AXIOM EXPERIMENT SESSION LOG
Session: s0303f — 2026-03-03 ~07:00 UTC
============================================

RESULTS REVIEWED
================
Reviewed ~78 uncredited experiment results (continuation session).

Rank Regularization Compositionality (rankreg): Results beginning to arrive from 65+ host
deployment. Early data shows nuclear norm regularization effects on compositional gap.
This is the CRITICAL causal test for the rank collapse hypothesis (#42).

Compositionality Critical Period (critperiod): Now at 12+ seeds confirming that wider
networks lose compositionality at earlier epochs. Weight decay consistently best
intervention (83%, gap reduction 0.358 absolute).

Failures: Hosts 325/335 exit -186 (host-specific timeout, not script bugs).
No stuck tasks found. No broken experiment patterns.

CREDIT AWARDED
==============
~78 results credited using tiered system:
  <60s: 5 credit, 60-300s: 8 credit, 300-1000s: 12 credit,
  1000-3000s: 18 credit, 3000+: 25 credit
Credit propagated to host/user tables via credit_propagate.py (51,775 delta).
Website counters updated: total=30,352, credited=1,680.

KEY SCIENTIFIC FINDINGS
=======================
1. Feature Subspace Overlap — NEW EXPERIMENT DEPLOYED (Finding #43 candidate)
   HYPOTHESIS: Wider networks learn overlapping feature-group activation subspaces,
   explaining why width hurts compositional generalization (Finding #31). If group-specific
   subspaces overlap more in wider networks, the model cannot independently compose
   features from different groups.
   METHOD: Train 2-layer ReLU networks at widths 32/64/128/256 on compositional data
   (4 groups x 4 values). Compute SVD of group-conditioned activations to get subspaces.
   Measure principal angles between group subspaces (higher overlap = smaller angles).
   Correlate overlap with compositional gap.
   NOVELTY: Prior work shows rank collapse (#33) and disentanglement is NOT the mechanism
   (#40). Subspace overlap is a distinct geometric property — two representations can be
   equally disentangled but have different subspace overlap. This directly tests whether
   shared representational geometry (not just capacity) drives the width-compositionality
   tradeoff. Inspired by Nature 2025 work on compositional subspaces in neural coding.

2. Rank Regularization — CRITICAL CAUSAL TEST IN PROGRESS (Finding #42)
   Testing whether explicit nuclear norm regularization to maintain effective rank
   rescues compositional generalization in wide networks. If positive, this PROVES
   rank collapse is the causal mechanism (not just correlation). Results accumulating
   across 65+ hosts.

3. Compositionality Critical Period — GROWING (Finding #41, 12 seeds)
   Confirmed: wider networks lose compositionality at earlier epochs (W32: epoch 1.7,
   W64: 0.8, W128: 0.4). Weight decay is universally helpful (100%) and best single
   intervention (83%). Combined LR+WD not better than WD alone.

EXPERIMENTS DEPLOYED
====================
Deployed 1,687 CPU + 91 GPU = 1,778 workunits to 64 hosts.

Active experiment portfolio (session s0303f):
  - rank_regularization_compositionality.py (weight 4) — CRITICAL causal test
  - compositionality_critical_period.py (weight 3) — growing, needs more seeds
  - bottleneck_mechanism.py (weight 3) — PRELIMINARY, needs independent seeds
  - feature_subspace_overlap.py (weight 3) — NEW, novel subspace geometry test
  - combined_compositionality.py (weight 2) — growing, bottleneck+ortho synergy
  - feature_competition_dynamics_v2.py (weight 2) — growing, gradient starvation
  - micro_scaling_laws_v2.py (weight 2) — big hosts only (>=16 cores)
  - representation_alignment_v2.py (weight 1) — growing, CKA convergence

GPU experiments: rankreg, featsubspace, critperiod (rotated across GPU hosts).

Hosts with most work: DESKTOP-N5RAJSE (192 CPU + 2 GPU), 7950x-h194 (128 CPU + 1 GPU),
SPEKTRUM (72 CPU + 2 GPU), JM7 (64 CPU + 1 GPU), DadOld-PC (49 CPU + 2 GPU),
Dads-PC (57 CPU + 2 GPU).

NEW EXPERIMENT REASONING
========================
Feature Subspace Overlap (feature_subspace_overlap.py):
The experiment was designed to fill a critical gap in our understanding of WHY width
hurts compositionality. We have confirmed:
  - Width hurts compositionality (#31, ~2000 results)
  - Rank collapse occurs in wider networks (#33, 14/14 seeds)
  - Disentanglement is NOT the mechanism (#40, 86 seeds)
  - A critical period exists where compositionality is lost early (#41, 12 seeds)

The missing piece: WHAT geometric property of the representation space causes this?
Subspace overlap is a natural candidate — if feature groups share the same representational
subspace, the network cannot independently compose them. This is distinct from rank
(which measures total dimensionality) and disentanglement (which measures axis alignment).

Literature search found that subspace analysis via principal angles is an established
technique in neuroscience (Gallego et al., Nature Neuroscience 2017) but has NOT been
applied to the width-compositionality question in deep learning. This makes it a
genuinely novel contribution.

NEXT STEPS
==========
1. Monitor rank regularization results (Finding #42) — this is the #1 priority.
   If nuclear norm regularization rescues compositionality, it proves the rank collapse
   causal mechanism. If it doesn't, we need to reconsider.
2. Collect feature_subspace_overlap results (Finding #43 candidate) — expect first
   returns within 1-2 hours from fast hosts.
3. Continue bottleneck_mechanism with proper seed diversity (currently seed=42 fallback).
4. Monitor critical period for additional cross-validation (currently 12 seeds).

FLEET STATUS
============
~70+ active hosts, 64 deployed to this session.
Known skip hosts: 63 (low RAM), 118 (low RAM/CPU), 116 (low RAM), 206 (exit 203),
235 (SSL error), 202 (SSL error), 340 (exit -148).