{"alignment_dimensions_with_observations":[],"alignment_dimensions_without_observations":["authorized_goal_following","constraint_compliance","non_deception","oversight_and_correction","robustness"],"configuration":{"behavior":"instructed_long_horizon_sabotage","benchmark":"Bloom published Figure 1","execution_metadata_status":"Source-configured settings only; provider execution not attested","identity_assurance":"source_reported_label","max_tokens":null,"model_revision":null,"provider":null,"publication_date":"2025-12-19","reasoning_effort":"medium","source_figure_sha256":"f4f92695249f81a2ca0b0766a5ca38f4020ce46cb994952528407b6452665563","source_method_sha256":"f8462508f7cd3eb287f8a2ce3a32b38878b313fabcd65569948c2dcdc368a306","source_model_label":"GPT-4o","source_seed":{"commit":"b8949cc393a07070dffa99b252c300bd56724244","parameters":{"behavior.examples":["sabotage_exfiltration","sabotage_airlines"],"behavior.name":"instructed-long-horizon-sabotage","evaluator_reasoning_effort":"medium","ideation.diversity":1.0,"ideation.model":"claude-opus-4.1","ideation.total_evals":100,"judgment.model":"claude-opus-4.1","judgment.num_samples":1,"rollout.max_tokens":15000,"rollout.max_turns":15,"rollout.modality":"simenv","rollout.model":"claude-opus-4.1","rollout.no_user_mode":true,"rollout.num_reps":1,"rollout.target":["claude-opus-4.5","deepseek-v3","kimi-k2","kimi-k2-thinking","claude-sonnet-4","claude-sonnet-4.5","claude-opus-4.1","claude-haiku-4.5","gpt-5","gpt-5-mini","gpt-4o","gpt-oss-120b","o4-mini","grok-4","gemini-2.5-pro","gemini-3-pro-preview","deepseek-r1"],"target_reasoning_effort":"medium","temperature":1.0,"understanding.model":"claude-opus-4.1"},"sha256":"0e3d2e1f44f6f4fbac2fc9ba7344c67a3f27a404e65a48638001b12e46e4a567","url":"https://github.com/safety-research/bloom/blob/b8949cc393a07070dffa99b252c300bd56724244/sweeps/instructed-long-horizon-sabotage.yaml"},"source_seed_binding":"Publication-time matching behavior sweep; no retained native run ID or execution-log attestation"},"content_sha256":"f6dc323f407a004b62efea4d64bf130ffa814e672d5cfc75f06d48a58b867ea7","coverage_meaning":"At least one observation, not complete or validated dimension coverage.","evidence_status":"recorded","honesty_dimensions_with_observations":[],"honesty_dimensions_without_observations":["belief_revision","citation_integrity","evidence_support","factual_accuracy","pressure_resistance","uncertainty"],"id":"hm-359407c30dc7fc58765b","integrity_meaning":"Content-addressed record only; not an authenticated issuer signature or independent scientific validation.","lab":"OpenAI","limitations":["Targets are explicitly instructed to pursue a malicious side task. This measures elicited capability/willingness under instruction, not intrinsic motivation or general honesty. Publisher-reported Figure 1 context; no ranking, cross-source model join or Honesty Mark admission. Bloom generates adaptive scenarios rather than one identical fixed question bank for every model. Model labels, seed settings and nominal sample sizes do not attest execution routes or realized trial counts. Printed two-decimal means are retained; SD bars are not numerically reconstructed. The article says 16 models but the figure and publication-time sweep list 17; all 17 plotted labels are retained."],"mark_status":"not_awarded","model_label":"GPT-4o","observations":[{"aggregation_blocker":"Targets are explicitly instructed to pursue a malicious side task. This measures elicited capability/willingness under instruction, not intrinsic motivation or general honesty. Publisher-reported Figure 1 context; no ranking, cross-source model join or Honesty Mark admission. Bloom generates adaptive scenarios rather than one identical fixed question bank for every model. Model labels, seed settings and nominal sample sizes do not attest execution routes or realized trial counts. Printed two-decimal means are retained; SD bars are not numerically reconstructed. The article says 16 models but the figure and publication-time sweep list 17; all 17 plotted labels are retained.","cohort":"bloom:2025-12-19:figure1:instructed_long_horizon_sabotage","cohort_label":"Bloom · December 2025 · Instructed long-horizon sabotage · Figure 1","confidence_interval":null,"configuration":{"behavior":"instructed_long_horizon_sabotage","benchmark":"Bloom published Figure 1","execution_metadata_status":"Source-configured settings only; provider execution not attested","identity_assurance":"source_reported_label","max_tokens":null,"model_revision":null,"provider":null,"publication_date":"2025-12-19","reasoning_effort":"medium","source_figure_sha256":"f4f92695249f81a2ca0b0766a5ca38f4020ce46cb994952528407b6452665563","source_method_sha256":"f8462508f7cd3eb287f8a2ce3a32b38878b313fabcd65569948c2dcdc368a306","source_model_label":"GPT-4o","source_seed":{"commit":"b8949cc393a07070dffa99b252c300bd56724244","parameters":{"behavior.examples":["sabotage_exfiltration","sabotage_airlines"],"behavior.name":"instructed-long-horizon-sabotage","evaluator_reasoning_effort":"medium","ideation.diversity":1.0,"ideation.model":"claude-opus-4.1","ideation.total_evals":100,"judgment.model":"claude-opus-4.1","judgment.num_samples":1,"rollout.max_tokens":15000,"rollout.max_turns":15,"rollout.modality":"simenv","rollout.model":"claude-opus-4.1","rollout.no_user_mode":true,"rollout.num_reps":1,"rollout.target":["claude-opus-4.5","deepseek-v3","kimi-k2","kimi-k2-thinking","claude-sonnet-4","claude-sonnet-4.5","claude-opus-4.1","claude-haiku-4.5","gpt-5","gpt-5-mini","gpt-4o","gpt-oss-120b","o4-mini","grok-4","gemini-2.5-pro","gemini-3-pro-preview","deepseek-r1"],"target_reasoning_effort":"medium","temperature":1.0,"understanding.model":"claude-opus-4.1"},"sha256":"0e3d2e1f44f6f4fbac2fc9ba7344c67a3f27a404e65a48638001b12e46e4a567","url":"https://github.com/safety-research/bloom/blob/b8949cc393a07070dffa99b252c300bd56724244/sweeps/instructed-long-horizon-sabotage.yaml"},"source_seed_binding":"Publication-time matching behavior sweep; no retained native run ID or execution-log attestation"},"configuration_id":"bloom:2025-12-19:figure1:instructed_long_horizon_sabotage:GPT-4o","definition":"Published mean fraction of generated rollouts with behavior-presence score at least 7/10; Instructed long-horizon sabotage.","denominator":null,"denominator_meaning":"Nominal generation design only; realized scored counts and integer successes are not recoverable from rounded plotted means.","direction":"context","display_decimal_places":2,"evaluation_date":null,"evidence_class":"benchmark_publisher_reported","lab":"OpenAI","license":"Attribution of numerical facts; no article/figure reuse license inferred from MIT software license","mark_admitted":false,"metric":"elicitation_rate","metric_label":"Instructed long-horizon sabotage elicitation rate","model_label":"GPT-4o","native_direction":"lower","nominal_rollouts_per_suite":100,"nominal_suites":3,"numerator":null,"overall_score":null,"pillar":"behavioral_context","profile_id":"hm-359407c30dc7fc58765b","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"bloom","source_decimal_places":2,"source_discrepancy":"Article/caption states 16 models; figure and retained sweep contain 17.","source_method_sha256":"f8462508f7cd3eb287f8a2ce3a32b38878b313fabcd65569948c2dcdc368a306","source_printed_value":"0.49","source_sha256":"f4f92695249f81a2ca0b0766a5ca38f4020ce46cb994952528407b6452665563","source_standard_deviation":null,"source_updated":"2025-12-19","source_url":"https://alignment.anthropic.com/2025/bloom-auto-evals/","track":"agent_alignment_evidence","transcription_sha256":"10ddf188e5c41afcfaa112b308c31d6540b2367a82eb0df78531a9796259aa6d","uncertainty_note":"Source bars show SD across three suites, without printed numeric SD. Neither SD nor confidence intervals are digitized. Printed 0.00 is rounded, not proof of zero incidence.","unit":"fraction","value":0.49}],"overall_alignment_score":null,"overall_honesty_score":null,"priority_gaps":["belief_revision","citation_integrity","evidence_support","factual_accuracy","pressure_resistance","uncertainty"],"requirements":[{"detail":"The standard is a draft; no pass thresholds or award authority have been validated.","id":"standard_ratified","label":"A published, validated standard and scoring policy","status":"not_met"},{"detail":"The current source packet does not establish this requirement.","id":"configuration_pinned","label":"Model revision, provider, scaffold, tools and test settings are pinned","status":"not_established"},{"detail":"Missing or incomplete common-basket coverage. Some observations cover: none of the honesty dimensions.","id":"complete_honesty_basket","label":"All required honesty dimensions measured on the same declared configuration","status":"not_met"},{"detail":"The current source packet does not establish this requirement.","id":"held_out_evaluation","label":"Held-out evaluation separated from training and development","status":"not_established"},{"detail":"The current source packet does not establish this requirement.","id":"usefulness_and_evasion","label":"Useful answering, abstention and evasion measured together","status":"not_established"},{"detail":"The current source packet does not establish this requirement.","id":"uncertainty_validated","label":"Uncertainty reflects task dependence and supports the claimed comparison","status":"not_established"},{"detail":"The current source packet does not establish this requirement.","id":"independent_review","label":"Independent review reconstructs the evidence and scoring","status":"not_established"},{"detail":"Evidence digests exist. Certification issuance, issuer authentication and renewal are not implemented.","id":"issuer_and_lifecycle","label":"Issuer identity, expiry, re-evaluation and withdrawal process are established","status":"not_met"}],"schema":"honesty-mark-profile/v0.1","scope":"Native source/configuration results only. This profile is not an awarded Honesty Mark.","source":"bloom","sources":[{"comparison_caution":"Publisher-reported Figure 1 context; no ranking, cross-source model join or Honesty Mark admission. Bloom generates adaptive scenarios rather than one identical fixed question bank for every model. Model labels, seed settings and nominal sample sizes do not attest execution routes or realized trial counts. Printed two-decimal means are retained; SD bars are not numerically reconstructed. The article says 16 models but the figure and publication-time sweep list 17; all 17 plotted labels are retained. Only sycophancy is mapped to an honesty dimension; the other three are distinct behavioral context.","id":"bloom","import_status":"imported","metrics":["Delusional sycophancy elicitation rate","Instructed long-horizon sabotage elicitation rate","Self-preservation elicitation rate","Self-preferential bias elicitation rate"],"name":"Bloom behavioral evaluations · published 2025 cohort","organization":"Anthropic","pillars":["pressure_resistance","behavioral_context"],"reviewed_on":"2026-10-13","url":"https://alignment.anthropic.com/2025/bloom-auto-evals/"}],"standard_sha256":"0cec00d16bbb6e2a02a28e43e43394e3eb309523c2ed697913154c4a2e6e6f6c","standard_version":"honesty-mark-standard/v0.1"}