SPECTER LABS

Cortical-Engagement Scoring Contract

This contract pins the experimental commitments of A-013, the lenia-swarm TRIBE overlay addendum. Changes to it require an addendum-internal review.

Frame

TRIBE-engagement is a relative score over Lenia creatures. We do not claim absolute neural realism. The contract is about reproducibility of the scoring procedure, not about cortex-as-ground-truth.

Inputs

  • A TRIBE v2 checkpoint identified by HuggingFace revision hash.
  • A stimulus manifest. Each entry is content-addressed and carries a class label drawn from {lenia, biomotion_positive, scrambled, noise, static, rigid_motion, grating} plus stimulus-specific metadata.
  • For lenia-class stimuli, a reference into the lenia-swarm compendium giving the source trajectory and the morphospace coordinates already computed there. We do not recompute morphospace features inside this addendum; we consume them.

Outputs

For every (checkpoint, stimulus) pair:

  • A vector of predicted per-voxel activations on fsaverage5 (20484 vertices, [left|right] hemispheres stacked per TRIBE’s convention).
  • A reduced vector of per-ROI mean activations for the registered ROI bundle (initial set: sts, lateral_ot, v1_proxy, defined via Destrieux atlas labels in lenia_tribe_overlay.rois).
  • A provenance record (TRIBE revision, codebase commit, torch and CUDA versions, machine identifier).

For lenia-class stimuli with a manifest-supplied specimen_id, an overlay report joins the ROI scores to the lenia_terminal_v1 feature space in the lenia-swarm morphospace warehouse and emits one row per linked specimen with ROI scores and the 16 lenia descriptor axes (normalized_value per feature_values.axis_id). Specimens whose specimen_id is not present in the warehouse cause the overlay to fail loudly; the warehouse is the source of truth for axis values.

Outputs are written under $SPECTER_ARTIFACT_ROOT/lenia-tribe-overlay/ if set, otherwise under addenda/lenia-tribe-overlay/.artifacts/.

Redundancy verdict

For lenia-class stimuli with linked specimen_ids, lenia-tribe-correlate computes Pearson r between every ROI score and every lenia_terminal_v1 axis. An ROI is tagged REDUNDANT if its top abs(r) crosses a threshold (default 0.85), candidate-new otherwise. The threshold and the n_specimens floor (currently 4) are part of the contract because they decide the verdict; changing either requires reissuing prior verdicts.

Sanity gate

A run of the experiment is invalid unless the sanity gate, defined in lenia_tribe_overlay.sanity, has been executed against the same TRIBE checkpoint within the last seven days and emitted a pass record with the same checkpoint hash. The gate confirms only one thing:

  1. Across-stimulus variance over a small OOD probe set exceeds VARIANCE_FLOOR = 1e-4. TRIBE must produce visibly differentiated whole-cortex predictions for visually distinct probes.

If (1) fails, the gate hard-errors with a diagnostic. Region-specific claims (e.g. STS responds more to a creature than to noise) are not part of the gate, because the bio-motion ROIs are ~1-2% of cortex and a localized signal can be drowned out in any whole-cortex aggregate. Those claims belong to the experiment proper.

Determinism

  • Stimulus tensors are precomputed and stored content-addressed; the same stimulus produces the same input tensor every run.
  • TRIBE inference is run with deterministic kernels where supported. Where it is not, the non-determinism is recorded and the experiment reports the standard deviation across N replicate runs.

What this contract does not promise

  • We do not claim the predicted activations correspond to any particular individual human’s response.
  • We do not claim that elevated predicted activation in any ROI for a Lenia stimulus implies that humans perceive that stimulus a particular way. The model predicts a hemodynamic proxy.
  • ROI masks are anatomical (Destrieux), not functional. sts, lateral_ot, and v1_proxy are loose proxies for STS, MT/EBA-lateral OTC, and V1 respectively; small-effect contrasts are not interpretable at this resolution.
  • We do not commit to any specific clustering or significance test upfront. The analysis layer is allowed to evolve. The contract pins what is measured and recorded, not what is concluded.