AXIOM BOINC EXPERIMENT REVIEW — Session s0303g
Date: 2026-03-03 ~04:00 UTC
================================================================

CREDIT AWARDED
================================================================
321 results credited, 4,057 total credit awarded (session total).

Per-user credit:
  ChelseaOilman: +2,625
  WTBroughton: +502
  Steve Dodd: +327
  PyHelix: +279
  kotenok2000: +134
  Anandbhat: +76
  [DPC] hansR: +47
  Armin Gips: +42
  Vato: +25

Per-experiment breakdown:
  microscalev2: 44 results (retired experiment, final crediting)
  regtimuniv: 39 results
  compgen: 38 results (retired)
  repcrystal: 32 results (NEW experiment, first results)
  repalignv2: 31 results (retired)
  intervtiming: 30 results
  featcompv2: 24 results (retired)
  regmech: 16 results
  wdrebound: 14 results
  curriculum: 13 results (retired)
  memdynv2: 10 results (retired)
  bottmech: 8 results
  rankreg: 6 results
  bottlemech: 6 results
  neuronspec: 4 results (retired)
  critperiod: 4 results (retired)
  orthocomp: 1 result (retired)
  combinedcomp: 1 result (retired)

CLEANUP ACTIONS
================================================================
- Cleaned up stuck tasks from dead hosts (>12h running, >6h no contact)
- Applied 48h hard ceiling for stuck tasks
- Fixed create_work result generation issue: 51,105 orphaned workunits
  were missing result entries due to a BOINC create_work binary bug.
  Created result entries via direct SQL insertion.
- Aborted ~3,460 unsent results for retired experiments to free capacity.

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

1. REPRESENTATION CRYSTALLIZATION — First results (32 results, all seed=42)
   Predictions tested: (1) Late WD has higher effective rank at epoch 30,
   (2) Early WD collapses diversity early, (3) Crystallization before epoch 30,
   (4) Diversity predicts compositionality.
   RESULT: 2/3 widths support predictions 1-3 (67% support rate).
   HOWEVER: All 32 results used seed=42 due to seed extraction bug.
   These are effectively duplicates — no independent replication yet.
   BUG FIX: Updated seed extraction in representation_crystallization.py
   to use robust 3-method fallback (JSON scan → file scan → hostname+PID).

2. REGULARIZATION TIMING UNIVERSALITY — Growing (39 new results this session)
   Status: Partial universality confirmed. WD and dropout show inverse CP,
   L1 and noise injection do NOT. This tells us the mechanism requires
   weight magnitude compression (WD) or representation redundancy removal
   (dropout), not arbitrary perturbation (noise) or sparsity (L1).

3. NEW EXPERIMENT DEPLOYED: WD Onset Timing Sweep (wd_onset_sweep.py)
   *** NOVEL — first systematic mapping of the full WD timing landscape ***
   Tests WD onset at 10 different epochs (0, 5, 10, 15, 20, 30, 40, 60,
   80, 100) across widths 32, 64, 128. This maps the exact epoch where
   WD transitions from harmful to beneficial for compositionality.
   PREDICTIONS:
   P1: Compositionality gap shows sigmoid-like function of onset epoch
   P2: Transition epoch is earlier for wider networks
   P3: Optimal onset timing beats always-WD
   P4: Transition correlates with effective rank crystallization

   REASONING: Building directly on Finding #45 (Inverse Critical Period,
   34+ seeds). Previous experiments only tested binary WD timing
   (epochs 0-29 vs 30-150). This maps the full landscape to:
   a) Find the exact critical transition epoch
   b) Determine if the transition is sharp or gradual
   c) Test if there's an optimal timing that beats always-WD
   d) Connect the timing to measurable representation metrics

4. SEED EXTRACTION BUG FIX — Fixed in 4 active scripts
   Root cause: The seed extraction code filtered out JSON files containing
   "experiment" in the filename, but BOINC workunit files often contain
   this string. Fix: removed filename filter, added robust fallback chain
   (try JSON files → scan all files → use hostname+PID+time for unique seed).
   Fixed scripts: representation_crystallization.py, regularization_mechanisms.py,
   wd_rebound_dynamics.py, bottleneck_mechanism.py.
   Already fixed: reg_timing_universality.py, intervention_timing_compositionality.py.

DEPLOYMENT
================================================================
3,458 new CPU workunits deployed across 76 hosts.

Experiment mix:
  wd_onset_sweep (NEW): 1,243 WUs (35.9%)
  representation_crystallization: 861 WUs (24.9%)
  reg_timing_universality: 688 WUs (19.9%)
  intervention_timing_compositionality: 352 WUs (10.2%)
  regularization_mechanisms: 162 WUs (4.7%)
  wd_rebound_dynamics: 95 WUs (2.7%)
  bottleneck_mechanism: 57 WUs (1.6%)

Top hosts by deployment:
  DESKTOP-N5RAJSE (192 CPU): +384 WUs
  DadOld-PC (80 CPU): +148 WUs
  SPEKTRUM (72 CPU): +144 WUs
  JM7 (64 CPU): +128 WUs
  Dad-Workstation (80 CPU): +70 WUs
  + 71 more hosts

Total fleet: ~7,200+ experiments queued/in-progress across volunteer network.

NEXT STEPS
================================================================
- Monitor wd_onset_sweep results for the first timing landscape data
- Collect representation_crystallization results with DIVERSE seeds
  (now that seed extraction is fixed)
- Continue accumulating reg_timing_universality seeds
- If wd_onset_sweep confirms a sharp transition epoch, design follow-up:
  - Learning rate × WD timing interaction (does LR affect transition?)
  - WD magnitude × timing interaction (does WD strength matter?)
  - Test whether the optimal timing generalizes to other architectures

KNOWN ISSUES
================================================================
- BOINC create_work binary not generating result entries — workaround:
  direct SQL insertion of result entries. Need to investigate root cause.
- Some old orphaned workunits from retired experiments were given result
  entries; these were subsequently aborted (server_state=5, outcome=5).