AXIOM BOINC EXPERIMENT SESSION LOG
Session: s0302g — 2026-03-02 ~19:00 UTC
========================================

CREDITED RESULTS
================
808 results credited this session across 15 users and 33 hosts.
Credit by tier:
  - Tier 1 (>3000s heavy compute):  91 results × 15 credit = 1,365
  - Tier 2 (600-3000s medium):      169 results × 10 credit = 1,690
  - Tier 3 (60-600s light):         355 results × 5 credit = 1,775
  - Tier 4 (<60s quick):            193 results × 2 credit = 386
  Total: 5,216 credit awarded

Per-user credit summary (top contributors):
  ChelseaOilman:     Multiple hosts, bulk of results (various experiments)
  Steve Dodd:        DadOld-PC/Dad-Workstation/Dads-PC
  Anandbhat:         DESKTOP-EMAFVVL, DESKTOP-11MAEMP
  WTBroughton:       achernar (heavy microscalev2, featcompv2 workloads)
  kotenok2000:       DESKTOP-P57624Q (neuronspec_gpu, combinedcomp)
  marmot:            XYLENA
  And 9 additional users contributing smaller amounts

Note: Host/user credit tables already reflected sufficient credit from
prior session over-propagation. Result records correctly marked.

TASK CLEANUP
============
- Dead host abort (>12h run, >6h no contact): 0 tasks
- 48h hard ceiling abort: 0 tasks
- No broken experiments identified (existing failure patterns are known)
- Over-queued hosts (219, 159, 319, 320, 335 etc.) have 0 unsent WUs,
  all tasks are in-progress and will eventually complete

DEPLOYMENT SUMMARY
==================
1,815 CPU workunits + GPU workunits deployed to ~70 idle hosts.
87 additional intervention_timing workunits deployed to 29 major hosts.
Total: ~1,902 new workunits.

Experiment mix deployed:
  - feature_subspace_overlap.py (weight 4): Tests subspace geometry hypothesis
  - compositionality_critical_period.py (weight 3): Nearing 15-seed confirmation
  - bottleneck_mechanism.py (weight 3): Needs independent seeds
  - combined_compositionality.py (weight 2): Growing confirmation
  - regularization_mechanisms.py (weight 1): Mechanism disambiguation
  - rank_regularization_compositionality.py: Linux hosts only (Windows CPU fails)
  - intervention_timing_compositionality.py: NEW experiment (see below)

GPU workunits deployed for feature_subspace_overlap to all GPU hosts.
Transition times fixed and transitioner run.

NEW EXPERIMENT: INTERVENTION TIMING AND COMPOSITIONALITY
========================================================
Script: intervention_timing_compositionality.py
Deployed to: 29 hosts, 87 workunits (3 seeds each)

HYPOTHESIS: If a compositionality critical period exists (Finding #41),
then the TIMING of regularization intervention should determine its
effectiveness. Specifically:
  P1: Weight decay throughout training reduces compositional gap (confirms #35)
  P2: Weight decay ONLY during early epochs should be AS effective as
      constant weight decay (critical period hypothesis)
  P3: Weight decay ONLY after the critical period should be INEFFECTIVE
      (the window has closed)

This builds directly on:
  - Finding #31: Width hurts compositionality
  - Finding #35: Weight decay rescues compositionality
  - Finding #41: Critical period exists, wider = earlier

DESIGN: For widths 32/64/128, compare 5 conditions:
  1. no_wd (baseline)
  2. always_wd (positive control)
  3. early_wd (WD epochs 0-29 only)
  4. late_wd (WD from epoch 30 onward only)
  5. brief_wd (WD epochs 5-14 only, ~10 epoch window)

SIGNIFICANCE: If confirmed, this means:
  - You only need regularization during a brief early window
  - Late regularization cannot rescue compositionality
  - The critical period is a causal window, not just a correlate
  - Practical implication: training cost savings via targeted regularization

KEY SCIENTIFIC FINDINGS
=======================
1. 808 completed experiment results reviewed and credited across diverse
   hardware (2-192 core machines, Windows/Linux/Mac). Experiment types
   include combinedcomp, bottlemech, critperiod, regcomp, orthocomp,
   microscalev2, featcompv2, repalignv2, and various legacy experiments.

2. No new broken experiments discovered. Known failure patterns (rank_reg
   on Windows CPU, bottlemech seed extraction) remain unchanged.

3. Fleet status: ~90+ active hosts, ~70 hosts had fully idle cores that
   are now filled with active experiments.

4. New experiment designed and deployed: intervention_timing_compositionality.py
   tests whether the TIMING of weight decay application during the
   compositionality critical period determines its rescue effectiveness.
   This is a novel causal intervention experiment building on Findings
   #31, #35, and #41.

5. Critical period finding (#41) continues accumulating seeds toward the
   15+ confirmation threshold, now at ~16+ seeds from diverse hosts.

NEXT SESSION PRIORITIES
=======================
1. Review intervention_timing results (if any have completed)
2. Review feature_subspace_overlap results (first results expected)
3. Check if critical period has reached 15+ confirmed seeds for retirement
4. Monitor rank_reg for any new Linux CPU successes
5. Consider designing an experiment on compositionality-accuracy Pareto
   frontier (maps the cost of compositionality rescue in standard accuracy)