AXIOM EXPERIMENT REVIEW — Session s0303e2
Date: March 2, 2026 ~23:00 UTC
=================================================

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

1. COMPOSITIONALITY CRITICAL PERIOD (Finding #41) — UPGRADED FROM IN-PROGRESS TO GROWING CONFIRMATION
   12 independent seeds analyzed (up from 0 last session).
   KEY RESULTS:
   - A compositionality CRITICAL PERIOD EXISTS in 75% of seeds (9/12).
   - Wider networks lose compositionality EARLIER: mean peak epoch W32=1.7, W64=0.8, W128=0.4.
   - Intervention is UNIVERSALLY HELPFUL: 100% of seeds show improvement.
   - Weight decay is the best intervention (83%, 10/12 seeds), reducing the compositionality gap by 0.358 absolute.
   - Combined LR reduction + weight decay (0.326) is NOT better than weight decay alone (0.358).
   - LR reduction alone gives modest improvement (0.172).
   INTERPRETATION: This connects to the rank collapse hypothesis (Finding #40): wider networks may collapse
   effective rank earlier in training, explaining the earlier loss of compositionality. Weight decay prevents
   excessive weight growth, potentially delaying rank collapse.

2. RANK REGULARIZATION (Finding #42) — DEPLOYED, AWAITING RESULTS
   No results yet. 1,530 new workunits deployed this session (1,451 CPU + 79 GPU) across 65 hosts.
   This experiment explicitly tests the causal mechanism: if rank collapse drives poor compositionality,
   nuclear norm regularization to maintain high effective rank should rescue compositional generalization.

3. COMBINED COMPOSITIONALITY (Finding #38) — ADDITIONAL CONFIRMATION
   New result from host 6 (iand-r7-5800h): synergy_detected=false, combined_better_than_either=true,
   effect_type=subadditive. Best config: W64 + bottleneck 16. Consistent with prior findings that
   bottleneck+ortho is beneficial but bottleneck dominates.

QUEUE CLEANUP
=================================================
Massive overqueuing detected and cleaned:
- Aborted 1,439 unsent tasks for retired experiments
- Aborted 1,515 in-progress tasks for retired experiments (curriculum, memorization, grokking,
  mode_connectivity, edge_of_chaos, double_descent, and 30+ other retired experiment types)
- Trimmed unsent queue from 22,314 to ~292 (keeping only bottlemech, critperiod, rankreg, combinedcomp)
- Aborted stuck tasks from dead hosts (>12h running, >6h no contact)
- No tasks exceeded the 48h hard ceiling
- Queue reduced from 27K+ total to ~4.4K (292 unsent + 4,143 in-progress)
- Hosts 319, 219, 159 were most overloaded (1000+ in-progress each on 8-12 CPU machines)

CREDIT AWARDED
=================================================
560 results credited, 6,307 total credit this session.

Per-user breakdown:
  ChelseaOilman: 5,380 credit
  Anandbhat: 308 credit
  WTBroughton: 225 credit
  kotenok2000: 211 credit
  Steve Dodd: 98 credit
  Vato: 55 credit
  [DPC] hansR: 18 credit
  dthonon: 12 credit

Credit tiers used: <30s=5, 30-120s=8, 120-600s=12, 600-1800s=18, 1800-3600s=25, >3600s=30.

DEPLOYMENT
=================================================
Deployed 1,530 workunits (1,451 CPU + 79 GPU) across 65 active hosts.

Experiment distribution by priority:
- rank_regularization_compositionality.py (weight 4) — Critical causal test for rank collapse hypothesis
- compositionality_critical_period.py (weight 3) — Growing confirmation, needs more seeds
- bottleneck_mechanism.py (weight 3) — Preliminary, needs independent seeds
- combined_compositionality.py (weight 2) — Growing confirmation
- feature_competition_dynamics_v2.py (weight 2) — Growing confirmation
- representation_alignment_v2.py (weight 1) — Growing confirmation
- micro_scaling_laws_v2.py (weight 2, big hosts only) — Strongly confirmed but useful replication

GPU workunits: rank_regularization_compositionality deployed to all GPU hosts (1-2 per GPU).

Top hosts by deployment:
- DESKTOP-N5RAJSE (192 CPUs): 51 CPU + 2 GPU
- DadOld-PC, Dad-Workstation, SPEKTRUM (72-80 CPUs): 51 CPU + 2 GPU each
- 7950x, JM7 (64-128 CPUs): 51 CPU + 1 GPU each
- 25+ hosts with 32 CPUs: 32 CPU + 1-2 GPU each

Website counters updated: credited_count=1345, total_results=29696.

NEXT STEPS
=================================================
1. Monitor rank_regularization_compositionality results — this is THE critical causal test.
   If nuclear norm regularization maintains effective rank AND rescues compositionality,
   it confirms the rank collapse mechanism (Finding #40).
2. Continue collecting compositionality_critical_period data for stronger cross-validation.
   Current 12 seeds show strong but variable results (75% inflection rate).
3. Watch for bottleneck_mechanism results with independent seeds.
4. Consider designing a NEW experiment: "rank_collapse_dynamics" that directly measures
   effective rank at each training epoch across widths, to verify the temporal relationship
   between rank collapse and compositionality loss predicted by the critical period results.