Axiom BOINC Session Log — s0303r/s (~18:00-19:00 UTC, March 3, 2026) ===================================================================== PART 1: VALIDATION & CREDIT (Session s0303r, ~18:00 UTC) ========================================================= Credit Awarded: 148 results, ~669 credit, 9 users, 10 hosts ---------------------------------------------------------------------- | User | Credit | Key Hosts & Details | |-----------------|--------|---------------------------------------------------------| | Steve Dodd | +380 | h123 Dads-PC (31 svdrank+perc+wdwidthmech) + h87 (5) | | kotenok2000 | +84 | h29 (25 percolation + wdlr from s0303q deployment) | | Anandbhat | +79 | h222 (3 wdwindow 2012-2025s HEAVY + wdlr + repcrystal) | | ChelseaOilman | +38 | h319 GPU wdtimscale+mpuniv + h330 retired microscalev2 | | matszpk | +24 | h322 gigas (24 GPU SSL errors — all failed) | | PyHelix | +20 | h1 (14 retired + 1 wdtimscale_gpu) | | marmot | +16 | h113 XYLENA (10 retired experiments) | | [DPC] hansR | +14 | h9 (3 wdtimscale_gpu) | | Armin Gips | +12 | h345 Andre-WEBK (1 svdrank 1154s) | Counters updated: credited_count.txt: 2654 -> 2802 (+148) total_results.txt: 159,180 Cleanup: 362 tasks aborted ---------------------------------------------------------------------- - 7 stuck tasks from dead host (Delta-2 h332, 12.6h running, 12.6h no contact) - 355 retired experiment tasks on live hosts: regcomp, rankreg, lottery_ticket, grokking, compgen, wdwidthmech, wdcritwidth, and other experiments retired in previous sessions - 0 broken experiments, 0 >48h tasks New Issue Discovered: gigas (h322, matszpk, Fedora Linux 16-core): SSL certificate verification failure prevents downloading experiment scripts. All 24 GPU results are SSL CERTIFICATE_VERIFY_FAILED errors. Credited minimally. Low priority. KEY SCIENCE FROM VALIDATED RESULTS ---------------------------------------------------------------------- Finding #51 (SVD Rank Intervention): +27 results -> 255+ total 26 from Dads-PC (h123, 807-1370s): ALL P1=TRUE (100%), P2=FALSE (0%). 1 from Andre-WEBK (h345, Armin Gips, 1154s heavy compute). Late SVD truncation CONSISTENTLY reduces gen gap vs baseline (w32/64/128). Perfectly consistent. STRONG CAUSAL EVIDENCE for inverse CP mechanism. Finding #52 (Percolation Phase Transition): +23 results -> 156+ total 11 from Dads-PC (h123, 319-520s), 11 from kotenok2000 (h29, 239-323s), 1 from Dad-Workstation (h87, 580s). ER p_c ~0.343 consistent across all new seeds. ROBUST cross-disciplinary result. Finding #54 (WD Timing Scale GPU): +6 new seeds -> 30+ total, ALL FALSE h319 Dell-XPS (2), h9 dbgrensenh27 (2), h1 Pyhelix (1), +1 prior uncredited. 30+/30+ ALL FALSE. 100% negative consistency. PUBLICATION-QUALITY scope-limiting result: Inverse CP does NOT generalize to w64+. Finding #49 (WD x LR Interaction): +8 results -> 126+ total ALL 8 new results: P3=TRUE (interaction measurable). Completely consistent. Finding #50 (WD Window Duration): +5 results -> 74+ total P4 (diminishing returns): SUPPORTED. 3 HEAVY results from h222 (2012-2025s). Finding #53 (MP Universality GPU): +2 -> 31+ total From Dell-XPS h319 (GTX 1050, ~588-598s). H1 (wider departs later) NOT supported. +5 wdwidthmech (#57, RETIRED), +1 wdrebound (#46), +1 repcrystal (#47, now 69+). PART 2: RESEARCH & DEPLOYMENT (Session s0303s, ~19:00 UTC) ============================================================ New Experiment Designed: Finding #60 (WD Timing Task Generality) ---------------------------------------------------------------------- Script: wd_timing_task_sweep.py CRITICAL FOR PUBLICATION: All 34+ seeds confirming inverse CP (#44) used only ONE task (compositional). This tests whether the effect generalizes across structurally distinct tasks at width 32. Design: - IDENTICAL hyperparams to Finding #44 (WD=0.05, LR=0.01, w32, depth=2, 150 epochs, SGD) - 4 structurally distinct tasks: 1. COMPOSITIONAL (exact #44 replication): group-structured binary classification 2. TEACHER-STUDENT: random teacher network generates labels for 20D input 3. HIDDEN MANIFOLD: low-rank structure in 20D with nonlinear boundary 4. SPIRAL: 2D spiral embedded in 20D with noise features - 4 conditions: no_wd, always_wd, early_wd, late_wd - Run duration: 1800s Novelty check: Builds directly on confirmed Finding #44 (34+ seeds, 100% at w32) and scope-limiting Finding #54 (30+ seeds, ALL FALSE at w64+). No prior experiment tested task generality of inverse CP. All existing results used the same compositional task. This is critical: if inverse CP holds across tasks, it's a GENERAL narrow-network principle suitable for a strong paper. Predictions: P1: inverse_cp holds for compositional (replicates #44) P2: inverse_cp holds for >= 2 other tasks (task generality) P3: task complexity modulates inverse CP strength Deployment Summary: 749 WUs total (735 CPU + 14 GPU), 0 errors ---------------------------------------------------------------------- CPU Experiments (735 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Weight | |--------------------------------|-----|---------|--------| | svd_rank_intervention | 150 | #51 | 25 | | wd_width_transition | 120 | #59 | 20 | | wd_timing_task_sweep (NEW) | 105 | #60 | 18 | | percolation_scaling | 90 | #52 | 15 | | wd_lr_interaction | 72 | #49 | 12 | | wd_window_duration | 54 | #50 | 9 | | wd_optimizer_interaction | 36 | #56 | 6 | | representation_crystallization | 36 | #47 | 6 | | regularization_mechanisms | 30 | — | 5 | | wd_rebound_dynamics | 24 | #46 | 4 | | bottleneck_mechanism | 18 | — | 3 | GPU Experiments (14 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Notes | |--------------------------------|-----|---------|---------------------------------| | wd_timing_scale_gpu | 6 | #54 | 30+ seeds all FALSE, robustness | | mp_universality_test | 8 | #53 | 31+ results, continuing | GPU Checkpoint: Active GPU hosts: 6+ confirmed - Pyhelix (h1): 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) Host Uptake: Dads-PC (h123): Pulled 156 tasks immediately XYLENA (h113): Pulled 45 tasks immediately 548 unsent WUs waiting for remaining 16 active hosts to check in Skipped: Latitude/Athlon (<6GB RAM), Foxtrot-2/3, Delta-2 (dead >12h) Transitioner Note: Errored on exp_percolation_s0303s_r39 (Error 1, likely duplicate key). Re-ran transitioner; all WUs processed and results created successfully. 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%). SCOPED AS NARROW-NETWORK PHENOMENON. 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 #54 (WD Scale Dependence): 30+ seeds, ALL FALSE. Inverse CP does NOT generalize to widths 64+. PUBLICATION-QUALITY scope-limiting result. 4. Finding #51 (SVD Rank Intervention): 255+ seeds, DIVERSE. Late SVD truncation consistently reduces gen gap. STRONG CAUSAL EVIDENCE. 5. Finding #52 (Percolation): 156+ seeds, ER p_c robustly 0.337-0.343. Cross-disciplinary statistical physics result. Growing / Active: - Finding #60 (Task Generality, NEW): 0 seeds, 105 WUs deployed. FIRST DATA expected within 30-60 minutes. - Finding #59 (Width Transition): 0 seeds, 220 WUs total (s0303q+s0303s). Supersedes #57/#58 with correct hyperparams. First data expected soon. - Finding #49 (WD x LR): 126+ seeds. P3 (interaction) at 83%. - Finding #50 (WD Window): 74+ seeds. P4 (diminishing returns) supported. - Finding #53 (MP Universality GPU): 31+ seeds. H1 NOT supported. - Finding #56 (WD x Optimizer): 36+ seeds. Consistent with #54. - Finding #47 (Rep Crystallization): 69+ seeds. Timing confirmed. - Finding #46 (WD Rebound): Preliminary negative. Emerging Theme — Inverse CP Width Dependence: w32: 100% confirmed (34+ seeds in #44) w64+: 0% confirmed (30+ seeds in #54, ALL FALSE) w64/128: ~5-7% per optimizer (36+ seeds in #56) Transition: Finding #59 will pinpoint exact boundary [24-64] Task generality: Finding #60 tests across 4 tasks at w32 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 (tasks cleaned up this session) NEXT SESSION PRIORITIES ======================== 1. Analyze first wd_timing_task_sweep (#60) results — is inverse CP task-general? 2. Analyze first wd_width_transition (#59) results — where does inverse CP break? 3. Continue SVD rank intervention (#51) for publication dataset 4. Monitor GPU experiments (#54 robustness, #53 continuing) 5. Track gigas SSL issue if matszpk reaches out Session duration: ~18:00-19:15 UTC March 3, 2026 Parts completed: Validation (s0303r) + Research/Deployment (s0303s) + Save (this log) Cumulative credited: 2,802 results | 159,180 total results in database