AXIOM SESSION RESULTS - 2026-03-04 00:33
=========================================

SESSION OVERVIEW
- Completed end-to-end validation/deploy/review cycle for session `s0304_part1_validate` + `s0304_part2_deploy`.
- Reviewed 617 completed `exp_%` results across 4 moving-queue validation passes.
- Credit plan remained under safety cap (planned 2,641; cap 10,000).
- Likely session-applied credit from validation transaction set: 403 results, +1,671 total credit.
- Residual uncredited rows at final validation check were live new arrivals during runtime.

RESULTS REVIEWED AND CREDIT AWARDED
- Bulk review covered interaction/mechanism experiments plus legacy-tail completions; payload quality was valid in reviewed rows (no missing/parse-error `experiment_result` payloads in this reviewed batch).
- Highest-volume reviewed families included:
  - `exp_wd_imbalance_interaction`
  - `exp_wd_labelsmooth_interaction` / `exp_wdlsint`
  - `exp_wd_batchnoise_interaction` / `exp_wdbnoise`
  - `exp_wd_anisotropy_trigger_gpu_gpu`
- Credited users this session (net increase):
  - ChelseaOilman +869
  - WTBroughton +531
  - Steve Dodd +141
  - kotenok2000 +45
  - zombie67 [MM] +31
  - Anandbhat +24
  - marmot +15
  - Vato +10
  - [DPC] hansR +5

STUCK/BROKEN TASK CLEANUP
- Broken legacy lines `exp_power_law_forgetting` and `exp_power_law_forgetting_v2` were remediated by aborting 6 active tasks (including second sweep for new arrivals).
- No dead-host stuck-task candidates were found under >12h runtime and >6h no-contact criteria.
- No >48h running experiment tasks were found.

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK DOCUMENTATION
1. `wd_curriculum_interaction.py` (CPU)
- Novelty check queries included curriculum learning, anti-curriculum learning, and WD-schedule interaction prior art.
- Prior work exists on curriculum effects, but no direct controlled late-WD timing interaction test with sample-order regime in this BOINC compositional-shift setting.
- Novel contribution: causal interaction test of WD timing gains vs representation-formation pace induced by data ordering.

2. `wd_noise_trigger_gpu.py` (GPU-aware)
- Novelty check queries included gradient-noise scale, adaptive/scheduled WD, and variance-schedule regularization prior art.
- Prior work exists on adaptive WD and gradient-noise concepts, but no direct trigger rule using noise-proxy stabilization versus fixed-late WD in this replicated BOINC framework.
- Novel contribution: mechanism test evaluating noise-stability as causal gate for WD activation.

DEPLOYMENT SUMMARY (PART 2)
- CPU workunits: 4,380 total (2,231 targeted + 2,149 fallback).
- GPU workunits: 561 total (280 targeted + 281 fallback).
- GPU idle-capacity hosts detected: 78.
- GPU-aware scripts deployed included anisotropy/curvature/timing-scale/subspace/noise-trigger lines.

KEY SCIENTIFIC FINDINGS
1. `wd_imbalance_interaction` remained strongly positive in this session’s reviewed batch, with mean interaction excess around +0.095 and minority-recall lift around +0.205 versus CE baseline.
2. `wd_labelsmooth_interaction` / `wdlsint` continued to show high interaction-presence rates (~94-95%) while attenuation held in a minority subset (~25%), consistent with non-universal damping.
3. `wd_batchnoise_interaction` / `wdbnoise` retained near-saturated positive interaction rates (~0.999) with positive mean gain-difference signal in reviewed runs.
4. `wd_anisotropy_trigger_gpu` showed stable trigger timing (epoch mean ~24) with only small adaptive-vs-fixed effect magnitude (near-zero to slight positive), supporting a nuanced mechanism rather than broad gain.
5. Across validation and deployment, inverse critical-period WD timing behavior remains supported, motivating continued falsification-oriented mechanism tests.

OPERATIONAL NOTES
- Over-seeded active-running lines still exist in catalog, but unsent count was 0 at retirement check time, so no additional abort-update pass was required in deployment step.
- Targeted dispatch can still underfill intermittently; untargeted fallback remained necessary to maintain host utilization.