AXIOM BOINC SESSION RESULTS - PART 3 COMBINED LOG
Session timestamp: 2026-03-03 23:55 (America/Denver)

PART 1 - VALIDATION, CREDIT, AND CLEANUP SUMMARY
- Completed+credited this session: 83 results across two credit passes, 371 total credit (under 10,000 cap).
- Session credit recipients: Steve Dodd +152, ChelseaOilman +113, kotenok2000 +45, Anandbhat +40, PyHelix +17, [DPC] hansR +4.
- Payload QC: uncredited-success sample 49/49 valid `experiment_result`; failed-prefix spot-check 27/30 with `experiment_result`, 0 parse errors, 3 missing files.
- Stuck task cleanup: 0 aborted by dead-host >12h rule; 0 aborted by >48h hard ceiling.
- Broken-prefix active aborts applied (server_state IN (2,4) -> server_state=5,outcome=5):
  - metapop_corridor_delay_forecast: 170
  - rps_delay_jitter_adaptive_mobility: 110
  - wd_mixup_interaction: 31
  - gpu_matmul_stress_gpu: 27
  - Total broken-prefix aborts: 338
- Website counters after Part 1 refresh: credited_count=54,562; total_results_count=55,181.

PART 2 - DEPLOYMENT AND EXPERIMENT DESIGN SUMMARY
CPU deployment
- CPU queue fill pass completed.
- Hosts processed for CPU fill: 81
- Hosts skipped for low RAM (<6 GB): 2
- CPU workunits created: 2,937
- CPU scripts used for deployment:
  - wd_batchnoise_interaction.py (new this session)
  - wd_labelsmooth_interaction.py

GPU deployment checkpoint
- GPU deployment pass was started with GPU-aware scripts and 3x GPU queue targeting.
- GPU scripts configured in deployment pass:
  - wd_curvature_trigger_gpu.py
  - wd_timing_scale_gpu.py
- Run log ended before GPU summary counters printed, so final GPU host/workunit totals were not emitted in this run log.

New experiments designed and novelty check documentation
- New CPU experiment script added: wd_batchnoise_interaction.py
- Research novelty search terms logged during Part 2:
  - 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
- Novel angle tested: whether late weight decay gains are stronger under small-batch gradient noise versus large-batch regimes (interaction-focused mechanism test, not a plain WD sweep).
- Deployment policy actions logged: retirement check run, no additional unsent aborts triggered in that pass (ABORT_TOTAL=0).

KEY SCIENTIFIC FINDINGS
1. Fresh completed ecological-delay and WD-interaction uploads remained science-bearing in this session, with payload quality checks showing valid `experiment_result` structure across sampled rows.
2. High-failure prefixes still frequently show outcome-level failures with empty stderr and often-valid payloads, supporting continued suspicion of validator/outcome mapping mismatch over pure runtime crash behavior.
3. Targeted broken-prefix aborts reduced further wasted volunteer compute while preserving produced files for diagnosis, and new interaction-focused WD experiment design (`wd_batchnoise_interaction`) extends the mechanism-testing line beyond prior timing-only probes.