Axiom BOINC Session Log - Part 3 Save/Upload
Timestamp: 2026-03-04 10:09 (America/Denver)

SESSION SUMMARY
- Combined summary from Part 1 validation/credit pass and Part 2 research/deployment pass.
- Validation completed and credits awarded on live DB; research pass added/deployed WD-interaction experiments.

PART 1 - RESULTS REVIEWED AND CREDIT AWARDED
- Total credit awarded this session: 5880 (session cap 10000; compliant).
- Total completed rows credited: 460 (all outcome=1 in credited batches).
- Credited mix: CPU(appid=1)=441, GPU(appid=2)=19.
- Website counters after pass: credited_count.txt=1613, total_results_count.txt=1613.
- Per-user increments:
  - ChelseaOilman: +4563
  - Steve Dodd: +1164
  - _Scandinavian_: +38
  - Orange Kid: +27
  - makracz: +26
  - Vato: +25
  - marmot: +24
  - vanos0512: +13

STUCK/BROKEN TASK CLEANUP
- Dead-host cleanup (>12h running and >6h host silence): 0 aborted.
- Hard >48h running-task cleanup: 0 aborted.
- Broken-experiment broad abort: none applied (no qualifying active multi-host crash pattern in screening window).

PART 2 - EXPERIMENT DEPLOYMENT
CPU DEPLOYMENT
- CPU deployment pass completed.
- Host targeting policy: 3x CPU queue target per active host, skipped hosts with <6 GB RAM.
- CPU hosts seen: 81
- CPU hosts skipped for RAM: 2
- CPU workunits deployed: 2937
- CPU scripts used in deployment:
  - wd_batchnoise_interaction.py
  - wd_labelsmooth_interaction.py

GPU DEPLOYMENT CHECKPOINT
- GPU deployment launched with scripts:
  - wd_curvature_trigger_gpu.py
  - wd_timing_scale_gpu.py
- Run log was interrupted during GPU deployment, so final in-log completion counters were not printed.
- Live DB snapshot for deployed GPU prefixes:
  - GPU hosts observed (curvature script): 12 (host IDs: 1, 9, 29, 57, 126, 159, 249, 340, 341, 347, 353, 355)
  - GPU hosts observed (timing-scale script): 11 (host IDs: 1, 9, 57, 126, 159, 249, 299, 341, 347, 353, 355)
  - GPU workunits present (curvature script): 1576
  - GPU workunits present (timing-scale script): 1523
  - GPU total workunits across these two scripts: 3099

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK
- New CPU experiment added in Part 2: wd_batchnoise_interaction.py
  - Hypothesis: late weight decay gains should be stronger under high-noise (small-batch) training than low-noise (large-batch) training.
  - Includes explicit run-duration budget loop and host-derived seeding.
- Existing/newly active mechanism lines reviewed/deployed:
  - wd_labelsmooth_interaction.py
  - wd_curvature_trigger_gpu.py
  - wd_timing_scale_gpu.py
- Novelty check documentation captured in run log:
  - Web/arXiv search queries executed (examples):
    1. weight decay batch size interaction neural networks
    2. arxiv weight decay batch size interaction deep learning
    3. site:arxiv.org weight decay label smoothing interaction
    4. Scheduled Weight Decay paper arxiv 2021
  - Decision: proceed with interaction/mechanism-focused variants rather than repeating settled baseline WD lines.

KEY SCIENTIFIC FINDINGS
1. Validated completion flow remains dominated by delay/ecology/control families with healthy payload integrity in sampled rows, indicating stable throughput in the active science portfolio.
2. No experiment prefix met emergency broad-abort criteria (high fail-rate plus multi-host active crash evidence), so compute was preserved for healthy queues.
3. The current research direction is now explicitly mechanism-focused: label-smoothing interaction, curvature-triggered timing, and batch-noise interaction were prioritized as the next discriminative tests of WD timing behavior.
4. GPU WD mechanism families are actively deployed across multiple GPU hosts, enabling near-term cross-host comparison of curvature-triggered vs timing-scale effects.

END-OF-SESSION NOTES
- Uncredited completed success rows at validation cutoff: 1 (live queue race after final credit pass).
- No cumulative credited ID list retained here; DB remains source of truth for credited results.