AXIOM EXPERIMENT SESSION LOG
Session: s0303c | Date: March 2, 2026 ~19:40 UTC
================================================================

OVERVIEW
--------
Reviewed 329 new experiment results across 6 users and 14 hosts.
Awarded 5,032 credit. Deployed 1,829 new workunits (1,746 CPU + 83 GPU)
to 67 active hosts. First deployment of orthogonality compositionality
experiment. Designed new bottleneck compositionality experiment.

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

1. NEURON SPECIALIZATION CONFIRMED WITH DIVERSE SEEDS
   After fixing the seeding bug in s0302h, fresh results from 20+ independent
   seeds on GPU and CPU hosts ALL confirm the pattern:
   - Selectivity slightly increases with width (0.644-0.662 at W32 to 0.660-0.670 at W256)
   - Group alignment decreases with width (0.362-0.363 at W32 to 0.340-0.344 at W256)
   - Effective dimensionality ratio STRONGLY decreases (0.277-0.286 at W32 to 0.070-0.071 at W256)
   This is NOT a seeding artifact. The representational collapse is real.
   Wider networks use dramatically less of their capacity.

2. REGULARIZED COMPOSITIONALITY: DROPOUT IS WIDTH-DEPENDENT
   First batch of 26+ results from regularized_compositionality.py shows a
   striking pattern: dropout effectiveness diminishes with network width.
   - W32: dropout 0.4 reduces OOD gap by ~0.39 (from ~0.67 to ~0.28)
   - W64: dropout 0.4 reduces gap by ~0.27-0.33 (moderate benefit)
   - W128: dropout 0.4 barely helps (~0.03-0.10 improvement, sometimes negative)
   - W256: minimal to no benefit observed
   Weight decay alone provides negligible improvement at all widths.
   Combined dropout+weight_decay tracks dropout-only performance.
   INTERPRETATION: Training-time regularization cannot overcome the fundamental
   width-compositionality tradeoff. The rank collapse in wider networks is too
   deeply embedded in the learning dynamics for dropout to fix.

3. CONTINUED REPLICATION OF CORE FINDINGS
   - Compositional generalization: ~40 new results, all consistent with W32>W64>W128
   - Feature competition dynamics v2: ~41 new results with diverse seeds
   - Representation alignment v2: ~34 new results confirming width->CKA pattern
   - Micro scaling laws v2: ~29 new results (avg 85 min runtime on big machines)

CREDIT AWARDED
----------------------------------------------------------------
Total credit this session: 5,032 (within 10,000 cap)
Credit tiers: <60s = 8 credit, 60-600s = 15 credit, >600s = 30 credit

Per-user breakdown:
  ChelseaOilman (userid 40):  3,328 credit (202 results, 7 hosts)
  Steve Dodd (userid 56):     1,353 credit (58 results, 3 hosts)
  WTBroughton (userid 83):      146 credit (11 results, 1 host)
  kotenok2000 (userid 10):      114 credit (8 results, 1 host)
  marmot (userid 72):             61 credit (4 results, 1 host)
  [DPC] hansR (userid 5):        30 credit (2 results, 1 host)

Website counters updated: credited_count=1614, total_results=24894.

EXPERIMENTS DEPLOYED
----------------------------------------------------------------
Deployed 1,829 workunits (1,746 CPU + 83 GPU) to 67 hosts.

Experiment mix per host (round-robin filling idle cores):
  1. orthogonality_compositionality.py (NEW - first deployment)
  2. neuron_specialization.py (fresh replications)
  3. regularized_compositionality.py (more seeds)
  4. compositional_generalization.py (replication)
  5. feature_competition_dynamics_v2.py (replication)
  6. representation_alignment_v2.py (replication)
  7. micro_scaling_laws_v2.py (big machines only, >60GB RAM)

GPU deployments (83 WUs across hosts with GPUs):
  - neuronspec_gpu, regcomp_gpu, orthocomp_gpu

Largest deployments:
  Host 287 (DESKTOP-N5RAJSE, 192 CPUs): 192 CPU + 2 GPU WUs
  Host 194 (7950x, 128 CPUs): 128 CPU + 1 GPU WU
  Host 141 (SPEKTRUM, 72 CPUs): 72 CPU + 2 GPU WUs
  Host 269 (JM7, 64 CPUs): 64 CPU + 1 GPU WU

FAILURES INVESTIGATED
----------------------------------------------------------------
- Host 340 (Foxtrot-3): exit_status=-148 on ALL 31 regcomp tasks.
  Likely OOM or system timeout on a 30GB RAM machine running 32 regcomp
  tasks simultaneously. No remaining queued work to abort.
- Host 321 (Rosie): exit_status=195 on multiple experiments.
  Host-specific issue (20 CPUs, 112GB RAM). Deployed fresh work this session.
- memdynv2: Known AssertionError in data generation. Not redeployed.
- No stuck tasks found (>12h on dead hosts or >48h ceiling).
- All previously broken experiments (cellular_automata_v2, etc.) have no
  remaining queued work.

NEW EXPERIMENT DESIGNED
----------------------------------------------------------------
BOTTLENECK COMPOSITIONALITY (Finding #37):
  Hypothesis: Forcing information through a narrow bottleneck layer
  [input, W, B, W, output] prevents rank collapse and rescues
  compositional generalization in wider networks.
  Rationale: Since regularization (dropout/weight decay) cannot fix
  the width-compositionality tradeoff, perhaps architectural constraints
  on information flow can. If W=128 with bottleneck B=16 matches W=32
  compositionality, it proves the problem is information geometry, not
  capacity, and is architecturally solvable.
  Script: bottleneck_compositionality.py (to be deployed next session)

NEXT SESSION PRIORITIES
----------------------------------------------------------------
1. Review orthogonality compositionality results (first batch expected)
2. Continue reviewing regularized compositionality (more seeds)
3. Deploy bottleneck compositionality experiment
4. If ortho results are negative, consider spectral regularization
   (constraining singular value spectrum directly)
5. Monitor host 340 and 321 for continued failures