AXIOM BOINC SESSION RESULTS
Timestamp: 2026-03-03 (America/Denver)
Session sources: validate_2026-03-03_1909.txt + run_2026-03-03_1352.log

PART 1 VALIDATION AND CREDIT SUMMARY
- Reviewed and credited completed exp_* results in ID span 1573025 through 1587373.
- Total reviewed+credited this pass: 8439.
- Outcome mix: success=1605, failed/error=6834.
- Upload payload quality in this credited batch: missing=8439, experiment_result=0, error=0.
- Credit awarded: 10000 total (session cap respected), across 46 users.
- Top session additions: ChelseaOilman +2020, Steve Dodd +1508, Amapola +639, mmonnin +593, Landjunge +586.
- Website counters updated: credited_count 64711 -> 73150; total_results_count remained 54848 (non-decreasing rule preserved).

CLEANUP AND STUCK/BROKEN TASK ACTIONS
- Validation cleanup: dead-host stuck-task aborts (>12h run and >6h no contact): 0.
- Validation cleanup: hard-ceiling aborts (>48h): 0.
- Broken-prefix blanket aborts during validation: none applied (no active unsent/in-progress exp_* at review time).
- Part 2 retirement pass was rerun from live state: retirement candidates found but unsent backlog for listed candidates was 0, so ABORT_TOTAL=0.

PART 2 DEPLOYMENT SUMMARY
CPU deployment
- Deployment scripts: wd_batchnoise_interaction.py, wd_labelsmooth_interaction.py.
- Host-targeted fill policy used in run: queue to 3x CPU per active host, skip hosts with <6 GB RAM.
- CPU hosts considered: 81.
- Hosts skipped for RAM floor: 2.
- CPU workunits created: 2937.

GPU deployment checkpoint
- Intended GPU scripts in run: wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py.
- The logged GPU deployment command was interrupted before completion in run_2026-03-03_1352.log.
- Post-run checkpoint query for this session window (since 2026-03-03 13:52) found: GPU hosts assigned=0, GPU workunits created=0.

EXPERIMENT DESIGN AND NOVELTY CHECK NOTES
- New experiment script created in this session: wd_batchnoise_interaction.py.
- Novelty/related-work checks were documented in the run log via targeted searches, including:
  - 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
- Design intent of wd_batchnoise_interaction.py: test whether late weight decay gains are stronger under small-batch gradient-noise conditions than large-batch conditions.

KEY SCIENTIFIC FINDINGS
1. No new reversal signal was observed against the established inverse critical-period (ICP) weight-decay conclusions in the validated batch.
2. The credited validation batch was dominated by missing upload payloads (8439/8439), so this pass primarily advanced fairness/credit maintenance rather than new payload-level mechanistic interpretation.
3. Mechanism-focused experiment design progressed via wd_batchnoise_interaction.py to isolate batch-noise x late-weight-decay interaction effects, with novelty checks documented before deployment actions.
4. Operationally, CPU queue fill succeeded at scale (2937 WUs), while GPU deployment for the same run did not complete and produced zero new GPU WUs in the tracked session window.

KEY OPEN QUESTIONS
- Root cause of repeated missing upload payloads for completed rows (retention timing, path/layout mismatch, or ingestion pipeline behavior).
- Whether GPU deployment interruption was transient execution/tooling interruption versus scheduler-state saturation at run time.