SPECTER LABS

Lenia TRIBE Overlay: Design Overview

Question

lenia-swarm already gives each creature 16 shape-and-motion numbers (lenia_terminal_v1: spread, compactness, boundary complexity, symmetry, displacement, etc.). This addendum computes a 17th number a totally different way: feed the video to TRIBE v2 and average the predicted cortical activation in a few regions.

Is the 17th number basically a duplicate of one of the 16, or actually new? If it correlates ~0.9 with boundary_complexity, drop it. If it’s uncorrelated with all 16, keep it as a new way to sort creatures.

We do not claim Lenia creatures are alive because the cortex says so, and the score is not a “lifelikeness” measurement.

Tool

TRIBE v2 is a multimodal transformer (LLaMA 3.2 + V-JEPA2 + Wav2Vec-BERT) trained on >1000 hours of fMRI from 720 subjects to predict cortical responses to video, audio, and language. We use only the video pathway; audio and text are stubbed to silence and empty string. The model emits per-voxel predictions on fsaverage5 (20484 vertices, [left|right] hemispheres stacked).

Pipeline

  1. Sanity gate. Push a small set of OOD-but-controlled probes through TRIBE and confirm whole-cortex predictions vary across them. Hard-error on variance collapse.
  2. ROI bundle. Build named anatomical-proxy masks on fsaverage5 via the Destrieux atlas: an STS region, a lateral-occipitotemporal region (proxy for MT / EBA), and a V1 region as control.
  3. Lenia corpus loading. Existing lenia-swarm MP4 renders are re-timed to TRIBE’s 32-frame, 8-fps, 4-second probe window — the brain does not know what one Lenia timestep means in seconds, so the entire source clip is compressed into the probe window.
  4. Inference. Forward pass per stimulus through TRIBE. Per-vertex predictions persist to .artifacts/predictions/. ROI summaries persist alongside under .artifacts/lenia-scores/.
  5. Warehouse overlay. When a scoring run supplies a manifest with specimen_id per Lenia clip, lenia-tribe-overlay joins the ROI scores against the lenia_terminal_v1 feature space in the lenia-swarm morphospace warehouse and emits one row per linked specimen with ROI scores plus the 16 descriptor axes. Manifest entries without specimen_id are scored and recorded but skipped by the overlay.
  6. Comparative analysis. lenia-tribe-correlate computes Pearson r between each ROI score and each of the 16 lenia_terminal_v1 axes, prints the matrix, and tags each ROI as REDUNDANT (any abs(r) >= threshold, default 0.85) or candidate-new.

Determinism and provenance

Every inference run records: TRIBE weight hash, codebase commit, stimulus manifest derived from content-addressed MP4s, torch and CUDA versions, seed for any preprocessing randomness. Reports are timestamped under .artifacts/.

Limitations recorded up front

  • TRIBE predicts fMRI, not perception. A predicted activation in any ROI is evidence of model-internal engagement, not of conscious perception.
  • Single observer. TRIBE averages 720 subjects into a learned mapping; individual and cultural differences are absent.
  • OOD substrate. Lenia is far from TRIBE’s training distribution. The sanity gate is necessary but not sufficient guard against garbage.
  • ROI masks are anatomical (Destrieux), not functional. sts, lateral_ot, and v1_proxy are loose proxies; small-effect contrasts at this resolution are not interpretable.
  • License. CC BY-NC. No downstream commercial surface.