AXIOM SESSION RESULTS LOG - 2026-03-04 00:55
Session context: Part 3 save/upload, based on Part 1 validation + Part 2 experiment operations

RESULTS REVIEWED AND CREDIT AWARDED
- Reviewed 1,392 completed uncredited exp_% results.
- Credited 1,060 results this session under runtime-tiered crediting (science payloads confirmed).
- Total credit awarded: +10,000 (session cap reached, not exceeded).
- Per-user awards this session:
  - ChelseaOilman: +4,152 across 433 results
  - WTBroughton: +2,214 across 248 results
  - Anandbhat: +1,224 across 127 results
  - zombie67 [MM]: +802 across 63 results
  - Steve Dodd: +716 across 66 results
  - Amapola: +452 across 89 results
  - kotenok2000: +312 across 19 results
  - marmot: +128 across 15 results
- Most-credited experiment family this batch: wd_labelsmooth_interaction (+89 credited runs).
- Remaining completed uncredited exp_% results were deferred due to session credit cap.

STUCK/BROKEN TASK CLEANUP
- Dead-host stuck-task cleanup (>12h run and >6h no contact): 0 tasks aborted.
- Hard-ceiling cleanup (>48h runtime): 0 tasks aborted.
- Broken experiment aborts: none (no cross-host deterministic crash pattern requiring blanket abort).

NEW EXPERIMENTS AND NOVELTY CHECK DOCUMENTATION
1) wd_mixup_interaction.py (CPU) designed and deployed.
- Novelty-check queries executed:
  - "weight decay mixup interaction deep learning study"
  - "arxiv mixup: Beyond Empirical Risk Minimization"
  - "AdamW Decoupled Weight Decay Regularization arxiv"
  - "Delayed Weight Decay Regularization arxiv"
- Novelty conclusion: prior work exists on mixup and decoupled/late weight decay individually, but no direct controlled test of late-WD timing x mixup interaction under compositional covariate shift with repeated independent host-seeded BOINC trials.
- Mechanistic target metric: interaction_delta = (late_wd_gain_with_mixup - late_wd_gain_no_mixup).

DEPLOYMENT SUMMARY (PART 2 CONTEXT)
- CPU host-targeted top-up: 2,248 WUs (2 low-RAM hosts skipped).
- CPU new experiment injection (wd_mixup_interaction): 240 WUs.
- CPU total created: 2,488 WUs.
- GPU host-targeted top-up: 269 WUs.
- GPU hosts with idle capacity identified: 72.
- GPU-aware scripts used: wd_noise_trigger_gpu.py, wd_curvature_trigger_gpu.py, wd_timing_scale_gpu.py, grad_subspace_wd_gpu.py, wd_anisotropy_trigger_gpu.py.

KEY SCIENTIFIC FINDINGS
1. Inverse critical-period behavior for weight-decay timing remains established; strategy remains focused on mechanism/interaction testing rather than further seed-only replication.
2. In wd_labelsmooth_interaction (newly credited +89 host-runs), the interaction-present prediction held in 86.5% of runs, supporting a positive interaction signal.
3. The attenuation prediction held in only 23.6% of those runs, indicating attenuation is currently weak/inconsistent and should be targeted for follow-up.
4. Mean additivity residual for wd_labelsmooth_interaction in this batch was +0.0393, consistent with a modest positive non-additive component.
5. New wd_mixup_interaction deployments create a direct test of whether late-WD gains are mechanistically redundant with vicinal smoothing under compositional shift.

OPERATIONAL NOTES
- No experiment-script code patch was required from this validation pass.
- Website counters were updated during validation operations.
- Queue-gap readings may transiently lag immediately after WU creation due to scheduler/transition propagation.