AXIOM BOINC SESSION LOG (PART 3 SAVE)
Session timestamp: 2026-03-04 03:36:11 -07:00
Sources: validate_2026-03-04_0327.txt, run_2026-03-03_1352.log, findings_summary.txt

SCOPE
- Consolidated Part 1 (validation/credit/cleanup) and Part 2 (deployment/research/design) outcomes into one publishable session report.

PART 1: VALIDATION, CREDIT, AND CLEANUP SUMMARY
- Results reviewed/credited (summarized by type):
  - Core pass reviewed targeted completed experiment rows across WD-family and follow-up lines, with conservative time-tiered crediting.
  - Backlog pass credited 2,823 completed-uncredited rows (predominantly outcome=6 timeout/abort completions).
  - Final catch-up pass credited 190 additional completed-uncredited rows.
- Total rows credited this session: 3,069
- Total credit awarded this session: 1,277.30 (within 10,000 cap)
- Top credited users this session-id range: Amapola +909.70, Orange Kid +260.00, makracz +40.50, ChelseaOilman +33.50.
- Stuck/broken cleanup:
  - Dead-host stuck-task aborts (>12h running, >6h silent): 0
  - Hard-ceiling aborts (>48h running): 0
  - Broken-prefix fail-rate auto-abort: none triggered at check time
- Website counters updated:
  - credited_count.txt = 3247
  - total_results_count.txt = 254
- Cutoff state after Part 1: uncredited completed exp rows = 0; uncredited completed success rows = 0.

PART 2: DEPLOYMENT AND RESEARCH SUMMARY
- Retirement guard pass executed against over-seeded lines; no unsent tasks found for abort (ABORT_TOTAL=0).

CPU DEPLOYMENT
- CPU deployment run completed with host-targeted queue fill:
  - CPU_HOSTS_SEEN = 81
  - CPU_SKIPPED_LOW_RAM (<6GB) = 2
  - CPU_WU_CREATED = 2937
- CPU scripts deployed:
  - wd_batchnoise_interaction.py (newly designed this session)
  - wd_labelsmooth_interaction.py (active interaction line)
- Targeting policy used: fill toward ~3x CPU queued tasks per eligible active host with duplicate checks and per-host balancing.

GPU DEPLOYMENT CHECKPOINT
- GPU deployment command in run log was interrupted before final printed summary lines.
- Independent DB checkpoint by workunit naming on server indicates active GPU deployment for:
  - wd_curvature_trigger_gpu.py: 737 workunits, 85 hosts
  - wd_timing_scale_gpu.py: 731 workunits, 84 hosts
- GPU checkpoint totals from this snapshot:
  - GPU hosts represented: 85
  - GPU workunits represented: 1,468
  - GPU scripts: wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py

NEW EXPERIMENT DESIGN + NOVELTY CHECK DOCUMENTATION
- Newly added experiment script:
  - wd_batchnoise_interaction.py
  - Hypothesis: late weight decay improves generalization more under high-noise (small-batch) than low-noise (large-batch) training.
  - Implementation notes: numpy-only, seed derivation from experiment_name, run_duration-driven repeated-trial loop, outputs aggregated interaction metrics.
- Novelty/literature search log captured in Part 2 run log:
  - "weight decay batch size interaction neural networks"
  - "arxiv weight decay batch size interaction deep learning"
  - "site:arxiv.org weight decay label smoothing interaction"
  - "site:arxiv.org batch size weight decay generalization"
  - "arxiv 1711.05101 decoupled weight decay regularization"
- Novel angle retained for deployment:
  - Explicit interaction-mechanism test (late WD timing x gradient-noise regime) rather than single-factor WD schedule comparison.

KEY SCIENTIFIC FINDINGS
1. No new evidence in this session overturns the prior inverse critical-period / late-WD timing signal.
2. The dominant operational signal remains reliability pressure (large outcome=6 completion waves), not a contradictory scientific trend.
3. Missing upload JSON artifacts persist even among successful completions, limiting payload-level interpretation and emphasizing need for ingestion-path debugging.
4. The science queue was pivoted toward mechanism/interaction tests (batch-noise interaction, label-smoothing interaction, curvature-triggered GPU timing) rather than further over-seeding of settled WD families.

NOTES
- This report intentionally omits cumulative credited result-ID lists; database state is the source of truth.