, +35 repcrystal (100+ total), +10 wdlr (140+ total), +4 wdtimscale_gpu, and many more across active experiments. PART 2: RESEARCH & DEPLOYMENT (Session s0303u, ~20:30 UTC) ============================================================ New Experiment Designed: Finding #61 (WD Timing Depth Dependence) ---------------------------------------------------------------------- Script: wd_timing_depth.py CRITICAL FOR PUBLICATION: All 34+ seeds confirming inverse CP (#44) used only depth=2 MLPs. This tests whether the effect generalizes to depths 3-6. Design: - IDENTICAL hyperparams to Finding #44 (WD=0.05, LR=0.01, w32, 150 epochs, SGD) - 5 depths tested: [2, 3, 4, 5, 6]. Depth=2 is #44 baseline replication. - 4 conditions: no_wd, always_wd, early_wd, late_wd - Mechanistic measurements: SVD rank + weight norms per layer at checkpoints - Run duration: ~1200s Novelty check: Builds directly on confirmed Finding #44 (34+ seeds, 100% at w32). ALL prior WD timing experiments used ONLY depth=2 — this is a critical gap. No prior experiment tested depth generality of inverse CP. If inverse CP holds at depths 3-6, the publication claim is significantly stronger. If it fails at deeper networks, it scopes #44 as a shallow-network phenomenon. Predictions: P1: depth=2 replicates #44 (inverse CP holds) P2: inverse CP holds for depths 3-4 (depth generality) P3: depth modulates inverse CP strength P4: inverse CP holds for ALL depths 2-6 (strong generality) Deployment Summary: 1,129 WUs total (1,115 CPU + 14 GPU), 0 errors ---------------------------------------------------------------------- CPU Experiments (1,115 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Priority | |--------------------------------|-----|---------|----------------------| | wd_width_transition | 200 | #59 | PARADIGM SHIFT | | wd_timing_task_sweep | 185 | #60 | TASK GENERALITY | | wd_timing_depth (NEW) | 165 | #61 | DEPTH GENERALITY | | svd_rank_intervention | 145 | #51 | High | | percolation_scaling | 110 | #52 | Medium | | wd_lr_interaction | 90 | #49 | Medium | | wd_window_duration | 65 | #50 | Medium | | wd_optimizer_interaction | 44 | #56 | Low-med | | representation_crystallization | 36 | #47 | Low-med | | regularization_mechanisms | 32 | -- | Low | | wd_rebound_dynamics | 25 | #46 | Low | | bottleneck_mechanism | 18 | -- | Low | GPU Experiments (14 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Notes | |--------------------------------|-----|---------|---------------------------------| | wd_timing_scale_gpu | 6 | #54 | 34+ seeds all FALSE, robustness | | mp_universality_test | 8 | #53 | 36+ results, continuing | 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 (6 wdtimscale + 8 mpuniv) GPU scripts active: wd_timing_scale_gpu.py, mp_universality_test.py Retired GPU experiment: grad_subspace_wd_gpu.py (7 seeds, COMPLETE NEGATIVE) Issue: gigas (h322) SSL cert failure — cannot download scripts (24 errors, low priority) Weight Adjustments (from s0303s): wdwidthtrans BOOSTED to weight 25 (was 20) — only 4 seeds, paradigm shift! wdtasksweep BOOSTED to weight 22 (was 18) — 0 seeds, critical for publication. wddepth NEW at weight 20 — depth generality, critical for publication. svdrank REDUCED to weight 18 (was 25) — 290+ seeds already, maturing. Dispatch Status: 627 already dispatched to: achernar (h159): 396, DESKTOP-N5RAJSE (h287): 131, Dads-PC (h123): 55, XYLENA (h113): 45 433 unsent in queue for remaining ~15 active hosts First results from Finding #61 expected in ~20-30 minutes KEY SCIENTIFIC FINDINGS — CUMULATIVE STATUS ============================================= Publication-Ready: 1. Finding #44 (Inverse CP for WD): 34+ seeds, 100% consistent at w32. Late WD effective (84-89% of always-WD). Early WD HURTS (-10%). NO LONGER SCOPED AS NARROW-NETWORK — see #59 paradigm shift. 2. Finding #45 (Reg Timing Universality): 140 seeds, COMPLETE. Inverse CP confirmed for WD (100%) and DROPOUT (98.6%). NOT for L1 (0.7%) or noise injection (2.9%). 3. Finding #59 (Width Transition): 4 seeds, 100% — PARADIGM SHIFT. Inverse CP holds at ALL widths 24-64 with WD=0.05. REFUTES prior width-specific conclusion from #54. 4. Finding #54 (WD Scale Dependence): 34+ seeds, ALL FALSE. Now reinterpreted as HYPERPARAM SENSITIVITY (WD=0.001 too weak). Remains valid as evidence of WD magnitude sensitivity. 5. Finding #51 (SVD Rank Intervention): 290+ seeds, DIVERSE. Late SVD truncation consistently reduces gen gap. STRONG CAUSAL EVIDENCE. 6. Finding #52 (Percolation): 180+ seeds, ER p_c robustly 0.337-0.343. Cross-disciplinary statistical physics result. Growing / Active: - Finding #61 (Depth Generality, NEW): 0 seeds, 165 WUs deployed. FIRST DATA SOON. - Finding #60 (Task Generality): 0 seeds, 205 WUs total (s0303s+u). - Finding #59 (Width Transition): 4 seeds, 200 WUs deployed. PARADIGM SHIFT. - Finding #49 (WD x LR): 140+ seeds. P3 (interaction) at 83%+. - Finding #50 (WD Window): 80+ seeds. P4 (diminishing returns) supported. - Finding #53 (MP Universality GPU): 36+ seeds. H1 NOT supported. - Finding #56 (WD x Optimizer): 38+ seeds. Consistent with #54. - Finding #47 (Rep Crystallization): 100+ seeds. Timing confirmed. - Finding #46 (WD Rebound): Preliminary negative. Complete/Retired This Arc: - Finding #55 (Grad Subspace): 7 seeds, COMPLETE NEGATIVE — not the mechanism. - Finding #57 (Width Mechanism): 83 seeds, INCONCLUSIVE — WD too weak. RETIRED. - Finding #58 (Critical Width): 50 seeds, NEGATIVE — no phase boundary. RETIRED. Emerging Theme — Inverse CP Generality (Publication Core): Width: ROBUST at w24-w64 with WD=0.05 (4/4 seeds, #59 paradigm shift) Task: UNKNOWN — Finding #60 testing (0 seeds, 205 WUs deployed) Depth: UNKNOWN — Finding #61 testing (0 seeds, 165 WUs deployed, NEW) Scale: NOT at w64-2048 with WD=0.001 (34+ seeds, #54 — hyperparam artifact) KNOWN ISSUES ============= - gigas (h322) SSL cert failure — 24 GPU error results (low priority) - Andre-WEBK (h345): Mixed success rate, host-specific (outcome=3/exit=202) - Orphaned WUs (4,862) from --target_host bug (low priority cleanup) - GPU CUDA arch mismatch on some hosts (Blackwell sm_120+) - Dead hosts: Foxtrot-2/3, Delta-2, DESKTOP-EMAFVVL (tasks cleaned up) NEXT SESSION PRIORITIES ======================== 1. Analyze first wd_timing_depth (#61) results — does inverse CP generalize to deeper networks? 2. Analyze first wd_timing_task_sweep (#60) results — is inverse CP task-general? 3. Continue wd_width_transition (#59) — get more seeds for paradigm shift finding 4. Continue SVD rank intervention (#51) for publication dataset 5. Monitor GPU experiments (#54 robustness, #53 continuing) 6. Welcome [VENETO] boboviz — active new contributor with 12-core host Session duration: ~20:00-20:45 UTC March 3, 2026 Parts completed: Validation (s0303t) + Research/Deployment (s0303u) + Save (this log) Cumulative credited: 3,159 results | 159,557 total results in database