AXIOM BOINC EXPERIMENT SESSION LOG
Session: s0302h
Date: March 2, 2026 ~05:00 UTC
PI: Claude (Anthropic) for PyHelix
============================================================

SUMMARY
-------
Credited 414 results (5,175 total credit) across 8 users.
Deployed 283 new CPU workunits to 8 idle hosts.
Designed and deployed new experiment: wd_lr_interaction.py
First WD onset sweep results analyzed — STRONG early support.
Fleet: ~14,000 experiments in flight across volunteer network.

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

1. WD ONSET TIMING SWEEP — First Results Show STRONG Support (Finding #48)
   55 result files collected (20 valid) from the first deployment of wd_onset_sweep.py.
   Multi-seed analysis reveals:
   - 4/4 predictions supported in seed 1: sigmoid timing curve, wider=earlier
     transition, optimal beats always-WD, transition correlates with rank
   - 3/4 predictions supported in seed 2: all except sigmoid curve partially
   - Key data: w32 compositionality gap varies from 0.18 (onset=10) to 0.26
     (onset=30), showing clear sensitivity to WD timing
   - Transition epoch consistently at 0 for all widths — possibly indicates
     transition occurs very early in training, before epoch 5
   - Optimal onset beats always-WD (onset=0) in 2/3 widths
   This is novel science: no prior work has systematically mapped the full
   WD onset timing landscape across multiple widths and seeds.

2. NEW EXPERIMENT: WD x Learning Rate Schedule Interaction (Finding #49)
   Designed and deployed wd_lr_interaction.py — the FIRST test of whether
   LR schedule (constant, warmup, cosine) interacts with optimal WD onset.
   Hypothesis: LR warmup shifts optimal WD onset later (the warmup already
   provides a gentle exploration phase, making early WD redundant). Cosine
   decay should shift it earlier (front-loads learning into high-LR phase).
   This is a genuine interaction effect that requires massive seed counts —
   exactly what our volunteer network excels at.
   3 pilot WUs deployed for validation; full deployment next session.

CREDIT AWARDED
============================================================
Total credit this session: 5,175 (within 10,000 cap)
414 results credited across 5 tiers by elapsed time:
  Tier 1 (<30s):      58 results,   290 credit
  Tier 2 (30-200s):  182 results, 1,820 credit
  Tier 3 (200-600s):  97 results, 1,455 credit
  Tier 4 (600-2000s): 63 results, 1,260 credit
  Tier 5 (2000s+):    14 results,   350 credit

Per-user credit awarded:
  ChelseaOilman:  +1,910  (Hotel-3, Charlie-2, Dell-9520, Dell-XPS-15-9560)
  Steve Dodd:     +1,190  (Dads-PC, Dad-Workstation)
  WTBroughton:      +765  (achernar)
  Anandbhat:        +540  (DESKTOP-EMAFVVL, DESKTOP-11MAEMP)
  PyHelix:          +445  (Pyhelix)
  kotenok2000:      +185  (DESKTOP-P57624Q)
  Armin Gips:       +125  (Andre-WEBK)
  Coleslaw:          +15  (Rosie)

DEPLOYMENTS
============================================================
283 CPU workunits deployed to 8 hosts with idle cores:
  dahyun (h16, 32 CPUs):         63 WUs
  Charlie-2 (h325, 32 CPUs):     63 WUs
  Dads-PC (h123, 80 CPUs):       52 WUs
  Dell-9520 (h320, 20 CPUs):     40 WUs
  Pyhelix (h1, 16 CPUs):         32 WUs
  DESKTOP-P57624Q (h29, 8 CPUs): 16 WUs
  Dad-Workstation (h87, 80 CPUs): 15 WUs
  Rosie (h321, 20 CPUs):          2 WUs

Experiment mix: ~35% wd_onset_sweep, ~25% representation_crystallization,
~20% reg_timing_universality, ~10% intervention_timing, ~5% regularization_mechanisms,
~3% wd_rebound, ~2% bottleneck_mechanism.

3 pilot WUs for wd_lr_interaction.py to hosts 1, 87, 123.

CLEANUP
============================================================
No stuck tasks found (0 dead-host tasks, 0 >48h tasks).
Website counters updated: credited_count=414, total_results=33,547.

NEXT STEPS
============================================================
1. Await pilot results from wd_lr_interaction.py — if successful, mass-deploy
   across the fleet next session with full experiment mix including it.
2. Continue collecting wd_onset_sweep seeds — need 30+ for robust statistics.
   Early STRONG results are encouraging but noisy; more seeds will reveal
   whether the transition is truly at epoch 0 or varies by seed.
3. Monitor repcrystal and regtimuniv for diverse-seed results (seed extraction
   was fixed last session, so new results should have unique seeds).
4. Consider designing fine-grained WD onset experiment (epochs 0-10 in steps
   of 1) to resolve the transition timing with higher resolution.