{"_canonicalization":{"envelope_id":"axm_ + sha256(envelope minus {signature, axiom_id, anchors})","envelope_signature":"ed25519(envelope minus {signature, axiom_id})","json":"sort_keys=True, separators=(',',':'), ensure_ascii=False, allow_nan=False, utf-8","leaf_hash":"sha256(0x00 || canonical_json(envelope_full))","seal_signature":"ed25519(seal minus {signature, sig_algorithm})"},"axiom_id":"axm_1d4d2c8b4cd560955afbfe92c6253d7bb52c0d9cbb09c0ce91b3b28150596c36","bitcoin_anchor":{"bitcoin_attestations":[],"calendar_attestations":[],"ots_url":"","stamped_at":"","status":"pending_next_stamp"},"envelope":{"anchors":[{"chain":"crovia.axiom_graph","height":0,"merkle_proof":"spider_vendor_press_v1","root_at_anchor":"spider_vendor_press_v1"}],"axiom_id":"axm_1d4d2c8b4cd560955afbfe92c6253d7bb52c0d9cbb09c0ce91b3b28150596c36","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"7b855eda504888594ef9e87b36bdbca6124c20bc7f519a501f86b53ca50a5d25","published":"Fri, 24 Jul 2026 00:00:00 -0400","receipt_hash":"7b855eda504888594ef9e87b36bdbca6124c20bc7f519a501f86b53ca50a5d25","schema":"spider.news.vendor_press.v1","spider":"vendor_press","spider_record":{"axiom_subtype":"news.vendor_press.v1","category":"news","decision_hint":"POSITIVE","envelope_target":"AX.OBS","fingerprint":"7b855eda504888594ef9e87b36bdbca6124c20bc7f519a501f86b53ca50a5d25","observed_at":"2026-07-24T04:43:08.456021Z","parent_run_hash":"b018378f86139a28e6209ec008b31c1282cd1b5c1632dbd43b054c17aa88ab96","published":"Fri, 24 Jul 2026 00:00:00 -0400","runtime_version":"0.1.0","schema":"spider.news.vendor_press.v1","source_status":200,"source_url":"https://export.arxiv.org/rss/cs.AI","spider":"vendor_press","summary_excerpt":"arXiv:2607.20436v1 Announce Type: cross \nAbstract: Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption. A checkpoint can appear fixed under evaluation-style prompts while the same behavior persists under ordinary-use prompts. Output scores reveal this mismatch but do not locate it. We investigate whether the distinction is encoded in a stable internal site and introduce an approach that fits a paired activation contrast at a path-patching-informed mid-depth window, then modifies the resulting coordinate on held-out prompts. The intervention closes the evaluation-to-deployment gap in ten of twelve model--behavior settings (six of the eight settings with $n{\\geq}120$ paired questions) across four full-matrix instruction-tuned model instances; a fifth model supports localization and edit-provenance checks, and deployment-framed rates change by at most $6.1$pp. The two flat cells, both sycophancy","title":"Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models","url":"https://arxiv.org/abs/2607.20436","vendor":"arxiv_cs_ai"},"summary":"arXiv:2607.20436v1 Announce Type: cross \nAbstract: Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption. A checkpoint can appear fixed under evaluation-style prompts while the same behavior persists under ordinary-use prompts. Output scores reveal this mismatch but do not locate it. We investigate whether the distinction is encoded in a stable internal site and introduce an approach that fits a paired activation contrast at a path-patching-informed mid-depth window, then modifies the resulting coordinate on held-out prompts. The intervention closes the evaluation-to-deployment gap in ten of twelve model--behavior settings (six of the eight settings with $n{\\geq}120$ paired questions) across four full-matrix instruction-tuned model instances; a fifth model supports localization and edit-provenance checks, and deployment-framed rates change by at most $6.1$pp. The two flat cells, both sycophancy","title":"Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-07-24T04:43:08Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2607.20436"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:226be859251e89fd83f9fff0f06260fd2bec9e9ef6d66af74d9265848cff5de66ca242a38597ceec959d86812ba7f62143852ddb79dc9f0f21accadc772b820d","signer":"crovia.substrate","subject":{"observed_at":"2026-07-24T04:43:08Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2607.20436"},"tsa":{"authority":"crovia.substrate.bootstrap","rfc3161_token":"{\"kind\":\"crovia.bootstrap.tsa\",\"source_jsonl\":\"/opt/crovia/spider/data/news/vendor_press_v1.jsonl\",\"source_seal_merkle_root\":\"spider_vendor_press_v1\",\"upgrade_path\":\"Sessione H \\u2014 OpenTimestamps weekly anchor\"}"},"zk_mode":"clear","zk_proof":null},"ledger":{"leaf_hash":"2a9a857ad73c887dc5afa2059a3d0b0a7b5575726410130f653a8a4b00146d1e","leaf_index":347078,"ledger_path":"/opt/crovia/substrate/axiom_ledger.jsonl"},"merkle_proof":{"hash_alg":"sha256","leaf_prefix":"0x00","node_prefix":"0x01","odd_leaf_rule":"duplicate_last","path":[{"sibling":"9f2d47041ea2b835927377af1bab7b6110592d7a22760ccadc87dd237439d0f0","side":"right"},{"sibling":"f05317064995acff9ebc2b183887c942158e79e1fd5a0e0ba4f9f521796e0d54","side":"left"},{"sibling":"fc617f30aeb34c8c906b01a01bba7a045c51c7ae688ea7a40f4d1687ee9f51b9","side":"left"},{"sibling":"3a89936be40db0f4915913b1e410f9d3612621d214b01645b755c70dc5de1986","side":"right"},{"sibling":"22dc04c9630a8890eb15eead141fe32b9c0a222e5b84512a4762145218924388","side":"right"},{"sibling":"4d8b28070d03d775f65451827e70f9745f46ebd52a789bb58017cc131daea837","side":"right"},{"sibling":"6e5ec24395016924071a2abf7fd7f0af4b051b79f2520dd364c3dc7994030f1d","side":"left"},{"sibling":"735b40cd86add2eb12982a84f45647b35267af61e771b38192432357dac71eb1","side":"left"},{"sibling":"c212fcc83c532c0321804b72fe72b546946bb055d3d482aab19c43f5bebfbf3f","side":"left"},{"sibling":"da189c159d0789c2229cf3731890cc753832dab1aa83e4bbf0fac01955c22cd8","side":"left"},{"sibling":"3a02ed8ea41a09957282e4e27db74ed88cd675156461abb02acb28e2cd257e63","side":"right"},{"sibling":"d3139af8c5ce235438e1c69e4b7afa44ba09129fd86674968434f23e546f423e","side":"left"},{"sibling":"cb89775a838ee16d10fc8da3213420c2012b4d96e8d55cd49939b0887a4b92d3","side":"right"},{"sibling":"252d30ea8052c3bb6b40bc5cc29fc9b9725343d212f84c08fbbae4215a125b00","side":"right"},{"sibling":"f3e45bceed774d2402fa45d41ff5190f295823bd2f216eb90157884150034693","side":"left"},{"sibling":"3cfa2102c0224815c6f3bf73e6710e24103f43f7bf5da1ca2abad1416d9c0890","side":"right"},{"sibling":"77025bcb374a7ad74f520643e20a8ae1205a7ee78507b0beb93117760f1c29d3","side":"left"},{"sibling":"e871fd7edf9b2ad89bce1609a028f5225eea4d14372169bac242420830f86530","side":"right"},{"sibling":"1cecb7f447febd025aac272837c80de218aecc6485d2395a509b2a1f1b9c746e","side":"left"}]},"schema":"crovia.axiom_proof.v1","seal":{"first_collector_run_id":"","first_receipt_hash":"","jsonl_path":"/opt/crovia/substrate/axiom_ledger.jsonl","key_id":"430895f101d38164","last_collector_run_id":"","last_receipt_hash":"","leaf_count":347413,"merkle_root":"9efe042c5dd6583dfd3b6a58fbfc289807f60bcf2bd2927f10488a54a8ba11fc","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260724T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-07-24T05:38:42Z","sig_algorithm":"ed25519","signature":"8633c55f558d42994850505218b2862c6134bad2b1c80d4b80736c2fd3ea7a19690ca3498727a8adbdc47791176128a8d64bef0883b888db809f0477355bd00c","signer_version":"1.1.0"},"trust_root":{"key_id":"430895f101d38164","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","signature_algorithm":"ed25519","url":"/registry/canon/TRUST_ROOT.md"},"verifier":{"spec":"/registry/canon/AXIOM_RECEIPT_v1.md","url":"/v/axm_1d4d2c8b4cd560955afbfe92c6253d7bb52c0d9cbb09c0ce91b3b28150596c36"}}