AXIOM BOINC SESSION RESULTS
Session timestamp: 2026-03-04 00:40 CET (validation) / 2026-03-04 deployment cycle
Session label(s): s0304_part1_validate_b, s0304_part2_deploy_b

RESULTS REVIEWED AND CREDIT AWARDED
- Reviewed completed uncredited exp results: 134
- Payload summary: 110 science payloads, 24 explicit error payloads, 0 missing JSON, 0 parse errors.
- Credit policy applied: elapsed-time tier for science payloads; reduced-tier compensation for error payloads.
- Total incremental credit awarded: +1007 (safety cap check PASS, <=10000)
- Database updates: 134 result rows, 6 user rows, 15 host rows.
- Per-user awards this session:
  - ChelseaOilman: +612
  - WTBroughton: +234
  - Steve Dodd: +95
  - Armin Gips: +36
  - zombie67 [MM]: +18
  - marmot: +12

STUCK/BROKEN TASK CLEANUP
- Broken stale line `exp_regtimuniv_gpu_g1%` cleaned: 19 unsent tasks aborted (server_state=5, outcome=5).
- Dead-host stuck cleanup (>12h running and >6h host silence): 0 tasks.
- Hard-ceiling cleanup (>48h running): 0 tasks.

NEW EXPERIMENTS DESIGNED AND DEPLOYED
- New experiment script added: `wd_clipping_interaction.py` (CPU, numpy-only, run-duration compliant).
- Conditions tested: baseline, late_wd, clip_only, clip_late_wd, always_wd_clip.
- Primary metric: interaction_delta = (late_wd_gain_with_clip - late_wd_gain_no_clip).

NOVELTY CHECK DOCUMENTATION
- Literature/query sweep performed for:
  - weight decay x gradient clipping interaction
  - delayed WD schedule prior work (including arXiv:1905.08694)
  - adaptive gradient clipping prior work (including arXiv:1905.11881)
  - weight decay schedule + clipping combinations
- Novelty conclusion: prior work treats clipping and decay mostly separately; no direct replicated controlled test found for late-WD timing x clipping interaction under compositional covariate shift.
- Mechanistic hypothesis targeted: late WD may act via suppression of rare high-norm updates that clipping could partially substitute.

DEPLOYMENT SUMMARY (PART 2)
- CPU targeted pass: 2,181 WUs
- CPU top-up pass: 2,153 WUs
- CPU new-experiment injection (`wd_clipping_interaction`): 234 WUs
- CPU total this session: 4,568 WUs
- GPU targeted pass: 278 WUs
- GPU top-up pass: 276 WUs
- GPU total this session: 554 WUs
- GPU hosts with idle capacity found: 74

KEY SCIENTIFIC FINDINGS
1. `wd_imbalance_interaction` remained strongly positive in this increment (+40 runs): mean interaction_excess = +0.0949 and mean minority-recall lift (late_wd_class_balanced - CE) = +0.2053.
2. `wd_labelsmooth_interaction` (+9 runs) supported interaction in 100.0% of runs, with attenuation predicted in 66.7%; mean additivity residual = +0.0020 (near-additive at batch level).
3. `wd_anisotropy_trigger_gpu` (+2 GPU runs) stayed near parity with fixed-late WD: adaptive_minus_fixed mean = +0.0005 with stable trigger epoch mean = 24.0.
4. `wd_timing_depth` (+2 runs) was mixed in this increment: inverse-critical-period all-depth flag true in 50.0% and depth-vs-late-effectiveness correlation mean = -0.0286; more replication required.
5. Confirmed program-level status remains consistent: inverse critical period behavior is established; current priority stays on interaction/mechanism tests that can falsify causal timing hypotheses.

OPERATIONS NOTES
- Website counters updated during validation: credited_count advanced by +134; total results counter synchronized to live success maximum.
- No new experiment-script bug requiring immediate patching was identified during this validation batch.