AXIOM BOINC EXPERIMENT REVIEW — Session s0302e
Date: March 2, 2026 ~19:30 UTC
================================================================

SUMMARY
-------
Credited 323 completed results (1,594.7 total credit).
Deployed ~1,866 new workunits (1,777 CPU + 89 GPU) to ~70 hosts.
No stuck or broken tasks requiring cleanup.
Fleet: ~85 active hosts, now fully loaded.

CREDIT AWARDED (1,594.7 total, 323 results)
--------------------------------------------
Per user:
  Steve Dodd: 801.7 credit (Dads-PC + DadOld-PC, heavy microscalev2 runs)
  ChelseaOilman: 620.2 credit (Hotel-3, Dell-9520, Charlie/Delta/Echo/Foxtrot fleet)
  Anandbhat: 83.5 credit (DESKTOP-EMAFVVL)
  WTBroughton: 35.2 credit (achernar — legacy grokking + lottery results)
  marmot: 32.2 credit (XYLENA)
  kotenok2000: 15.7 credit (DESKTOP-P57624Q)
  dthonon: 3.6 credit (thonon-meylan)
  [DPC] hansR: 1.6 credit (dbgrensenh27)
  Vato: 1.0 credit (iand-r7-5800h)

Per experiment type:
  compgen: 44, bottleneck: 41, regcomp: 36, curriculum: 34
  repalignv2: 34, featcompv2: 34, memdynv2: 31, microscalev2: 25
  orthocomp: 24, neuronspec: 18, grokking: 1, lottery: 1

DEPLOYMENT (1,866 workunits: 1,777 CPU + 89 GPU)
-------------------------------------------------
Experiment priority (weighted):
  bottleneck_compositionality.py — weight 3 (strongest finding, needs 50+ seeds)
  orthogonality_compositionality.py — weight 2 (growing confirmation)
  regularized_compositionality.py — weight 1
  compositional_generalization.py — weight 1
  feature_competition_dynamics_v2.py — weight 1
  representation_alignment_v2.py — weight 1
  neuron_specialization.py — weight 1
  micro_scaling_laws_v2.py — weight 1 (big hosts only, >=32 cores & >=60GB RAM)

GPU workunits: bottleneck_gpu, orthocomp_gpu, neuronspec_gpu, regcomp_gpu
  1 GPU WU per GPU on each GPU-capable host.

Major hosts filled:
  DESKTOP-N5RAJSE (192 cores, 2 GPU) — 192 CPU + 2 GPU WUs
  7950x (128 cores, 1 GPU) — 128 CPU + 1 GPU WUs
  SPEKTRUM (72 cores, 2 GPU) — 72 CPU + 2 GPU WUs
  JM7 (64 cores, 1 GPU) — 64 CPU + 1 GPU WUs
  DadOld-PC (80 cores, 50 idle) — 50 CPU + 2 GPU WUs
  Dads-PC (80 cores, 38 idle) — 38 CPU + 2 GPU WUs
  Plus ~65 additional hosts ranging from 4-32 cores each

Skipped hosts: Latitude (4GB RAM), Athlon-x2-250 (3GB RAM), alix (SSL),
  archlinux (SSL), MSI-B550-A-Pro (exit_status=203), Rosie (exit_status=195)

SYSTEM STATUS
-------------
No stuck tasks (>12h on dead hosts or >48h on live hosts).
No new crash patterns in active experiments.
Over-queued hosts still draining from previous sessions: 113, 137, 159, 219, 222, 319.
Website counters updated: credited=2901, total_results=26175.

KEY SCIENTIFIC FINDINGS
=======================
1. BOTTLENECK COMPOSITIONALITY continues strong positive signal. This session adds ~41 bottleneck results to the growing pool. The bottleneck layer [Wide→Bottleneck→Wide] remains the STRONGEST intervention discovered for rescuing compositional generalization in wide networks. Previous finding: gap reduction of 4-8% absolute (vs 1-2% for orthogonality, 0% for pruning). Best configs: w128_b16, w256_b32. Hierarchy confirmed: bottleneck >> ortho >> dropout >> pruning.

2. ORTHOGONALITY COMPOSITIONALITY accumulates ~24 more seeds this session. Soft orthogonality penalty gives modest but real rescue of compositionality. ortho_rescues_compositionality=true consistent; ortho_helps_wide_networks remains MIXED.

3. REGULARIZED COMPOSITIONALITY adds ~36 regcomp results. Dropout is width-dependent: helps narrow networks more than wide ones. Growing confirmation.

4. MICRO SCALING LAWS continue to show mixed signal on data scaling at micro scale. Parameter scaling consistently does NOT hold. ~25 more results this session from heavy compute hosts.

5. Cross-validation status update: Findings 27 (feature competition ~160+), 28 (rep alignment ~150+), and 31 (compositional generalization ~160+) continue to mature with additional seeds from this session.

NEXT SESSION PRIORITIES
-----------------------
- Continue bottleneck compositionality replication toward 100+ seed target
- Monitor orthocomp for convergence on the "ortho_helps_wide_networks" mixed signal
- Consider designing new experiment: "random projection bottleneck" to test whether bottleneck benefit comes from dimensionality reduction per se or from learned compression. If a frozen random projection [W→Random_Fixed→W] also helps compositionality, the mechanism is purely architectural constraint. If only trained bottleneck helps, it's about learning the right compression.
- Retired experiments remain retired. No new experiments deployed this session.