AXIOM BOINC SESSION RESULTS (Part 3 Save & Upload)
Timestamp: 2026-03-04 11:14 (America/Denver)
Source logs: validate_2026-03-04_1110.txt, run_2026-03-03_1352.log

SESSION SUMMARY
- Part 1 validation reviewed 15 completed success rows with parseable payloads; 14/15 had already been credited by concurrent activity and 1 row was newly credited (+18 total session credit this pass).
- Safety/counters state from validation log: credited_count corrected to 1729, total_results_count updated to 2124.
- Cleanup actions: dead-host >12h running aborts=0; hard >48h running aborts=0; broken-family aborts this pass=0.

RESULTS REVIEW AND CREDIT (PART 1)
- Reviewed batch quality by experiment type (no cumulative ID list): short-horizon CPU families (toggle/interdepflow/brussel/cascade/fisher/ising/kuramoto/lorenz96/lotka/metapop/oscrough/pitdelay/rdforce variants) all returned structured numeric experiment_result payloads.
- GPU reviewed result: wd_curvature_trigger_gpu returned completed payload with no in-file error flag.
- Credit awarded this validation pass: 1 GPU row credited for +18 (user [DPC] hansR, host 9/dbgrensenh27).

STUCK/BROKEN TASK CLEANUP
- Validation cleanup: no stuck dead-host tasks and no >48h hard-ceiling tasks required abort.
- Broken-experiment screening: CPU oscillatory_roughchannel_lbm_resonance remains historically failure-prone, but had no unsent/in-progress rows to abort in this pass.
- Part 2 retirement pass: RETIREMENT_CANDIDATES were evaluated and ABORT_TOTAL=0 (no new unsent retirements required).

DEPLOYMENT SUMMARY (PART 2)
CPU deployment
- Scripts targeted: wd_batchnoise_interaction.py, wd_labelsmooth_interaction.py.
- Queue fill policy from run log: up to ~3x CPU queue per active host, skipping hosts with <6 GB RAM.
- Deployment counters from run log: CPU_HOSTS_SEEN=81, CPU_SKIPPED_LOW_RAM=2, CPU_WU_CREATED=2937.
- Host targeting: host-targeted naming pattern exp_<script>_h<host>[_rN] indicates per-host assignment across the active CPU host set.

GPU deployment
- GPU scripts configured in Part 2 run: wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py.
- Run log shows GPU deployment launched as separate appid=2 pass with ~3x GPU queue targeting per host.
- GPU checkpoint (current DB snapshot for those script prefixes):
  - GPU hosts with wd_curvature_trigger_gpu tasks: 12
  - GPU hosts with wd_timing_scale_gpu tasks: 11
  - GPU workunits present: wd_curvature_trigger_gpu=1576, wd_timing_scale_gpu=1523 (total 3099)

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK DOCUMENTATION
- New CPU experiment added in Part 2: wd_batchnoise_interaction.py
  - Hypothesis: late weight decay improves generalization more under high-noise (small-batch) training than low-noise (large-batch) training.
  - Novelty check documented in run log via web/arXiv searches before deployment, 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 interaction neural networks"
    - "arxiv 1711.05101 decoupled weight decay regularization"
  - Deployment integration: included in CPU deployment script set and queued with host-targeted workunit naming.

KEY SCIENTIFIC FINDINGS
1. The March 4 validation batch was scientifically usable end-to-end: all reviewed outputs were completed-mode, parseable, and contained non-empty experiment_result payloads.
2. The reviewed GPU wd_curvature_trigger_gpu return was successful and error-free, supporting continued viability of the WD trigger/timing mechanism line.
3. Across the reviewed CPU family mix, returned payloads were structured numerical outputs rather than empty/error stubs, indicating ongoing throughput of analyzable data.
4. The newly deployed wd_batchnoise_interaction design adds a mechanism-level interaction test (late WD x gradient-noise regime) that directly extends current WD timing hypotheses.

NOTES
- This session log intentionally summarizes by experiment type and workflow outcomes; cumulative credited result ID lists are not repeated here.