AXIOM BOINC SESSION RESULTS (Part 3 Consolidation)
Timestamp: 2026-03-04 12:22
Sources: validate_2026-03-04_1218.txt, run_2026-03-03_1352.log, findings_summary.txt

SCOPE
- Consolidated summary of Part 1 (validation/credit/cleanup) and Part 2 (research/deployment).

PART 1: RESULTS REVIEWED AND CREDIT AWARDED
- Completed experiment results reviewed and credited: 479 (all outcome=1).
- Upload payload audit: 479/479 parseable JSON with experiment_result payloads.
- Credit awarded this session: +8,538 total (session cap 10,000 respected).
- Credited app mix: CPU (appid=1)=468, GPU (appid=2)=11.
- Incremental per-user credit additions:
  - ChelseaOilman: +6120
  - Steve Dodd: +1836
  - kotenok2000: +162
  - PyHelix: +126
  - Vato: +108
  - hwt: +56
  - marmot: +36
  - Orange Kid: +34
  - Henk Haneveld: +24
  - vanos0512: +18
  - [DPC] hansR: +18

CLEANUP / FAILURE MANAGEMENT
- Broken-experiment screening (CPU/GPU split): no high-confidence deterministic crash family with active unsent/in-progress queue; no broad abort applied.
- Stuck cleanup (>12h dead-host rule): 0 aborts.
- Hard ceiling cleanup (>48h runtime): 0 aborts.
- End-of-session uncredited completed: success=0, failure=0.

PART 2: DEPLOYMENT AND RESEARCH EXECUTION
- Retirement pass checked over-seeded families (svdrank, wdlr, wdwindow, wdoptim, percolation, repcrystal, etc.); unsent abort actions this pass: 0 (ABORT_TOTAL=0).

CPU deployment
- CPU host-targeted queue fill completed.
- CPU hosts seen by deployment logic: 81.
- Hosts skipped for low RAM (<6GB): 2.
- CPU workunits created: 2937.
- CPU scripts used:
  - wd_batchnoise_interaction.py (new in this session)
  - wd_labelsmooth_interaction.py
- Targeting policy used in run log: assign unassigned experiment types first, then backfill each host toward ~3x CPU queued tasks.

GPU deployment checkpoint
- GPU deployment command in run log was interrupted before final summary counters were printed.
- Live checkpoint after run:
  - Active GPU hosts with queued wd_curvature_trigger_gpu / wd_timing_scale_gpu work: 7
  - Queued GPU workunits (appid=2) for these scripts: 12 total
    - wd_curvature_trigger_gpu: 8
    - wd_timing_scale_gpu: 4
  - Observed queued-host mapping (checkpoint):
    - wd_curvature_trigger_gpu: host 353 (Thing0L_4000), host 1 (Pyhelix), host 9 (dbgrensenh27), host 29 (DESKTOP-P57624Q), host 159 (achernar)
    - wd_timing_scale_gpu: host 355 (DESKTOP-DUVULOS), host 29 (DESKTOP-P57624Q), host 299 (Raimund-PC)
- GPU scripts used in this cycle:
  - wd_curvature_trigger_gpu.py
  - wd_timing_scale_gpu.py

NEW EXPERIMENTS DESIGNED / NOVELTY CHECK DOCUMENTATION
- New script added and compiled:
  - wd_batchnoise_interaction.py
  - py_compile check passed on server.
- Mechanistic hypotheses emphasized in this cycle:
  - WD timing x label smoothing interaction.
  - Curvature-triggered WD vs fixed late-start WD.
  - WD timing x batch-noise interaction (small-batch vs large-batch gradient noise sensitivity).
- Novelty/prior-art search traces logged before deployment included:
  - 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
  - Scheduled Weight Decay paper arxiv 2021
- Session rationale: focus shifted from over-seeded main-effect WD lines toward interaction/mechanism tests expected to produce non-redundant findings.

KEY SCIENTIFIC FINDINGS
1. Validation quality was high in this session window: 479/479 credited completions produced parseable experiment_result payloads, indicating stable end-to-end experiment return integrity.
2. Throughput in this credited batch remained dominated by active delay/ecology/control experiment families, with no currently queued high-confidence deterministic crash family requiring broad abort.
3. CPU/GPU mixed validation remained healthy in reviewed outcomes (CPU 468, GPU 11), supporting continued parallel collection across both execution classes.
4. The research pipeline now prioritizes interaction/mechanism tests (label-smoothing interaction, curvature-triggered timing, batch-noise interaction) rather than additional replication of already over-seeded WD main-effect lines.

NOTES
- This log intentionally omits cumulative result-ID inventories; the database is the source of truth for credited state.