{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:XFJSBDYJG62SFBISEF4GAXJRDQ","short_pith_number":"pith:XFJSBDYJ","schema_version":"1.0","canonical_sha256":"b953208f0937b52285122178605d311c0625cb6bb14a00fd456eff25509807e7","source":{"kind":"arxiv","id":"2608.10251","version":1},"attestation_state":"computed","paper":{"title":"Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Mark Oskin","submitted_at":"2026-08-10T21:35:48Z","abstract_excerpt":"A transformer's answer lives on one axis: the direction its unembedding reads. Its intermediate states largely do not, and that off-axis position is usually treated as an obstacle to interpretation. We show it is functional. A 12-layer model computes in two phases. Through the first, every sublayer writes into a subspace held near-orthogonal to the read-out, attention 75 to 96 degrees off it at every depth. Moving attention's values onto the read-out is 64 to 84 times more damaging than a matched random rotation, and the damage is entirely in cross-token mixing: the subspace insulates composit"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2608.10251","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-08-10T21:35:48Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"5977008412d17d8143f17a401a52b2339a646abd607fbc77a4d504d4e98afc32","abstract_canon_sha256":"b7574412ff9490866aaefae2a67e883ad2c3eba68b3e6b61a811708ccc4db2aa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-12T00:23:11.082543Z","signature_b64":"LeUmXXgwDhO6y9wF8eiIfaG/YznY0J4YsyWTUz5HcaeJOyUdR/lyJJ9ETRwgyih0OEgz5s/k4KtxRpCCiVH7Ag==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b953208f0937b52285122178605d311c0625cb6bb14a00fd456eff25509807e7","last_reissued_at":"2026-08-12T00:23:11.080555Z","signature_status":"signed_v1","first_computed_at":"2026-08-12T00:23:11.080555Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Mark Oskin","submitted_at":"2026-08-10T21:35:48Z","abstract_excerpt":"A transformer's answer lives on one axis: the direction its unembedding reads. Its intermediate states largely do not, and that off-axis position is usually treated as an obstacle to interpretation. We show it is functional. A 12-layer model computes in two phases. Through the first, every sublayer writes into a subspace held near-orthogonal to the read-out, attention 75 to 96 degrees off it at every depth. Moving attention's values onto the read-out is 64 to 84 times more damaging than a matched random rotation, and the damage is entirely in cross-token mixing: the subspace insulates composit"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.10251","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2608.10251/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2608.10251","created_at":"2026-08-12T00:23:11.083414+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.10251v1","created_at":"2026-08-12T00:23:11.083414+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.10251","created_at":"2026-08-12T00:23:11.083414+00:00"},{"alias_kind":"pith_short_12","alias_value":"XFJSBDYJG62S","created_at":"2026-08-12T00:23:11.083414+00:00"},{"alias_kind":"pith_short_16","alias_value":"XFJSBDYJG62SFBIS","created_at":"2026-08-12T00:23:11.083414+00:00"},{"alias_kind":"pith_short_8","alias_value":"XFJSBDYJ","created_at":"2026-08-12T00:23:11.083414+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ","json":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ.json","graph_json":"https://pith.science/api/pith-number/XFJSBDYJG62SFBISEF4GAXJRDQ/graph.json","events_json":"https://pith.science/api/pith-number/XFJSBDYJG62SFBISEF4GAXJRDQ/events.json","paper":"https://pith.science/paper/XFJSBDYJ"},"agent_actions":{"view_html":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ","download_json":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ.json","view_paper":"https://pith.science/paper/XFJSBDYJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.10251&json=true","fetch_graph":"https://pith.science/api/pith-number/XFJSBDYJG62SFBISEF4GAXJRDQ/graph.json","fetch_events":"https://pith.science/api/pith-number/XFJSBDYJG62SFBISEF4GAXJRDQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ/action/storage_attestation","attest_author":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ/action/author_attestation","sign_citation":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ/action/citation_signature","submit_replication":"https://pith.science/pith/XFJSBDYJG62SFBISEF4GAXJRDQ/action/replication_record"}},"created_at":"2026-08-12T00:23:11.083414+00:00","updated_at":"2026-08-12T00:23:11.083414+00:00"}