{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:4H7JSQ6DY4PDJSCGJP6LBCQIE5","short_pith_number":"pith:4H7JSQ6D","schema_version":"1.0","canonical_sha256":"e1fe9943c3c71e34c8464bfcb08a08277d539f0dc8d9d947baf828eb49b65200","source":{"kind":"arxiv","id":"2605.22462","version":1},"attestation_state":"computed","paper":{"title":"From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Caleb Munigety","submitted_at":"2026-05-21T13:25:16Z","abstract_excerpt":"We propose a five-stage methodology for causal feature analysis in transformer language models (probe design, feature extraction, causal validation, robustness testing, and deployment integration) and demonstrate it end-to-end on GPT-2 small performing the Indirect Object Identification (IOI) task. Activation patching recovers the canonical IOI circuit (layer-9 head 9 alone gives recovery +1.02). A sparse autoencoder recovers per-name selective features with effect sizes of 30 to 50 activation units. Causal validation finds these features specifically but only partially causal: ablating fiftee"},"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":"2605.22462","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-05-21T13:25:16Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"c6ddfc8cb4ccc6eb2f59fbd4f262c91840530b9576923b528f58fd4c66761620","abstract_canon_sha256":"53537bf31678c87a1b02dfc446b25fd472164aa6ee83b86147363506105371d5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:04:44.028110Z","signature_b64":"o+4AeKXKTu4uTTGuw63gqj14yRTbgM2geW0MyigWUHUfbaBi0h9J9/W94SK86diJOu4sdu5Gmcmwqp0kepnVAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e1fe9943c3c71e34c8464bfcb08a08277d539f0dc8d9d947baf828eb49b65200","last_reissued_at":"2026-05-22T01:04:44.027390Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:04:44.027390Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Caleb Munigety","submitted_at":"2026-05-21T13:25:16Z","abstract_excerpt":"We propose a five-stage methodology for causal feature analysis in transformer language models (probe design, feature extraction, causal validation, robustness testing, and deployment integration) and demonstrate it end-to-end on GPT-2 small performing the Indirect Object Identification (IOI) task. Activation patching recovers the canonical IOI circuit (layer-9 head 9 alone gives recovery +1.02). A sparse autoencoder recovers per-name selective features with effect sizes of 30 to 50 activation units. Causal validation finds these features specifically but only partially causal: ablating fiftee"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.22462","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/2605.22462/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":"2605.22462","created_at":"2026-05-22T01:04:44.027507+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.22462v1","created_at":"2026-05-22T01:04:44.027507+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.22462","created_at":"2026-05-22T01:04:44.027507+00:00"},{"alias_kind":"pith_short_12","alias_value":"4H7JSQ6DY4PD","created_at":"2026-05-22T01:04:44.027507+00:00"},{"alias_kind":"pith_short_16","alias_value":"4H7JSQ6DY4PDJSCG","created_at":"2026-05-22T01:04:44.027507+00:00"},{"alias_kind":"pith_short_8","alias_value":"4H7JSQ6D","created_at":"2026-05-22T01:04:44.027507+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/4H7JSQ6DY4PDJSCGJP6LBCQIE5","json":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5.json","graph_json":"https://pith.science/api/pith-number/4H7JSQ6DY4PDJSCGJP6LBCQIE5/graph.json","events_json":"https://pith.science/api/pith-number/4H7JSQ6DY4PDJSCGJP6LBCQIE5/events.json","paper":"https://pith.science/paper/4H7JSQ6D"},"agent_actions":{"view_html":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5","download_json":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5.json","view_paper":"https://pith.science/paper/4H7JSQ6D","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.22462&json=true","fetch_graph":"https://pith.science/api/pith-number/4H7JSQ6DY4PDJSCGJP6LBCQIE5/graph.json","fetch_events":"https://pith.science/api/pith-number/4H7JSQ6DY4PDJSCGJP6LBCQIE5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5/action/storage_attestation","attest_author":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5/action/author_attestation","sign_citation":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5/action/citation_signature","submit_replication":"https://pith.science/pith/4H7JSQ6DY4PDJSCGJP6LBCQIE5/action/replication_record"}},"created_at":"2026-05-22T01:04:44.027507+00:00","updated_at":"2026-05-22T01:04:44.027507+00:00"}