SPECTER LABS

Lenia TRIBE Overlay: Findings (Incubation)

Status as of 2026-05-04. The addendum is staged: pipeline is wired end-to-end through correlation, sanity gate passes, but no real (non-fake) scoring batch with warehouse-linked specimens has been run yet.

What was built

  • TRIBE v2 sanity gate (lenia-tribe-sanity). Passes on 6 OOD probes (variance > 1e-4 floor) for facebook/tribev2. Hard-errors on variance collapse with a diagnostic, no fallback path.
  • ROI bundle on fsaverage5 via Destrieux atlas labels: sts, lateral_ot, v1_proxy. Masks are anatomical proxies, not functional.
  • ROI probe (lenia-tribe-roi-probe) caches per-vertex predictions for the OOD probe set under .artifacts/predictions/ for control-row reuse.
  • Corpus loader resamples arbitrary-length MP4s to TRIBE’s 32-frame, 8-fps probe window via even-spaced linear sampling and PIL bilinear resize.
  • Scoring CLI (lenia-tribe-score) accepts either --corpus (ad-hoc) or --manifest (warehouse-linkable). Manifest mode records specimen_id per row; corpus mode records null and is skipped by the overlay.
  • Overlay CLI (lenia-tribe-overlay) joins a score report against the lenia-swarm morphospace.duckdb warehouse on specimen_id, fetches the 16-axis lenia_terminal_v1 descriptor per specimen, and emits one row per linked specimen with ROI scores plus descriptor coordinates.
  • Correlation CLI (lenia-tribe-correlate) computes Pearson r for every (ROI, axis) pair, prints a matrix, and tags each ROI as REDUNDANT (any abs(r) >= threshold, default 0.85) or candidate-new.

Framing pivot

First probe round used a synthetic point-light walker as a positive bio-motion control. STS engagement went the wrong direction. Instead of chasing a clean control (Vanrie & Verfaillie etc.), we rescoped: ROI engagement is a relative score over Lenia creatures.

The question we keep is whether the TRIBE score is a duplicate of one of the 16 lenia_terminal_v1 axes or actually new. High correlation with an existing axis means drop it. Low correlation with all 16 means keep it. The score is not a “lifelikeness” measurement.

Smoke run (fake client)

End-to-end pipeline verified through correlation: lenia-tribe-score --fake on a 6-entry manifest, lenia-tribe-overlay against the real warehouse, then lenia-tribe-correlate on the resulting overlay. Five rows linked, 16 axes per row, 48-cell correlation matrix produced and verdict line emitted per ROI. Numbers are not interpretable (fake client emits class-conditioned constants plus tiny noise; n=5 is far below what would stabilize r).

What blocks the first real comparative analysis

The lenia-swarm warehouse and the renders/ MP4 tree are decoupled:

  • 514 specimens have lenia_terminal_v1 features and an export_dir of JSON manifests (base.json, meta.json, search.json); no MP4 alongside.
  • MP4s in renders/ are keyed on QD-cell labels and scene names, not specimen_ids. There is no automated linkage.

To produce a real warehouse-linked scoring batch we need to drive the lenia-swarm replay and media pipelines from a curated specimen list:

  1. Write an index.jsonl enumerating chosen specimen_ids and their export_dirs (read straight from morphospace.duckdb).
  2. Run swift run lenia-cli replay --input <index.jsonl> ... to materialize replay-capable run dirs per specimen.
  3. Run swift run lenia-cli media --input <replay-dir> ... to render MP4s labelled per specimen.
  4. Build a manifest mapping {name, mp4, specimen_id} and feed it to lenia-tribe-score --manifest.

Until that pipeline is driven, the “TRIBE-engagement vs descriptor” plot is not interpretable.

Open questions for the first real batch

  • How many specimens to render. ~30-50 covers the warehouse’s family diversity without taking days on CPU TRIBE inference.
  • Replay budget. TRIBE wants a 4-second probe window (32 frames at 8 fps). If replay produces shorter native trajectories we have to either pad or reject; the corpus loader already hard-errors on under-length sources.
  • Whether to mix scene composites with single-creature renders. Single-creature is cleaner for the warehouse join (one specimen per MP4); composites are interesting but conflate ROI signals across specimens.

Limitations carried forward

  • TRIBE predicts fMRI, not perception; predicted activation is evidence of model-internal engagement, not conscious perception.
  • Single observer: TRIBE averages 720 subjects into one mapping.
  • OOD substrate: Lenia is far from TRIBE’s training distribution. The sanity gate is necessary but not sufficient.
  • ROI masks are anatomical (Destrieux), not functional. Small-effect contrasts within-ROI are not interpretable at this resolution.
  • License: CC BY-NC; no commercial surface.