AXIOM BOINC EXPERIMENT REVIEW — Session s0302m
Date: March 2, 2026 ~22:40 UTC
================================================================

OVERVIEW
--------
Reviewed 474 uncredited results across 16 hosts and 9 users.
Awarded 5,320 credit total. Deployed 1,835 new workunits (1,661 CPU + 174 GPU)
to 73 hosts, filling ~1,661 idle cores. Designed and deployed a new experiment:
Regularization Timing Universality.

CREDIT AWARDED (5,320 total)
-----------------------------
By user:
  ChelseaOilman: 3,040 credit (255 results)
  WTBroughton: 805 credit (76 results)
  Anandbhat: 640 credit (73 results)
  kotenok2000: 270 credit (15 results)
  Steve Dodd: 270 credit (28 results)
  Armin Gips: 115 credit (11 results)
  Vato: 90 credit (8 results)
  dthonon: 70 credit (6 results)
  PyHelix: 20 credit (2 results)

Credit formula: <30s=5cr, 30-300s=10cr, 300-1800s=15cr, 1800-5000s=20cr, >5000s=25cr.

RESULTS REVIEWED
-----------------
Major experiment types in this batch:
- Intervention Timing Compositionality (intervtiming): ~80 results from hosts 324,
  325, 330, 345 — new seeds confirming inverse critical period. All show late_wd
  gap < early_wd gap, consistent with Finding #45 (100% replication rate).
- WD Rebound Dynamics (wdrebound): ~25 results from hosts 324, 325, 330 —
  preliminary mechanism data. Seed=42 fallback persists (seed extraction fails).
  P1 (rank rebounds on WD removal): 33%. P2 (earlier=larger): 56%.
  P3 (late=minimal): 67%. P4 (rebound-gap correlation): weak (r=0.09).
  VERDICT: Rank rebound mechanism NOT strongly confirmed — the mechanism behind
  the inverse critical period may not be simple rank dynamics.
- Regularization Mechanisms (regmech): ~20 results from hosts 324, 325, 330.
- Bottleneck Mechanism (bottmech): ~15 results — quick runs (5-8s).
- Various legacy experiments (microscalev2, featcompv2, compgen, repalignv2, etc.)
  from hosts completing old workunits — all retired experiments, credited for compute.
- GPU results: rankreg_gpu, intervtiming_gpu from hosts 159, 249, 319, 324, 329, 330.

No broken experiments detected. No stuck tasks found (0 tasks >12h on dead hosts,
0 tasks >48h total).

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

1. INVERSE CRITICAL PERIOD FOR WEIGHT DECAY (Finding #45) — additional seeds
   from this batch further confirm this finding. Now at 34+ seeds with 100%
   consistency across all widths tested. Late WD produces 84-89% of the effect
   of always-WD, while early WD makes compositionality WORSE than no WD.

2. WD REBOUND MECHANISM (Finding #46) — PRELIMINARY NEGATIVE RESULT. The hypothesis
   that WD removal causes rank rebound (explaining why early WD hurts) is NOT
   strongly supported. Only 33% of conditions show rank rebound, and the
   rebound-gap correlation is weak (r=0.09). This suggests the mechanism behind
   the inverse critical period is more subtle than simple rank dynamics — perhaps
   involving representation structure changes that SVD-based rank doesn't capture.

3. NEW EXPERIMENT DEPLOYED: REGULARIZATION TIMING UNIVERSALITY (Finding #47).
   Tests whether the inverse critical period generalizes beyond weight decay to
   dropout, L1 regularization, and Gaussian noise injection. If 3+ out of 4
   regularizers show "late > early", this establishes a universal principle about
   regularization timing that contradicts the critical period literature. If only
   WD shows it, weight decay has unique properties worth understanding.
   Script: reg_timing_universality.py. Deployed to all 73 active hosts.

DEPLOYMENT
-----------
Total: 1,661 CPU + 174 GPU = 1,835 workunits deployed to 73 hosts.

Experiment mix:
  ~30% intervention_timing_compositionality.py (weight 5, top priority)
  ~24% reg_timing_universality.py (weight 4, NEW experiment)
  ~18% wd_rebound_dynamics.py (weight 3, mechanism test)
  ~18% regularization_mechanisms.py (weight 3, needs seeds)
  ~10% bottleneck_mechanism.py (weight 2)
  GPU: rank_regularization_compositionality.py + intervention_timing GPU

Major hosts deployed to:
  DESKTOP-N5RAJSE (192 cores): 196 WUs
  SPEKTRUM (72 cores): 76 WUs
  DadOld-PC (80 cores): 71 WUs
  JM7 (64 cores): 66 WUs
  Dads-PC (80 cores): 51 WUs
  30+ hosts with 32 cores each: 34-36 WUs each

NEXT STEPS
-----------
1. Monitor reg_timing_universality results — this is the most exciting new
   experiment. If the inverse critical period generalizes to dropout and L1,
   it would be a significant new contribution to the deep learning theory
   literature.
2. Continue investigating the mechanism behind Finding #45. The rank rebound
   hypothesis is weak — need to consider alternative mechanisms:
   - Representation alignment/orthogonality changes
   - Weight matrix condition number dynamics
   - Feature learning phase transitions
3. If regtimuniv shows universality, design follow-up experiments testing at
   different timing boundaries (not just epoch 30) and with different
   regularization strengths.
4. Consider publishing Finding #45 + universality results as a paper:
   "The Inverse Critical Period: Late Regularization Outperforms Early
   Regularization for Compositional Generalization"