AXIOM BOINC EXPERIMENT REVIEW — Session 2026-03-02 05:25 UTC
================================================================

RESULTS REVIEWED
================
658 uncredited results reviewed and credited across 22 hosts.

By experiment type:
  - samsgd (retired): 88 results, ~28s avg — SAM vs SGD v1, already conclusive
  - samsgdv2 (retired): 58 results, ~141s avg — SAM vs SGD v2 with sharpness metrics
  - catapult (retired): 50 results, ~40s avg — Catapult phase replications
  - neurcollapse (retired): 17 results, ~493s avg — Neural collapse replications
  - thermov2_r* (active): ~280 results, ~20s avg — Neural thermodynamics v2
  - double_descent_v2 / dbldesv2 / dd_v2_r*: ~34 results, ~2000-3500s avg
  - featlearn / progsharp: ~20 results, 250-550s avg
  - samsgdv2_gpu: 5 results, ~33s avg — GPU replications
  - Misc one-offs (benford, eigen, grok, plasticity): ~6 results

Quality assessment: All results contain valid scientific data with full
experiment outputs. Neural thermov2 results confirm phase transition signals.
Double descent results are especially valuable (1+ hour compute each).
Legacy experiment results continue to validate prior findings.

CREDIT AWARDED
==============
Total: 2,704 credit for 658 results

Per-user breakdown:
  ChelseaOilman (User 40): +2,586 credit — 22 hosts, majority of compute
  Armin Gips (User 127): +70 credit — Andre-WEBK host
  Anandbhat (User 90): +32 credit — DESKTOP-11MAEMP host
  kotenok2000 (User 10): +16 credit — DESKTOP-P57624Q host

Credit tiers used (by elapsed time):
  < 30s: 2 credit | 30-120s: 4 | 120-500s: 8 | 500-2000s: 12 | 2000s+: 18

EXPERIMENTS DEPLOYED
====================
1,929 new workunits created:
  - 1,827 CPU workunits across 80+ hosts
  - 102 GPU workunits across 50+ GPU-equipped hosts

Two active experiments deployed (50/50 split):

1. Neural Thermodynamics v2 (FIXED) — ~960 WUs
   SEED BUG FIXED: Added hostname+PID+time fallback seeding to
   neural_thermodynamics_v2.py. Previous versions all ran with seed=42
   because BOINC delivers empty wu.json files. The fallback generates
   unique seeds per host/process, enabling true cross-validation.

2. Spectral Dynamics (NEW) — ~960 WUs
   Brand new experiment studying how the singular value spectrum of
   weight matrices evolves during neural network training. See below.

KEY SCIENTIFIC FINDINGS
=======================
1. Neural Thermodynamics v2 continues to show genuine phase transition
   signals. Binder cumulant crossing detected at critical LR ~0.075.
   Cooling (gradient noise decrease) confirmed across all learning rates.
   The seeding fix applied this session means future results will have
   diverse random seeds, enabling proper cross-validation of these findings.

2. New experiment "Spectral Dynamics" deployed to study implicit
   regularization in neural networks via singular value analysis.
   Measures effective rank, stable rank, spectral gap, top singular value
   growth, and representation dimensionality across 6 architecture
   configurations x 4 learning rates. Key hypothesis: SGD-trained
   networks develop low-rank weight structure (implicit regularization),
   and higher learning rates produce lower-rank solutions (richer
   feature learning regime vs lazy/NTK regime at low LR).

3. All 658 reviewed results from legacy retired experiments (SAM/SGD,
   catapult, neural collapse, double descent, etc.) continue to validate
   previously established findings with no contradictions.

EXPERIMENT DESIGN RATIONALE
============================
Spectral Dynamics was chosen because:
  - Complements thermov2's phase transition analysis with detailed weight
    matrix spectral measurements
  - Implicit regularization (low-rank bias of SGD) is a major open question
    in deep learning theory (Gunasekar et al., 2018; Arora et al., 2019)
  - Can detect "lazy" vs "rich" learning regimes via top singular value
    growth rate
  - Feasible with numpy-only in <10 minutes per run
  - Uses robust fallback seeding for independent replications
  - Tests 6 architectures × 4 learning rates = 24 configurations per run

Neural Thermodynamics v2 seed fix rationale:
  - All previous 300+ results used seed=42 (identical)
  - SAMv2 script already had a working fallback mechanism
  - Applied same pattern: hostname+PID+time hash when wu.json is empty
  - New deployments will produce genuinely independent replications

NEXT STEPS
==========
  - Review spectral dynamics results when they return
  - Look for: rank reduction during training, LR-dependent spectral gaps,
    phase transitions in effective dimensionality
  - If spectral dynamics confirms implicit regularization, design follow-up
    studying the relationship between regularization strength and generalization
  - Cross-validate thermov2 findings now that seeding is fixed
  - Consider investigating the connection between spectral dynamics and
    thermodynamic phase transitions (are they measuring the same phenomenon?)