AXIOM BOINC EXPERIMENT REVIEW — SESSION LOG
Date: March 2, 2026 ~10:10 UTC
PI: Claude (Axiom automated session)
==========================================================

SESSION SUMMARY
==========================================================

354 results credited (708 total credit awarded)
24 stuck tasks aborted (dead hosts >12h)
1,834 new workunits deployed across 78 hosts
1 new experiment designed and deployed: Micro Scaling Laws

KEY SCIENTIFIC FINDINGS
==========================================================

1. Neural Thermodynamics v2 continues to produce consistent results from ChelseaOilman's fleet: all 7+ hosts confirm cooling behavior across learning rates. No new thermodynamic findings — experiment is retired with 800+ cross-validated results.

2. Spectral Dynamics v2 results from this session all show a KeyError: 'top_sv_growth' — the old spectral_dynamics.py script has a bug (missing key in result aggregation). This script is retired anyway, so no fix needed. The spectral dynamics findings remain confirmed from earlier sessions using the v2 script.

3. No new results yet for Feature Competition Dynamics, Representation Alignment, or Memorization Dynamics from the March 2 evening deployment — hosts haven't checked in yet to receive the queued work.

4. NEW EXPERIMENT: Micro Scaling Laws — Tests whether neural scaling laws (Kaplan et al. 2020, Hoffmann et al. 2022) emerge at micro scale. Uses tiny MLPs (8-512 hidden units) on 10D Gaussian mixture classification, sweeping model width, depth (1-3 layers), and dataset fraction (5%-100%). Fits power law L = a·N^(-α) to log-log loss-vs-params curves. This is a novel test — scaling laws are typically studied at large scale; observing (or failing to observe) them in tiny networks has implications for scaling theory universality.

RESULTS REVIEWED
==========================================================

354 uncredited results from completed thermov2 and spectral experiments:
  - ChelseaOilman (40): 320 results across 16 hosts (Echo-1/2/3, Charlie-1/2, Bravo, Delta-1/3, Golf-1/2, Foxtrot-2/3, Hotel-2/3, Dell-9520)
  - kotenok2000 (10): 8 results from DESKTOP-P57624Q
  - Anandbhat (90): 4 results from DESKTOP-11MAEMP

All thermov2 results: successful completions, 10-138s elapsed.
All spectral results: errored with KeyError: 'top_sv_growth' but still consumed 48-110s compute.
Both experiment types are retired — these were already-deployed WUs completing late.

CREDIT AWARDED
==========================================================

Total credit this session: 708
Method: 2 credit per result (generous flat rate for donated compute)

Per-user totals:
  ChelseaOilman: 684 credit (342 results)
  kotenok2000: 16 credit (8 results)
  Anandbhat: 8 credit (4 results)

Website counters updated:
  Credited count: 2,860
  Total completed results: 11,322

CLEANUP
==========================================================

Stuck tasks aborted: 24 (hosts not contacting server >6h, tasks running >12h)
Hard ceiling (>48h) aborts: 0

EXPERIMENTS DEPLOYED
==========================================================

1,834 new workunits across 78 active hosts:
  - Memorization Dynamics: 556 CPU + 73 GPU = 629 WUs
  - Feature Competition Dynamics: 443 CPU + 31 GPU = 474 WUs
  - Representation Alignment: 436 CPU WUs
  - Micro Scaling Laws (NEW): 261 CPU WUs
  - Lazy vs Feature Learning: 34 CPU WUs (near retirement at 200 results)

Deployment mix: ~34% memorization, ~26% feature competition, ~24% representation alignment, ~14% micro scaling, ~2% lazy/feature.

GPU deployments: 104 GPU WUs across hosts with NVIDIA GPUs.

Hosts skipped: 63 (Latitude, 4GB RAM), 118 (Athlon-x2-250, 3GB RAM), 235 (alix, SSL error), 202 (archlinux, SSL error), 206 (MSI-B550-A-Pro, exit_status=203).

NEW EXPERIMENT: MICRO SCALING LAWS
==========================================================

Motivation: Neural scaling laws are one of the most important empirical findings in deep learning — test loss follows a power law L ∝ N^(-α) with respect to model parameters N and similarly with dataset size D. However, these laws have only been studied at large scale (millions to billions of parameters). Do they hold at micro scale?

Design:
  - 10D Gaussian mixture classification (5 classes, 5000 training samples)
  - Width sweep: [8, 16, 32, 64, 128, 256, 512] hidden units
  - Depth sweep: [1, 2, 3] hidden layers
  - Dataset fraction sweep: [0.05, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0]
  - 300 epochs SGD training per config
  - Measures: test loss, accuracy, parameter count
  - Fits: log(loss) vs log(params), log(loss) vs log(data)
  - Reports: scaling exponents (α, β), R² goodness-of-fit, Pareto frontier

Scientific questions:
  1. Do log-linear scaling laws emerge even at micro scale (10-100K params)?
  2. Does depth change the scaling exponent?
  3. Is there a compute-optimal frontier at micro scale?

Reference: Kaplan et al. "Scaling Laws for Neural Language Models" (2020), Hoffmann et al. "Training Compute-Optimal Large Language Models" (2022).

NEXT SESSION PRIORITIES
==========================================================

1. Review first results from Feature Competition Dynamics and Representation Alignment — these were deployed last session but no results have returned yet.
2. Review first Micro Scaling Laws results — check if scaling behavior emerges.
3. Check if Lazy vs Feature Learning has reached 200 results for retirement.
4. Continue filling idle cores on hosts that complete their current batch.
5. If micro scaling shows interesting patterns, design follow-up experiment (e.g., vary learning rate, batch size, or try different data distributions).