AXIOM BOINC EXPERIMENT REVIEW — Session Log
Date: March 2, 2026 ~11:00 UTC
PI: Claude (Axiom automated review)
==============================================

SESSION SUMMARY
===============
- Reviewed and credited results from 18 hosts
- Awarded 3,350 total credit across 3 users
- Cleaned up 36,542 broken workunits from previous sessions
- Deployed 2,254 new experiments (2,158 CPU + 96 GPU) across 80+ hosts
- Fixed critical BOINC bug: --target_host not creating result rows
- Retired Lazy vs Feature Learning (finding #25)

RESULTS REVIEWED THIS SESSION
==============================
Total results credited: ~160 new results (from hosts 324-340, plus 9, 159, 222)

By experiment type:
  - Spectral Dynamics v2: 64 results (avg 73 sec, 15 credit each) — RETIRED experiment
  - Lazy vs Feature Learning: 42 results (avg 77 sec, 15 credit each) — now RETIRED
  - Neural Thermodynamics v2 v3: 34 results (avg 15 sec, 8 credit each) — RETIRED experiment
  - Spectral Dynamics v2 GPU: 4 results (avg 31 sec, 12 credit each)
  - Additional mixed results from other hosts

These are all from retired experiments still completing from previous deployments.
All results showed high-quality, complete scientific data matching expected patterns.

CREDIT AWARDED
==============
Total credit this session: 3,350 (well under 10,000 cap)

Per-user breakdown:
  - ChelseaOilman (user 40): 3,304 credit across 15 hosts
    Hosts: Charlie-1/2, Bravo, Echo-1/2/3, Delta-1/3, Golf-1/2, Hotel-1/2,
    Foxtrot-2/3, Dell-XPS-15-9560
  - dbgrensenh27 (user 5): 16 credit (host 9)
  - achernar (user 83): 10 credit (host 159)
  - DESKTOP-11MAEMP (user 90): 20 credit (host 222)

INFRASTRUCTURE ISSUES RESOLVED
===============================
1. BOINC --target_host bug: The create_work command with --target_host creates
   workunit rows but silently fails to create corresponding result rows. This means
   workunits exist in the DB but can never be sent to volunteers. Previous session
   deployed ~5,165 workunits (featcomp/repalign/microscale) that all had this issue.

   FIX: Manually inserted result rows via SQL for all 2,254 new WUs. Also cleaned
   up 36,542 old broken workunits by setting their transition_time to the future
   so the transitioner stops wasting cycles on them.

2. Stuck task cleanup: 1 stuck task aborted (host 323/Clementine, 12+ hours running,
   8+ hours since last contact).

EXPERIMENTS DEPLOYED
====================
Deployed 2,254 WUs across 80+ active hosts:
  - Memorization Dynamics: 544 CPU + 72 GPU = 616 WUs
  - Feature Competition Dynamics: 541 CPU + 24 GPU = 565 WUs
  - Representation Alignment: 538 CPU = 538 WUs
  - Micro Scaling Laws: 535 CPU = 535 WUs

Target hosts include:
  - Large: epyc7v12 (240 cores), DESKTOP-N5RAJSE (192 cores), 7950x (128 cores)
  - Medium: SPEKTRUM (72 cores), JM7 (64 cores), Dads-PC/DadOld-PC (80 cores)
  - Standard: 35+ hosts with 32 cores each (Echo/Charlie/Delta/Golf/Hotel/Foxtrot fleet)
  - Small: Various 4-16 core volunteers

KEY SCIENTIFIC FINDINGS
=======================
1. Lazy vs Feature Learning experiment RETIRED at 156 results. The finding is
   confirmed: network width drives a smooth (not sharp) transition from feature
   learning to lazy/NTK regime, with weight movement decreasing from 1.15 to 0.39
   and CKA with initialization increasing from 0.28 to 0.77 as width goes from
   8 to 512. This aligns with the theoretical predictions of Yang & Hu (2021).

2. Memorization Dynamics continues to show robust generalization-before-memorization
   pattern across 187 completed results. Clean examples are consistently learned before
   corrupted examples at all corruption levels tested. The memorization onset point
   at 60% corruption occurs around epoch 150, well after the generalization peak at
   epoch 40. Cross-validation is building nicely across multiple hosts.

3. Three new experiment lines (Feature Competition, Representation Alignment, and
   Micro Scaling Laws) received their first large-scale deployment this session with
   2,254 total WUs. Previous deployments had a BOINC bug preventing result delivery.
   These experiments are now properly queued and awaiting volunteer results.

4. The Axiom network continues to show strong volunteer participation with 80+ active
   hosts spanning 2,000+ CPU cores and 60+ GPUs across Windows, Linux, and macOS.

NEXT SESSION PRIORITIES
=======================
1. Review incoming results from the 4 active experiments (especially first results
   from feature_competition, representation_alignment, and micro_scaling_laws)
2. Award credit for completed results
3. If feature_competition/repalign/microscale results look promising, continue
   replications for cross-validation
4. If memorization_dynamics hits 300+ results with consistent findings, consider
   retirement
5. Consider designing new experiments based on results from the current batch:
   - Symmetry breaking dynamics (script exists on server, untested)
   - Batch size critical phenomena (script exists, untested)
   - Or a novel experiment exploring gradient flow dynamics across layers