Lab

8 experiments · 5 pass · 2 fail · 1 partial

1. pass·M15 · allostatic regulation

   Allostatic regulation under resource pressure

   Agent maintains homeostatic targets under 3x resource scarcity.

   crafter · 45 min · apr 294/4 vs 0/4
2. pass·M11 · language-as-reasoning

   14-word self-learned vocabulary, parity with 3B LLM

   Internal symbolic protocol emerges without external supervision; reaches LLM-equivalent reasoning on held-out tasks.

   babyai · 2 hr · apr 285/5 vs 0/5
3. fail·M07 · feature decomposition

   Width-constrained variant fails to decompose

   Compositional representation collapses; agent memorizes scenes instead of decomposing.

   crafter · 1 hr · apr 280/4 vs 0/4
4. partial·M03 · gene-as-bias

   Compressed inheritance recovers 95% of full inheritance

   Compounding learning preserved under aggressive inheritance compression.

   babyai · 3 hr · apr 273/4 vs 0/4
5. pass·WMCP · cross-modal agreement

   V-JEPA 2 ↔ TinyLlama symbol convergence

   Two unaligned encoders agree on 92% of discrete tokens on a held-out task.

   wmcp · 6 hr · apr 2692.0%
6. pass·M01 · edge of chaos

   Binary phase transition, world-vs-agent disorder

   Confirms the substrate sustains complexity, not the policy. Agent stays critical near the boundary.

   crafter · 1.5 hr · apr 252/2 vs 0/2
7. fail·M11 · language-as-reasoning

   Vocabulary scaling beyond 32 tokens

   Symbol space saturates; reasoning collapses to surface-level pattern match.

   babyai · 4 hr · apr 240/5 vs 0/5
8. pass·WMCP · cross-modal agreement

   DINOv2 ↔ wav2vec2 protocol stability

   Visual and audio encoders converge on shared discrete protocol; 97.5% agreement on a held-out object set.

   wmcp · 5 hr · apr 2397.5%