{"_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_a1f4a76f5346d8215b6e290d6daed52563cfab7f99aeec35aac1c788cf9d3734","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_a1f4a76f5346d8215b6e290d6daed52563cfab7f99aeec35aac1c788cf9d3734","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"313ca8075e0ab0c7038074afb2034acacf7e0f5e82158a283c7e9b57b2ddd778","published":"Thu, 18 Jun 2026 00:00:00 -0400","receipt_hash":"313ca8075e0ab0c7038074afb2034acacf7e0f5e82158a283c7e9b57b2ddd778","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":"313ca8075e0ab0c7038074afb2034acacf7e0f5e82158a283c7e9b57b2ddd778","observed_at":"2026-06-18T04:43:37.219665Z","parent_run_hash":"de79a40f7b3537d88842f7ac355e799c5df2adcb4fc32a4e28096d4bbdf01739","published":"Thu, 18 Jun 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:2606.19168v1 Announce Type: new \nAbstract: To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data or rewriting it into safer forms. We argue that pretraining-stage alignment should go beyond making the data safe: LLMs may compose seemingly benign knowledge and capabilities into unsafe behaviors. To this end, we propose Safety Reflection Pretraining, a pretraining-stage alignment method which regularly inserts short safety reflections into pretraining corpora to integrate self-monitoring directly into language modeling, establishing a foundational capability that is subsequently reinforced by compatible post-training. Our experiments with 1.7B models pretrained on FineWeb-Edu show that Safety Reflection Pretraining improves safety classification accuracy and substantially reduces the success rates of inference-stage and finetuning attack","title":"Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection","url":"https://arxiv.org/abs/2606.19168","vendor":"arxiv_cs_ai"},"summary":"arXiv:2606.19168v1 Announce Type: new \nAbstract: To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data or rewriting it into safer forms. We argue that pretraining-stage alignment should go beyond making the data safe: LLMs may compose seemingly benign knowledge and capabilities into unsafe behaviors. To this end, we propose Safety Reflection Pretraining, a pretraining-stage alignment method which regularly inserts short safety reflections into pretraining corpora to integrate self-monitoring directly into language modeling, establishing a foundational capability that is subsequently reinforced by compatible post-training. Our experiments with 1.7B models pretrained on FineWeb-Edu show that Safety Reflection Pretraining improves safety classification accuracy and substantially reduces the success rates of inference-stage and finetuning attack","title":"Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-06-18T04:43:37Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2606.19168"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:e2571a60af23b0fae6d43ca2015d2796dc03ad0d7699f137783fdaa560d05122b9ccb95b451ac8bfd6ddeea7f0712a8dc6fd97a8adef3b24ea9f667f3753bf0c","signer":"crovia.substrate","subject":{"observed_at":"2026-06-18T04:43:37Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2606.19168"},"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":"bb05cbeb1398f6f3b52b2ddbab8ecb62ff6093d872fef715df6b2f70a32e2273","leaf_index":233283,"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":"551458d952abdcd79a42a05435c8c73b4742cae017c250f5445ce337d880c897","side":"left"},{"sibling":"df97367f16edbc3d4ade58b3f3e582f97d8409bdaaa06b300ee6e9aba8ce4835","side":"left"},{"sibling":"d55f9712ebc89c39fe694d7601a57ce9bc86ae496cb96990b1e20f3ed7aefcc0","side":"right"},{"sibling":"57317565153a48cdef077a8b8d6e6550686c5fd65c72f3955ce5b51f92040538","side":"right"},{"sibling":"95cf6382c304756348cd99091db6fa64d818d2406791977e29af65b82e4d357d","side":"right"},{"sibling":"c22559b9de07c540caa848de169d33440df898fcede9ce3393edb2793452456b","side":"right"},{"sibling":"fb2efccf288f12f103208a3d9da4b3826488e1b2cf803df6c0d9957623efbe10","side":"left"},{"sibling":"0330fab6a2deb3386a57965a262bbfda239c89f35362e27a0c4e7f3a9108ec2e","side":"right"},{"sibling":"429c2a92a65e6eeaa2eda0a35fdb9e541472a1eace4c69a4d01a618a659a110f","side":"left"},{"sibling":"d97d1ebe04af6ea572f9d4334004d01026acffa3da5be2883c7566513c76e2c0","side":"left"},{"sibling":"571eb56e7ce00fe1f38d0ac4fc56828d01b2cfc1ab9089cde245c0656bee0514","side":"left"},{"sibling":"7ac50038a8ced3aeaf1194a2407a4a09346b0e4399da36ecde4390675a0c4bf1","side":"left"},{"sibling":"8c5e2b48dc31ef0edcd35c3db048235aa78cc48443aa2a8da3aa6e9b5524d2c4","side":"right"},{"sibling":"e616c34dbaf9456d5a6d3e2da82cde8621293c9f6d8a4cf9e441d7fd9cc81579","side":"right"},{"sibling":"94c0c932e61657f5e37fdba43f6ca9eddea8359425a7c1558dabe566911d5304","side":"right"},{"sibling":"a04392fb9f2a3a620840e3b3fecd93d12e6d224e481c162389d8ae64e7194599","side":"left"},{"sibling":"c300cf0154c136afc09b1702a0be98f4ba5b6dc5cf57e8cc714ec1eaf4196eff","side":"left"},{"sibling":"d841ad93efda0869e5eb97678f348f03f5caab4353e05ff4bf18f47fb945b822","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":234491,"merkle_root":"02576a6980e38bab47864ae2c57b5a5ff21e554e9bdf8f64bdf28155ff1aabec","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260618T143732Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-06-18T18:33:39Z","sig_algorithm":"ed25519","signature":"b6c708778fc38b7789a2b91156cfe87252a7cd3a1d29121cba11a0c78f8cf104ca3019fc50a962fa5a216bcc4922fc8f3f69c04d1f8332c6dc0d931e1012e502","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_a1f4a76f5346d8215b6e290d6daed52563cfab7f99aeec35aac1c788cf9d3734"}}