AXIOM BOINC SESSION LOG - Part 3 Save & Upload
Session timestamp: 2026-03-04 13:55 -07:00
Sources used:
- Validation log: validate_2026-03-04_1352.txt
- Auto-review/deploy log: run_2026-03-03_1352.log
- Findings memory: findings_summary.txt

PART 1 SUMMARY - VALIDATION, CREDIT, CLEANUP
- Results reviewed/credited this session: 24 completed success results.
- Total credit awarded: 472 (under 10,000 cap).
- Per-user credit: ChelseaOilman 240; Armin Gips 80; Steve Dodd 54; WTBroughton 34; _Scandinavian_ 24; zombie67 [MM] 20; vanos0512 20.
- Payload integrity: 24/24 credited outputs resolved via xml_doc_out upload filenames and parsed as valid experiment_result JSON.
- Counter updates: credited_count.txt=1867; total_results_count.txt=1867.

STUCK/BROKEN TASK CLEANUP
- Dead-host >12h cleanup aborts: 0
- Hard >48h cleanup aborts: 0
- Broken experiment broad aborts: none this session
- Uncredited completed failures with elapsed_time > 0 at cutoff: 0

PART 2 SUMMARY - DEPLOYMENT / RESEARCH
CPU deployment (appid=1)
- Deployment mode: host-targeted queue fill to ~3x CPU cores per active host.
- Hosts evaluated for CPU fill: 81
- Hosts skipped for low RAM (<6 GB): 2
- CPU workunits created: 2937
- CPU scripts deployed: wd_batchnoise_interaction.py, wd_labelsmooth_interaction.py
- Targeting: --target_host assignment was used across active CPU hosts with queue deficits.

GPU deployment (appid=2)
- Planned GPU scripts: wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py
- GPU deployment command sequence started, but run log ended with interrupt (^C) before completion counters were emitted.
- GPU checkpoint status from this run log: partial/incomplete capture; final GPU host count and final GPU workunit count were not recorded in the captured output.

Retirement/queue hygiene during Part 2
- Retirement candidate scan executed.
- ABORT_TOTAL=0 (no additional unsent abort actions taken in this pass).

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK DOCUMENTATION
- New script added and syntax-checked: wd_batchnoise_interaction.py (py_compile OK on server).
- Hypothesis: late weight decay benefit is stronger under small-batch gradient-noise conditions (interaction effect test).
- Novelty search queries logged in run output:
  - site:arxiv.org weight decay label smoothing interaction
  - site:arxiv.org adaptive weight decay deep neural networks
  - site:arxiv.org batch size weight decay generalization
  - "weight decay" "batch size" "schedule" neural networks
  - arxiv 1711.05101 decoupled weight decay regularization

KEY SCIENTIFIC FINDINGS
1. Validation pass recovered all credited outputs via xml_doc_out filename resolution (24/24 valid experiment_result payloads), confirming the upload-name parsing fix is functioning.
2. GPU WD-family outputs in the credited batch (wd_noise_trigger_gpu, wd_timing_scale_gpu, wd_anisotropy_trigger_gpu) produced science-bearing JSON, extending GPU reliability evidence.
3. Delay-driven ecology/control families remained high-throughput with runtimes concentrated near ~844 seconds median in this credited batch.
4. At session cutoff, nonzero-elapsed uncredited failures remained zero, indicating no pending compute-time compensation for runtime crashes.

NOTES
- This session log intentionally summarizes by experiment family/type and does not include cumulative credited ID lists.
- GPU deployment totals are reported as partial because the recorded run log terminated before final GPU output lines.