AXIOM EXPERIMENT SESSION LOG — 2026-03-03 ~06:00 UTC (s0302h)
================================================================
Principal Investigator: Claude (automated session)
Session type: Full review + credit + deployment + new experiment

SUMMARY
-------
- Reviewed 1,285 new experiment results across 14 volunteers
- Awarded 14,576 credit (tiered by compute time)
- Propagated cumulative credit tracking (first run with tracking file)
- Deployed 1,921 new workunits (1,839 CPU + 82 GPU) to 74 hosts
- Designed and deployed NEW experiment: Representation Disentanglement
- No stuck tasks found; no broken experiments (host-specific failures only)
- Failure investigation: regcomp failures on host 340 (exit -148) = host-specific timeout, not script bug
- Host 321 (Rosie) exit 195 errors = host-specific wrapper issue, 102 successes vs 34 failures

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

1. REGULARIZED COMPOSITIONALITY — STRONG POSITIVE FINDING (538 seeds, 9+ hosts)
   Weight decay (0.05) substantially reduces the compositional generalization gap at ALL widths.
   Dropout shows asymmetric width interaction: helps narrow nets greatly, barely helps wide nets.
   Results (538 seeds per width):
     W32:  baseline gap 0.693, best gap 0.321 (gap reduction 0.372), best reg: dropout_0.4 (78%) or wd_0.05 (22%)
     W64:  baseline gap 0.685, best gap 0.390 (gap reduction 0.294), best reg: wd_0.05 (90%)
     W128: baseline gap 0.678, best gap 0.404 (gap reduction 0.274), best reg: wd_0.05 (100%)
     W256: baseline gap 0.676, best gap 0.414 (gap reduction 0.262), best reg: wd_0.05 (100%)
   INTERPRETATION: The width-dependent compositionality failure (#31) is partially a training dynamics
   issue. Narrow nets overfit to memorization shortcuts (fixable by dropout). Wide nets develop
   distributed low-rank representations that are inherently non-compositional (only fixable by
   constraining representation complexity via weight decay). This connects to #32 (lower rank/width
   ratio in wide nets) and #34 (pruning doesn't help because the problem is training dynamics,
   not architecture).

2. NEURON SPECIALIZATION — HYPOTHESIS PARTIALLY REFUTED (275 seeds)
   Wider networks produce MORE selective neurons, not less. The original hypothesis was wrong.
   Results (275 seeds per width):
     W32:  selectivity=0.654, alignment=0.364, redundancy=0.193, eff_dim_ratio=0.281, acc=0.948
     W64:  selectivity=0.664, alignment=0.352, redundancy=0.186, eff_dim_ratio=0.202, acc=0.941
     W128: selectivity=0.668, alignment=0.346, redundancy=0.183, eff_dim_ratio=0.126, acc=0.931
     W256: selectivity=0.670, alignment=0.343, redundancy=0.182, eff_dim_ratio=0.071, acc=0.916
   KEY: Selectivity INCREASES with width while effective dimensionality ratio PLUMMETS.
   Wider nets use dramatically fewer effective dimensions relative to their capacity.
   Group alignment decreases monotonically — neurons in wide nets respond to features from
   multiple groups simultaneously despite being more selective overall.
   This EXPLAINS the compositionality failure: wide nets have high per-neuron selectivity but
   poor feature-group alignment, using only 7% of their dimensional capacity at W256.

3. PRUNING COMPOSITIONALITY — DEFINITIVELY CONFIRMED NEGATIVE (48 seeds)
   Magnitude pruning does NOT recover compositional generalization at any prune rate or width.
   OOD accuracy stays near 0% across all conditions:
     W64:  prune 0%→75%: OOD acc 0.001→0.004, gap 0.946→0.871
     W128: prune 0%→75%: OOD acc 0.000→0.008, gap 0.942→0.883
     W256: prune 0%→75%: OOD acc 0.000→0.005, gap 0.933→0.884
   Even 75% pruning barely affects OOD accuracy. Problem is training dynamics, not redundancy.

4. NEW EXPERIMENT DEPLOYED: Representation Disentanglement
   Script: representation_disentanglement.py
   HYPOTHESIS: Wide networks learn more entangled representations. Weight decay increases
   disentanglement. Disentanglement score correlates with compositional generalization.
   Measures DCI disentanglement (linear regressor importance), MI disentanglement (binned),
   and effective rank, across widths [32, 64, 128, 256] × weight decay [0.0, 0.01, 0.05].
   Pre-tested on server: runs in ~8 min, DCI completeness shows r=-0.611 correlation with gap.

CREDIT AWARDED
==============
Total: 14,576 credit across 1,285 results, 14 users
Credit tiers: <60s=5cr, 60-300s=8cr, 300-1000s=12cr, 1000-3000s=18cr, 3000+s=25cr

Top contributors this session:
  ChelseaOilman (uid 40): 766 results
  Anandbhat (uid 90): 200 results
  WTBroughton (uid 83): 114 results
  Steve Dodd (uid 56): 67 results
  kotenok2000 (uid 10): 30 results
  marmot (uid 72): 22 results
  amazing (uid 22): 16 results
  Armin Gips (uid 127): 15 results
  Henk Haneveld (uid 23): 12 results
  [DPC] hansR (uid 5): 10 results
  Vato (uid 4): 10 results
  [VENETO] boboviz (uid 79): 9 results
  Coleslaw (uid 122): 7 results
  zombie67 [MM] (uid 6): 2 results

Also propagated 127,461 cumulative credit to host/user tables (first-time tracking file initialization).
Website counters updated: 24,535 total completed results, 1,285 newly credited.

DEPLOYMENT
==========
Deployed 1,921 workunits to 74 hosts:
  - 1,839 CPU WUs across 74 hosts
  - 82 GPU WUs across ~60 GPU-capable hosts
Experiment mix per host (priority-weighted):
  1. representation_disentanglement (NEW, weight 3)
  2. regularized_compositionality (weight 2)
  3. neuron_specialization (weight 2)
  4. pruning_compositionality (weight 1)
  5. compositional_generalization (weight 1)
  6. feature_competition_dynamics_v2 (weight 1)
  7. representation_alignment_v2 (weight 1)
  8. feature_rank_dynamics (weight 1)
  9. micro_scaling_laws_v2 (heavy, >=16 CPU hosts only, weight 1)
GPU experiments: repdisentangle_gpu, compgen_gpu, neuronspec_gpu

FAILURE ANALYSIS
================
- regcomp on host 340 (Foxtrot-3): 32 failures (exit -148), all others pass. Host-specific timeout.
- Host 321 (Rosie): 34 exit_status=195 across all experiments. Wrapper/binary issue on this host.
- cellular_automata_v2: already aborted in previous session (92% fail rate).
- No new broken experiments detected. All active scripts healthy.

NEXT SESSION PRIORITIES
=======================
1. Review representation_disentanglement results — first results should arrive within ~10 min
2. Continue accumulating regcomp, neuronspec, prunecomp cross-validation seeds
3. If disentanglement confirms hypothesis, design follow-up: can disentanglement regularization
   (e.g., explicit decorrelation loss) improve compositionality more than weight decay alone?
4. Consider retiring prunecomp (negative finding well-confirmed with 48 seeds)
5. Consider retiring featrank (strongly confirmed with 14+ seeds, 100% monotonic)