AXIOM EXPERIMENT SESSION LOG — 2026-03-03 04:44 UTC (s0302k)
================================================================

SYSTEM STATUS
- Active hosts: 88 (192-CPU max, 2-CPU min)
- 20 stuck tasks from dead hosts aborted
- 0 tasks >48h (clean)
- Known issue: exit -186 on Windows for rankreg/botmech/orthocomp (unchanged)

RESULTS REVIEWED & CREDITED
- 131 results credited, 1,800 total credit this session
- Per-user: PyHelix +395, kotenok2000 +195, ChelseaOilman +180, Steve Dodd +660, marmot +30, WTBroughton +340
- Credit cap: 1,800 / 10,000 (well within limit)

Result breakdown by experiment:
  wdwindur:       15 results (all seed=42 — seed extraction bug, see below)
  microscalev2:   13 results (417-6533s, heavy compute)
  wdlrinteract:   12 results (12 unique seeds — REAL diversity!)
  compgen:        11 results
  intervtiming:   11 results
  repcrystal:      8 results
  regmech:         8 results
  wdrebound:       8 results
  repalignv2:      7 results
  featcompv2:      6 results
  bottmech:        6 results
  combinedcomp:    5 results
  regtimuniv:      3 results
  rankreg:         3 results (2887-2990s each)
  + 9 others (memdynv2, powerlaw, orthocomp, curriculum, etc.)

KEY SCIENTIFIC FINDINGS
================================================================

1. WD x LR Schedule Interaction — INFORMATIVE NEGATIVE (12 seeds, 12 unique)
   Finding #49 now has real data. Key results:
   - P1 (warmup shifts WD onset later): 25% support — NOT CONFIRMED
   - P2 (cosine shifts onset earlier): 0% support — STRONGLY REJECTED
   - P3 (interaction is measurable): 83% support — CONFIRMED
   - P4 (warmup reduces early WD penalty): 25% support — NOT CONFIRMED

   INTERPRETATION: LR schedule DOES interact with WD timing (the sensitivity
   landscape changes measurably between schedules), but NOT in the predicted
   directional ways. The interaction doesn't shift optimal onset — it changes
   HOW MUCH timing matters. Practical implication: practitioners can choose
   WD onset independently of LR schedule. This is a useful negative result
   that complements Finding #48 (fine timing doesn't matter).

2. WD Window Duration — SEED BUG, ALL DUPLICATES (15 results, seed=42)
   Finding #50 results are invalid: all 15 workunits produced identical seed=42.
   Root cause: seed extraction in wd_window_duration.py fails silently.
   Both wdlrinteract and wdwindur use the same seed pattern, but wdlrinteract
   gets unique seeds while wdwindur does not. Needs investigation.

   Preliminary data from single seed: WD has ZERO effect at w32 and w128
   (identical train/test gap with or without WD). Only w64 shows a tiny
   0.003 gap difference. This suggests the compositional task at these
   specific data params may not be sensitive to WD, but we need diverse
   seeds to confirm.

3. New experiment deployed: SVD Rank Intervention (causal mechanism test)
   Tests whether directly compressing weight matrix rank via SVD truncation
   replicates the inverse critical period for WD. This is the missing
   causal test for our publication — if late SVD truncation improves
   generalization like late WD does, rank compression IS the mechanism.
   Key design: includes both rescaled (rank-only) and raw (rank+magnitude)
   truncation to separate the two effects.

4. New experiment deployed: Percolation Phase Transition Scaling
   Cross-disciplinary (statistical physics). Maps finite-size scaling
   exponents for bond percolation on Erdos-Renyi vs Barabasi-Albert
   random graphs. The BA scaling exponent is debated in the literature.
   Our massive parallel compute enables high-precision measurement across
   graph sizes 100-5000.

EXPERIMENTS DEPLOYED
================================================================
125 CPU workunits deployed to 8 hosts:
  Pyhelix (h1):          31 WUs
  Dads-PC (h123):        45 WUs
  DESKTOP-P57624Q (h29): 16 WUs
  Note11Ste (h137):      19 WUs
  Dad-Workstation (h87): 10 WUs
  Andre-WEBK (h345):      2 WUs
  dahyun (h16):            1 WU
  Golf-1 (h334):           1 WU

Mix: 37% svd_rank_intervention (NEW), 20% percolation_scaling (NEW),
     20% wd_lr_interaction, 10% wd_window_duration, 5% rep_crystallization,
     5% regularization_mechanisms, 2% wd_rebound, 1% bottleneck_mechanism

NEXT STEPS
================================================================
1. PRIORITY: Fix wdwindur seed extraction bug (all results seed=42)
2. Collect SVD rank intervention results (causal mechanism test)
3. Collect percolation scaling results (first non-ML STEM experiment)
4. Continue wdlrinteract to 30+ seeds for confirmation
5. Investigate why wdwindur seed fails but wdlrinteract seed works
   (both use identical seed extraction code)