The brain-benchmarking seam: nine methods, one boundary¶

TorchLens draws one explicit identity boundary for brain-benchmarking work (the seam): geometry ours, inferential statistics theirs. For each of Net2Brain's nine evaluation methods, TorchLens supplies every geometric prerequisite -- features, RDMs, alignment, interchange -- and refuses every inferential step with a typed error naming the tool that owns it. This page is the ledger of that boundary (brainpipe). Every TorchLens name here is experimental and may change in any release.

The working surfaces behind the "we supply" column:

• torchlens.neuro.plan_extraction / ExtractionPlan.run -- the planned whole-model sweep: two measured probes, a printable plan table (measured bytes, peak pairs with backend named, exact pass counts), and a run door onto the atomic, resumable extraction artifact.
• torchlens.neuro.export_npz -- npz interchange, including Net2Brain's per-stimulus layout with their lexicographic file-naming contract and a name-to-stimulus-id sidecar.
• tl.repgeom.rdm -- euclidean / manhattan / cosine / correlation RDMs, GPU compute device, row-blocked memory-bounded computation, condensed output, explicit batched [B, N, D] input.
• tl.repgeom.compare_rdms -- DESCRIPTIVE Pearson/Spearman/Kendall tau-a over aligned strict upper triangles (diagonal excluded). No p-values, no noise ceilings: asking for them refuses naming rsatoolbox.
• tl.repgeom.compare_rdms_interval / tl.repgeom.cka_interval -- the same scores with a DESCRIPTIVE percentile bootstrap interval over stimuli beside them (ScoreInterval: seeded, resample count a parameter). Not a test and not rsatoolbox's corrected bootstrap.
• tl.stats.CKA / tl.stats.cka -- one-shot and streaming linear CKA (memory-bounded over arbitrarily many stimuli), device=/dtype=.
• tl.neuro.compare(source, target, analyses=("rsa", "ridge", "cka")) -- one call that reads each site once and scores it against one target with descriptive RSA, a cross-validated ridge score and linear CKA, optional bootstrap intervals, and .disagreement() (tl.repgeom.metric_disagreement) for where the analyses rank sites differently. Descriptive only.
• tl.bridge.rsatoolbox.dataset -- the per-site rsatoolbox handoff.
• tl.bridge.brain_score -- the Brain-Score activations-model seam: a PytorchWrapper-shaped model over an ActivationsExtractorHelper that a model plugin hands to ModelCommitment; one neural and one behavioral benchmark were scored end to end through such a plugin (2026-10-14, brainscore-vision 2.4.1, Python 3.11, CPU). Brain-Score computes the scores; TorchLens supplies the activations. per_layer and score_parts report a ceiling-normalised score only beside its raw score (ScoreParts.raw, .ceiling_normalised, .ceiling, .error).

Two rows are genuinely stronger than the competition and are safe to lead with: CT-RSA (row 3: pass-qualified per-timestep RDMs fall out of TorchLens's recurrence grouping with zero per-model work) and stacked encoding (row 9: its prerequisite -- a complete, honestly-labeled, identically-ordered per-layer feature bank -- IS the sweep).

What the trade is, stated honestly (memo section 1): module-hook extractors (DeepJuice) have no capture floor and no overhead multiple precisely because module hooks do not see functional ops, fused attention, or per-pass recurrence. TorchLens pays a capture cost and in exchange sees every operation, prints a measured cost table before running, emits a manifest a reviewer can audit, and never silently drops a site. Coverage and provenance are the trade; this page names it rather than claiming dominance on both axes.

Which measure, exactly: similarity-repository identifiers¶

The similarity-repository is the catalogue for representational-similarity measures, including the ones TorchLens does not ship. Each TorchLens measure below carries the catalogue identifier of the exact variant it computes (* stands for the implementing repository, so similarity.make(identifier) selects every catalogued implementation of that variant), in these docs and in the result's provenance where the result has any: rdm_evolution(...) .similarity_repository_id, the similarity_repository_id descriptor of tl.neuro.rdms source mode, tl.stats.CKA.similarity_repository_id, the SimilarityScore that tl.stats.cka and tl.stats.CKA.compute() return (a float with similarity_repository_id; to_dict() keeps it), and tl.repgeom.cka_interval(...).similarity_repository_id (also in its to_dict()). Identifiers were taken from the repository at commit 035b3be and checked numerically against its implementations; a dash means the catalogue has no exact match, and no near match is substituted. The attribute and descriptor names are experimental.