Axiom BOINC Session Results Log
Session timestamp: 2026-03-03 22:18 (America/Denver)
Scope: Combined Part 1 (validation/credit/cleanup) + Part 2 (research/deploy)

PART 1 - VALIDATION / CREDIT / CLEANUP
- Validation source log: validate_2026-03-03_2213.txt.
- Results reviewed: latest failed/completed DB rows were inspected (including IDs 1658128, 1658115, 1658014, 1658001, 1657998), and payload audit of 60 newest upload JSON files was performed.
- Credit awarded this session: 0 newly credited results; total credit added = 0 (session cap respected).
- Counter refresh:
  - credited_count.txt remained 54,438
  - total_results_count.txt remained 55,009
- Stuck/broken cleanup:
  - Dead-host >12h cleanup: 0 rows aborted
  - Hard >48h cleanup: 0 rows aborted
  - Global broken-experiment aborts: none
- Operational note: boincadm MySQL auth failed in Part 1; validation proceeded with local MySQL root on server.

PART 2 - DEPLOYMENT / RESEARCH
- Part 2 run log source: run_2026-03-03_1352.log.
- Retirement pass executed against over-seeded candidates; ABORT_TOTAL=0 in this run.

CPU deployment
- CPU scripts used:
  - wd_batchnoise_interaction.py
  - wd_labelsmooth_interaction.py
- CPU deployment outcome recorded in run log:
  - CPU_HOSTS_SEEN=81
  - CPU_SKIPPED_LOW_RAM=2
  - CPU_WU_CREATED=2937
- Deployment policy used in run: target each eligible CPU host toward ~3x CPU queued tasks, skip hosts <6 GB RAM.

GPU deployment
- GPU scripts targeted:
  - wd_curvature_trigger_gpu.py
  - wd_timing_scale_gpu.py
- Run log shows GPU deployment pass started but interrupted before final aggregate counters were printed.
- GPU checkpoint (live queue snapshot after run):
  - Active queued/running GPU workunits for targeted scripts: 7
  - Active GPU hosts for targeted scripts: 4
  - Script split in active queue: wd_curvature_trigger_gpu=0, wd_timing_scale_gpu=7

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK
- New experiment script added and syntax-checked:
  - wd_batchnoise_interaction.py (uploaded to experiments directory; python3 -m py_compile passed)
- Purpose:
  - Test whether late weight-decay gain interacts with batch-size-driven gradient noise (interaction effect, not only main effects).
- Novelty check evidence documented in run log (queries executed):
  - 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
- Novelty conclusion used for deployment decisions:
  - Prior literature covers decoupled WD and batch-size generalization effects separately, but this specific late-WD x batch-noise interaction framing was treated as sufficiently novel for Axiom deployment.

KEY SCIENTIFIC FINDINGS
1. Recent upload payloads continue to contain valid experiment_result structures across active families, including long-runtime runs (~840-918s), indicating continued science throughput rather than widespread execution corruption.
2. Multiple rows marked outcome=5 still carry valid experiment_result payloads, reinforcing a validator/outcome classification mismatch signal rather than uniform experiment failure.
3. No new evidence in this session contradicts prior inverse critical-period weight-decay timing conclusions.
4. The wd_batchnoise_interaction line extends the mechanism program by explicitly testing whether late-WD benefit depends on batch-noise regime, creating a new high-value axis for causal discrimination.

FILES UPDATED THIS SESSION
- New session log: results_2026-03-03_2218.txt
- Context memory retained/updated separately: findings_summary.txt