AXIOM BOINC SESSION LOG — 2026-03-02 ~23:30 UTC (Session s0302m)
================================================================

OVERVIEW
--------
Routine credit-and-deploy cycle. Reviewed 324 new results, awarded 3,355 credit
to 10 users across 18 hosts. Deployed 1,872 new workunits (1,781 CPU + 91 GPU)
to 71 hosts. Key science update: first results from combined compositionality
experiment (Finding #38) show approximately additive effect — no synergy.

RESULTS REVIEWED THIS SESSION
------------------------------
324 uncredited results reviewed and credited. Breakdown by experiment type:
- Bottleneck compositionality: ~60 results (hosts 325, 332, 327, 339, 340)
  → Continues to confirm: 171/183 = 93.4% positive (bottleneck_helps=true)
- Combined compositionality: ~13 results (hosts 332, 327) — FIRST RESULTS
  → Finding: approximately additive, NOT synergistic (see below)
- Orthocomp: ~15 results (hosts 332, 339, 324, 327)
  → 175/175 rescues compositionality (100%), 132/175 helps wide (75%)
- Regcomp: ~8 results (host 345)
- Compgen: ~30 results (hosts 335, 320, 325, 219)
- FeatCompV2: ~35 results (hosts 219, 335, 320, 325)
- RepAlignV2: ~40 results (hosts 219, 335, 320, 325, 332, 327, 340)
- MemDynV2: ~35 results (hosts 219, 335, 320, 325, 332)
- NeuronSpec: ~15 results (hosts 325, 340, 332, 327)
- MicroScaleV2: ~10 results (hosts 335, 320, 332, 113, 327)
- Curriculum: ~20 results (hosts 219, 335, 320, 325)
- GPU experiments: ~8 results (hosts 9, 324, 319, 339, 327)
- Legacy experiments from Dad-Workstation/achernar: ~30 results (old experiment types)

CREDIT AWARDED
--------------
Total: 3,355 credit across 324 results

Per-user totals:
  ChelseaOilman: 2,295 credit (hosts: Charlie-2, Delta-2, Dell-9520, etc.)
  Anandbhat: 457 credit (host: DESKTOP-EMAFVVL)
  Steve Dodd: 305 credit (host: Dad-Workstation)
  Armin Gips: 108 credit (host: Andre-WEBK)
  WTBroughton: 81 credit (host: achernar)
  kotenok2000: 56 credit (host: DESKTOP-P57624Q)
  marmot: 25 credit (host: XYLENA)
  [DPC] hansR: 12 credit (host: dbgrensenh27)
  dthonon: 8 credit (host: thonon-meylan)
  Vato: 8 credit (host: iand-r7-5800h)

Credit tiers: 5cr (82 results, <30s), 8cr (86, 30-120s), 12cr (114, 120-600s),
              18cr (23, 600-2000s), 25cr (19, 2000s+)

Website counters updated: credited_count=3,555, total_results_count=26,814.

KEY SCIENTIFIC FINDINGS
-----------------------
1. COMBINED COMPOSITIONALITY (Finding #38) — FIRST DATA, APPROXIMATELY ADDITIVE
   13 seeds completed. The combination of bottleneck architecture + orthogonality
   regularization shows approximately additive effect, NOT synergistic.
   - Synergy detected: 10/13 True (but magnitude is near zero)
   - Mean gap synergy: +0.003 (essentially zero — neither helps nor hurts)
   - Mean OOD synergy: +0.002
   - Effect type breakdown: approximately_additive=7, superadditive=4, subadditive=2
   - Best intervention: "both"=5, "none"=4, "bottleneck_only"=3, "ortho_only"=1
   INTERPRETATION: Bottleneck and orthogonality address somewhat orthogonal aspects
   of the compositionality problem. Bottleneck constrains information flow
   structurally (hard constraint on dimensionality); orthogonality prevents
   eigenvalue collapse within each layer (soft regularization). They don't
   interfere, but they don't amplify each other either. The bottleneck remains
   the dominant intervention.

2. BOTTLENECK COMPOSITIONALITY (Finding #37) — DEFINITIVELY CONFIRMED
   Updated count: 171/183 positive results with key_results (93.4%).
   Now at ~230+ total result files. Promoted to definitively confirmed status.
   The bottleneck architecture remains the strongest known intervention for
   rescuing compositional generalization in wider networks.

3. ORTHOGONALITY COMPOSITIONALITY (Finding #36) — STRONGLY CONFIRMED
   Updated: 175/175 rescues compositionality (100%), 132/175 helps wide (75.4%).
   The effect is real and consistent but modest compared to bottleneck.
   The 75% rate on helps_wide indicates the effect is seed/config dependent
   for wider architectures, while narrow-net rescue is universal.

DEPLOYMENT
----------
1,872 new workunits deployed (1,781 CPU + 91 GPU) to 71 hosts.

Experiment weights: bottleneck(3), combined_compositionality(3),
  orthocomp(2), regcomp(1), compgen(1), featcompv2(1),
  repalignv2(1), neuronspec(1), microscalev2(1, big hosts only)

GPU experiments: bottleneck_gpu, orthocomp_gpu, neuronspec_gpu

Notable deployments:
- DESKTOP-N5RAJSE (h287, 192 CPUs): 192 CPU + 2 GPU WUs
- 7950x (h194, 128 CPUs): 128 CPU + 1 GPU WU
- SPEKTRUM (h141, 72 CPUs): 72 CPU + 2 GPU WUs
- JM7 (h269, 64 CPUs): 64 CPU + 1 GPU WU
- DadOld-PC (h85, 80 CPUs): 46 CPU + 2 GPU WUs (34 already running)
- Plus 66 other hosts fully loaded

No stuck tasks found (>12h dead host or >48h absolute). No new broken experiments.

NEXT SESSION PLAN
-----------------
1. Review combined compositionality with 50-70+ seeds (should be well-powered)
2. If combined comp stabilizes as additive, consider designing a new experiment:
   "Gradient Starvation → Compositionality" — test whether gradient starvation
   (Finding #27: scales with width) is the MECHANISM behind the width-compositionality
   tradeoff (Finding #31). This would connect two independent finding lines.
3. Continue monitoring bottleneck and orthocomp for stability
4. Consider promoting Finding #36 (orthocomp) to definitively confirmed if
   seed count reaches ~200+ with stable results