AXIOM BOINC SESSION RESULTS LOG
Session timestamp: 2026-03-03 21:49:34
Scope: Part 1 validation/credit + Part 2 research/deploy

RESULTS REVIEWED AND CREDIT AWARDED (PART 1)
- Reviewed 27 completed uncredited exp_* rows (all outcome=5, GPU appid=2, elapsed_time=0).
- Credited results this session: 27
- Credit policy used: 10 credit per completed uncredited error row
- Total session credit awarded: 270 (below 10,000 cap)
- Per-user credit awarded:
  - PyHelix: +100
  - ChelseaOilman: +70
  - zombie67 [MM]: +60
  - kotenok2000: +40
- Upload payload availability for reviewed batch: 0/27 JSON payload files present under /opt/axiom_boinc/upload/*/exp_*

STUCK/BROKEN TASK CLEANUP
- Broken active prefixes aborted from in-progress queue:
  - exp_grad_subspace_wd_gpu_gpu*: 7 tasks aborted
  - exp_wd_noise_trigger_gpu_gpu*: 10 tasks aborted
- Dead-host stuck-task cleanup (>12h running, >6h no host contact): 0 aborted
- Hard-ceiling cleanup (>48h runtime): 0 aborted
- Retirement pass during Part 2 found over-seeded families but unsent queue for those families was 0 at execution time (ABORT_TOTAL=0).

DEPLOYMENT SUMMARY (PART 2)
CPU deployment
- CPU scripts targeted: wd_batchnoise_interaction.py, wd_labelsmooth_interaction.py
- CPU hosts seen for fill logic: 81
- Hosts skipped for low RAM (<6GB): 2
- CPU workunits created in deployment pass: 2937
- CPU deployment policy used: fill toward 3x CPU queue per host, assign unassigned experiment types first, then replicate with duplicate-check naming.

GPU deployment checkpoint
- GPU scripts targeted: wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py
- Part 2 run log shows GPU deployment pass started but was interrupted before final totals were printed.
- Live queue checkpoint after session:
  - Active GPU hosts (appid=2, server_state in 1/2/4): 4
  - Active GPU workunits (appid=2, exp_*): 30
  - Active GPU WUs by target prefix:
    - exp_wd_timing_scale_gpu_gpu*: 9
    - exp_wd_curvature_trigger_gpu_gpu*: 0

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK (PART 2)
- New experiment script authored and compiled: wd_batchnoise_interaction.py
- Hypothesis tested: late weight decay gain should be stronger under small-batch gradient noise than large-batch settings (interaction test).
- Novelty/literature checks documented in run log via targeted search queries, including:
  - weight decay batch size interaction neural networks
  - arxiv weight decay batch size interaction deep learning
  - Scheduled Weight Decay paper arxiv 2021
  - 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
- Additional deployment-side script set included wd_labelsmooth_interaction.py (CPU) and GPU probes wd_curvature_trigger_gpu.py / wd_timing_scale_gpu.py.

KEY SCIENTIFIC FINDINGS
1. No new payload-backed scientific signal was extracted from the Part 1 credited batch because all 27 reviewed rows were missing upload artifacts.
2. Reliability evidence remains prefix-specific, not global: two GPU prefixes showed cross-host zero-time crash behavior and were cleaned from active queue, supporting prefix-level gating.
3. No reversal of established inverse critical-period weight-decay timing conclusions was observed in this session.
4. Part 2 introduced a new mechanistic hypothesis test (batch-noise x late-WD interaction) to probe whether gradient-noise regime modulates late-WD generalization gains.

KEY SCIENTIFIC RESULTS CONTEXT
- findings_summary.txt remains the persistent cross-session memory and should be read with this session log for longitudinal interpretation.