Axiom BOINC Session Log — s0303p/q (~16:00-17:15 UTC, March 3, 2026) ===================================================================== PART 1: VALIDATION & CREDIT (Session s0303p, ~16:00 UTC) ========================================================= Credit Awarded: 222 results, ~858.5 total credit, 8 users, 10 hosts ---------------------------------------------------------------------- | User | Credit | Key Hosts | |-----------------|--------|-------------------------------------------------| | ChelseaOilman | +281 | Echo-3 (h327), Dell-XPS (h319), Delta-1 (h330) | | Steve Dodd | +251 | Dads-PC (h123), Dad-Workstation (h87) | | Armin Gips | +172 | Andre-WEBK (h345): 27 heavy svdrank, 997-1603s! | | kotenok2000 | +133 | DESKTOP-P57624Q (h29): svdrank + wdcritwidth | | [DPC] hansR | +13 | dbgrensenh27 (h9): 3x mpuniv_gpu | | Anandbhat | +5 | DESKTOP-11MAEMP (h222) | | marmot | +3 | XYLENA (h113) | | PyHelix | +1 | Pyhelix (h1) | Counters updated: credited_count.txt: 2432 -> 2654 (+222) total_results.txt: 158,669 Uncredited remaining: 0 Cleanup: 862 tasks aborted ---------------------------------------------------------------------- - 204 stuck tasks from dead hosts: Foxtrot-2 (h339), Foxtrot-3 (h340), Delta-2 (h332), DESKTOP-EMAFVVL (h219) - 658 retired experiment tasks on live hosts: Dell-XPS had 567 alone! Plus DESKTOP-3OOKN65 (23), DESKTOP-11MAEMP (20), XYLENA (19) Experiments cleared: featcompv2, compgen, microscalev2, neuronspec, etc. Key Science from Validated Results ---------------------------------------------------------------------- Finding #51 (SVD Rank Intervention): +81 results -> 228+ total Andre-WEBK (h345, Armin Gips) contributed 27 heavy svdrank results (997-1603s each!) Plus Dads-PC (h123), Echo-3 (h327), DESKTOP-P57624Q (h29). Late SVD truncation CONSISTENTLY reduces gen gap vs baseline across w32/64/128. P1 (late SVD improves gap): SUPPORTED in virtually all seeds. STRONG CAUSAL EVIDENCE. Finding #54 (WD Timing Scale): +2 new GPU seeds -> 24+ total, ALL FALSE Seed 857746014 (Dell-XPS GTX 1050, 621s): FALSE, scale_corr=-0.77 Seed 857753717 (Dell-XPS GTX 1050, 568s): FALSE, wildly inconsistent 24+/24+ ALL FALSE. PUBLICATION-QUALITY scope-limiting result. Inverse CP does NOT generalize to widths 64+. Finding #58 (WD Critical Width): +20 s0303n seeds -> 50+ total NO consistent phase boundary: rates random 15-60% across all widths. w32 only 15%, w80 = 60%. Pure noise. Essentially COMPLETE (negative). Finding #52 (Percolation): +21 -> 133+ total ER p_c remains ~0.337-0.343 across all seeds (vs theory 0.25). ROBUST. Finding #53 (MP Universality): +3 GPU -> 29+ total H1 (wider departs later) still NOT supported (correlation=0.0). Finding #57 (WD Width Mechanism): +27 -> 83+ total Still noisy, hyperparameter-limited (WD=0.0001 too weak). Finding #49 (WD x LR): +11 -> 118+ total P3 (interaction measurable): 83% support. CONFIRMED. PART 2: RESEARCH & DEPLOYMENT (Session s0303q, ~17:15 UTC) ============================================================ New Experiment Designed: Finding #59 (WD Width Transition) ---------------------------------------------------------------------- Script: wd_width_transition.py SUPERSEDES Finding #57 (wd_width_mechanism, 83 seeds, RETIRED INCONCLUSIVE) SUPERSEDES Finding #58 (wd_critical_width, 50 seeds, RETIRED NEGATIVE) Root cause identified: #57/#58 used WD=0.0001, which is 500x weaker than Finding #44's WD=0.05. This made the inverse CP undetectable even at w32 (only 15-30% vs 100% in #44). Design: - IDENTICAL hyperparams to Finding #44 (WD=0.05, LR=0.01, compositional task, depth=2, 150 epochs) - Fine-grained width sweep: [24, 32, 40, 48, 56, 64] - 4 conditions: no_wd, always_wd, early_wd, late_wd - Mechanistic measurements: SVD effective rank + weight Frobenius norms per checkpoint - Run duration: 2700s (45 min) Novelty check: Builds directly on confirmed Finding #44 (34+ seeds, 100% consistent at w32) and scope-limiting Finding #54 (24+ seeds, ALL FALSE at w64+). No prior work identifies the EXACT width where inverse CP transitions from present to absent. #57/#58 attempted this but were compromised by weak hyperparams. This is the definitive experiment: matched hyperparams + fine-grained sweep. Predictions: P1: inverse_cp holds at w24, w32 (replicating #44) P2: inverse_cp weakens/fails at w48-64 (consistent with #54) P3: Mechanistic measurements change at transition width Deployment Summary: 549 WUs total, 0 errors ---------------------------------------------------------------------- CPU Experiments (535 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Priority | Weight | |-------------------------------|-----|---------|----------|--------| | svd_rank_intervention | 130 | #51 | Highest | 25 | | wd_width_transition (NEW) | 100 | #59 | Highest | 20 | | percolation_scaling | 80 | #52 | High | 15 | | wd_lr_interaction | 60 | #49 | Medium | 12 | | wd_window_duration | 45 | #50 | Medium | 9 | | wd_optimizer_interaction | 30 | #56 | Medium | 6 | | representation_crystallization| 30 | #47 | Medium | 6 | | regularization_mechanisms | 25 | -- | Low | 5 | | wd_rebound_dynamics | 20 | #46 | Low | 4 | | bottleneck_mechanism | 15 | -- | Low | 3 | GPU Experiments (14 WUs, all untargeted BOINC-scheduled): | Experiment | WUs | Finding | Notes | |-------------------------------|-----|---------|-----------------------------| | wd_timing_scale_gpu | 6 | #54 | 24+ seeds all FALSE, robustness | | mp_universality_test | 8 | #53 | 29+ results, continuing | GPU Checkpoint: Active GPU hosts: 6+ confirmed (Pyhelix RTX 3080, thonon-meylan RTX 4060 Ti, Dell-XPS GTX 1050 x2, W10-Home/Henk GPU, dbgrensenh27/hansR Quadro M4000M, DESKTOP-P57624Q/kotenok2000 GPU) GPU workunits this session: 14 GPU scripts: wd_timing_scale_gpu.py, mp_universality_test.py Retired GPU experiment: grad_subspace_wd_gpu.py (7 seeds, COMPLETE NEGATIVE) Immediate uptake: 169/549 WUs dispatched to hosts within minutes. 380 unsent results waiting in queue for hosts to check in. Active hosts needing work: Dad-Workstation (78), Dell-9520 (60), Hotel-3 (47), achernar (36), Charlie-1/2 (32 each), DESKTOP-3OOKN65 (29), Pyhelix (28), Delta-1 (27), Echo-3 (26), DESKTOP-P57624Q (24), Dell-XPS (16), Andre-WEBK (15), Thonon-PC (12), plus semi-active hosts. Experiments Retired This Session ---------------------------------------------------------------------- Finding #57 (wd_width_mechanism): 83+ seeds, COMPLETE INCONCLUSIVE Hyperparams too weak (WD=0.0001 vs #44's WD=0.05). Could not reliably produce inverse CP even at w32. Mechanistic hypotheses untestable. Superseded by Finding #59. Finding #58 (wd_critical_width): 50+ seeds, COMPLETE NEGATIVE No phase boundary found. Rates random 15-60% across all widths. Same hyperparameter issue as #57. Superseded by Finding #59. Key Discovery: Orphaned Workunit Bug ---------------------------------------------------------------------- 4,862 orphaned WUs found in database: regmech (1703), repcrystal (1382), wdrebound (945), bottmech (832) All have WU records but ZERO result records. Caused by --target_host bug creating WU records without result records. The queue was actually COMPLETELY EMPTY for all active experiments. This explains why hosts appeared idle despite "active" WU counts. KEY SCIENTIFIC FINDINGS SUMMARY ================================ Publication-Ready Results: 1. Finding #44 (Inverse CP for WD): 34+ seeds, 100% consistent at w32. Late WD is effective (84-89% of always-WD). Early WD HURTS (-10%). SCOPED AS NARROW-NETWORK PHENOMENON (see #54). 2. Finding #45 (Regularization 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): 24+ seeds, ALL FALSE. 100% consistency. Inverse CP does NOT generalize to widths 64+. Publication-quality scope-limiting result for #44. 4. Finding #51 (SVD Rank Intervention): 228+ seeds, STRONG CAUSAL EVIDENCE. Late SVD truncation consistently reduces generalization gap. Critical for explaining the mechanism behind inverse CP. Growing / Active: - Finding #59 (WD Width Transition): 100 WUs deployed, 0 results yet. FIRST DATA EXPECTED within 30-60 minutes. Will characterize exact transition width with matched hyperparams. - Finding #52 (Percolation): 133+ seeds. ER p_c robustly 0.337-0.343. - Finding #49 (WD x LR): 118+ seeds. Interaction confirmed (83%). - Finding #53 (MP Universality): 29+ GPU seeds. H1 NOT supported. Emerging Theme — Width Dependence of Inverse CP: w32: 100% confirmed (34+ seeds in #44) w64+: 0% confirmed (24+ seeds in #54, ALL FALSE) w64/128: ~5-7% per optimizer (36+ seeds in #56) Transition: Finding #59 will pinpoint the exact boundary [24-64] KNOWN ISSUES ============= - Orphaned WUs (4,862) from --target_host bug. Low priority cleanup. - Andre-WEBK (h345): Mixed success rate, host-specific failures (outcome=3/exit=202). - GPU CUDA arch mismatch on some hosts (Blackwell sm_120+). Needs binary rebuild. - Dead hosts: Foxtrot-2, Foxtrot-3, Delta-2, DESKTOP-EMAFVVL (tasks cleaned up). NEXT SESSION PRIORITIES ======================== 1. Analyze first wd_width_transition (#59) results — critical for width boundary 2. Continue SVD rank intervention (#51) seed collection for publication 3. Monitor GPU experiments (#54 robustness, #53 continuing) 4. Consider orphaned WU cleanup (4,862 dead records) Session duration: ~16:00-17:30 UTC March 3, 2026 Parts completed: Validation (s0303p) + Research/Deployment (s0303q) + Save (s0303r)