AXIOM BOINC EXPERIMENT SESSION LOG
Session: s0302p — March 2, 2026 ~22:30 UTC
===========================================

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

1. REGULARIZATION TIMING UNIVERSALITY — PARTIAL CONFIRMATION (Finding #47 updated):
   New result from Delta-1 (seed=1652859793, independent) reveals that the inverse critical
   period is NOT universal across all regularizers — it is PARTIAL, applying to 2 of 4 tested:

   - WEIGHT DECAY: Inverse CP confirmed (late beats early at all widths, early hurts).
   - DROPOUT: Inverse CP confirmed (late beats early at all widths, early hurts).
     At w32: no_reg=0.710, always=0.440, early=0.724, late=0.540.
     At w128: dropout less effective overall but inverse pattern still holds.
   - L1 REGULARIZATION: No inverse CP — L1 is NOT effective for compositionality at all.
     Gaps ~0.73 across all conditions, no meaningful variation.
   - NOISE INJECTION: No inverse CP — noise is NOT effective for compositionality.
     Gaps ~0.70 across all conditions, no meaningful variation.

   INTERPRETATION: The inverse CP is specific to regularizers that compress weight structure
   (weight decay) or reduce representation redundancy (dropout). L1 penalizes weight magnitude
   but doesn't compress rank structure. Noise injection adds stochasticity but doesn't constrain
   representations. This is a more interesting finding than simple universality — it identifies
   the class of regularizers for which timing matters.

2. INVERSE CRITICAL PERIOD (Finding #45) — continued confirmation with new independent seeds
   from Charlie-1 (h324) and Delta-1 (h330). Seed 445848802 from Delta-1 shows:
   w32: no_wd=0.612, always=0.328, early=0.703, late=0.335 (late ≈ always, early hurts)
   w64: no_wd=0.649, always=0.331, early=0.729, late=0.338 (same pattern)
   w128: no_wd=0.616, always=0.361, early=0.702, late=0.408 (same pattern)
   100% consistent with prior 34 seeds.

3. WD REBOUND DYNAMICS (Finding #46) — new result confirms preliminary negative:
   seed=42 (fallback) shows P1=0.417, P2=1.0, P3=0.667, P4=-0.404 (negative correlation).
   Rank rebound exists but is NEGATIVELY correlated with compositional gap, contradicting
   the simple rebound hypothesis. The mechanism is likely more subtle.

RESULTS REVIEWED & CREDIT AWARDED
==================================
Total results credited: 473
Total credit awarded: 6,830
Credit cap remaining: 3,170

By user:
  ChelseaOilman: 262 results, 3,755 credit (hosts: Dell-XPS-15-9560, Dell-9520, Hotel-3,
    Charlie-1, Delta-1, Golf-1, Foxtrot-2, Foxtrot-3, and many more)
  Anandbhat: 91 results, 1,155 credit (host: DESKTOP-EMAFVVL)
  Steve Dodd: 38 results, 580 credit (host: Dads-PC)
  WTBroughton: 29 results, 555 credit (host: achernar)
  kotenok2000: 18 results, 310 credit (host: DESKTOP-P57624Q)
  PyHelix: 23 results, 300 credit (host: Pyhelix)
  Vato: 7 results, 105 credit (host: iand-r7-5800h)
  dthonon: 5 results, 70 credit (host: thonon-meylan)

By experiment type (top 10):
  Retired experiments (compgen, featcompv2, microscalev2, neuronspec, etc.): ~350 results
  intervtiming: 33 results (new independent seeds from Charlie-1, Delta-1, Pyhelix, DESKTOP-P57624Q)
  regtimuniv: 18 results (new seeds from Charlie-1, Delta-1, DESKTOP-P57624Q)
  wdrebound: 11 results
  bottmech/bottlemech: 11 results
  regmech: 10 results
  rankreg: 10 results (from Dads-PC)
  critperiod: 9 results

CLEANUP ACTIONS
===============
- Aborted 2 stuck tasks from dead hosts (>12h running, >6h no contact)
- 0 tasks over 48h ceiling
- No new broken experiment patterns detected in active experiments

DEPLOYMENT
==========
Total WUs deployed: 1,596 CPU workunits to 73 hosts
No GPU workunits deployed (none of our active scripts use cupy/HAS_GPU)

Experiment distribution:
  representation_crystallization.py: 554 WUs (35%) — FIRST DEPLOYMENT of Finding #48
  reg_timing_universality.py: 486 WUs (30%) — highest priority for more seeds
  intervention_timing_compositionality.py: 245 WUs (15%) — continued replication
  regularization_mechanisms.py: 166 WUs (10%) — needs diverse seeds
  wd_rebound_dynamics.py: 92 WUs (6%) — mechanism test
  bottleneck_mechanism.py: 53 WUs (3%) — low priority, seed=42 issue

Top hosts by WUs deployed:
  DESKTOP-N5RAJSE (192-core): 192 WUs
  SPEKTRUM (72-core): 72 WUs
  DadOld-PC (80-core): 68 WUs
  JM7 (64-core): 64 WUs
  20+ hosts with 32 WUs each

EXPERIMENT DESIGN REASONING
============================
1. REPRESENTATION CRYSTALLIZATION (35% of deployment):
   This is our newest experiment (Finding #48) and needs initial seeds urgently.
   It tests the MECHANISM behind the inverse critical period — the hypothesis that
   networks need unrestricted exploration to discover compositional features before
   regularization can refine them. If confirmed, this explains WHY late WD works.

2. REGULARIZATION TIMING UNIVERSALITY (30% of deployment):
   The partial universality finding (WD + dropout show inverse CP; L1 + noise don't)
   is extremely interesting and needs replication across many seeds and hosts.
   This is potentially our second most publishable finding after #45 itself.

3. INTERVENTION TIMING (15%):
   Already strongly confirmed with 34+ seeds, but additional seeds from new hosts
   continue to strengthen the statistical case for publication.

4. REGULARIZATION MECHANISMS (10%):
   Still at seed=42 only. Needs diverse seeds to determine whether nuclear norm
   and other regularizers have consistent effects on compositionality.

5. WD REBOUND (6%) and BOTTLENECK MECHANISM (3%):
   Lower priority. WD rebound shows preliminary negative results (rank rebound
   is negatively correlated with gap improvement). Bottleneck mechanism still has
   seed extraction issues.

NEXT STEPS
==========
- Await representation crystallization results (first results expected in 1-2 hours)
- Monitor regtimuniv for more seeds confirming partial universality
- Consider designing a new experiment testing WHY dropout shows inverse CP but L1 doesn't
  (hypothesis: weight magnitude compression vs sparsity induction)
- Investigate whether seed extraction can be fixed in bottleneck_mechanism.py
- When repcrystal results arrive, analyze crystallization timing and diversity metrics