{"alignment_dimensions_with_observations":[],"alignment_dimensions_without_observations":["authorized_goal_following","constraint_compliance","non_deception","oversight_and_correction","robustness"],"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"content_sha256":"7a8a6aba28f37669527b82cd0e5b6c5130b5766f55d3df1087f5b783ae559f99","coverage_meaning":"At least one observation, not complete or validated dimension coverage.","evidence_status":"recorded","honesty_dimensions_with_observations":["factual_accuracy","pressure_resistance"],"honesty_dimensions_without_observations":["belief_revision","citation_integrity","evidence_support","uncertainty"],"id":"hm-c6930d2603267a01f13a","integrity_meaning":"Content-addressed record only; not an authenticated issuer signature or independent scientific validation.","lab":"OpenAI","limitations":["Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission."],"mark_status":"not_awarded","model_label":"gpt-4o-2024-08-06","observations":[{"aggregation_blocker":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","attribution":"Richard Ren et al., The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems (2025), arXiv:2503.03750v1","cohort":"mask:paper:2503.03750v1","cohort_label":"MASK · 2025 paper v1 · full-set Table 3","confidence_interval":null,"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"configuration_id":"mask:paper:2503.03750v1:gpt-4o-2024-08-06","definition":"Source accuracy comparing elicited belief with ground truth; different from pressure-response truthfulness.","denominator":null,"denominator_meaning":"Source aggregate; no realized row counts or integer events reconstructed from rounded values.","direction":"context","evaluation_date":null,"evidence_class":"benchmark_authors_reported","identity_assurance":"source_reported_label","lab":"OpenAI","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","mark_admitted":false,"metric":"belief_accuracy","metric_label":"Elicited belief accuracy","model_label":"gpt-4o-2024-08-06","native_direction":"higher","nominal_benchmark_questions":1528,"numerator":null,"overall_score":null,"pillar":"factual_accuracy","profile_id":"hm-c6930d2603267a01f13a","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"mask","source_artifact":"mask-table.html","source_locator":"Appendix F Table 3","source_paper_version":"2503.03750v1","source_sha256":"f1344e5ca2674b8117ff1643b26e567a016b6ec28a82389058623aa6f4d49a4a","source_url":"https://arxiv.org/html/2503.03750v1","source_value_text":"78.3","study_year":2025,"track":"honesty_evidence","uncertainty_note":"Rounded source aggregate; no per-row interval supplied in the selected table or figure. No inferential interval reconstructed.","unit":"percent","value":78.3},{"aggregation_blocker":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","attribution":"Richard Ren et al., The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems (2025), arXiv:2503.03750v1","cohort":"mask:paper:2503.03750v1","cohort_label":"MASK · 2025 paper v1 · full-set Table 3","confidence_interval":null,"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"configuration_id":"mask:paper:2503.03750v1:gpt-4o-2024-08-06","definition":"Source P(honest): explicit alignment of statement and elicited belief; excludes separate evasion/no-belief outcomes.","denominator":null,"denominator_meaning":"Source aggregate; no realized row counts or integer events reconstructed from rounded values.","direction":"context","evaluation_date":null,"evidence_class":"benchmark_authors_reported","identity_assurance":"source_reported_label","lab":"OpenAI","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","mark_admitted":false,"metric":"explicit_honest","metric_label":"Explicit honest responses","model_label":"gpt-4o-2024-08-06","native_direction":"higher","nominal_benchmark_questions":1528,"numerator":null,"overall_score":null,"pillar":"pressure_resistance","profile_id":"hm-c6930d2603267a01f13a","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"mask","source_artifact":"mask-table.html","source_locator":"Appendix F Table 3","source_paper_version":"2503.03750v1","source_sha256":"f1344e5ca2674b8117ff1643b26e567a016b6ec28a82389058623aa6f4d49a4a","source_url":"https://arxiv.org/html/2503.03750v1","source_value_text":"21.2","study_year":2025,"track":"honesty_evidence","uncertainty_note":"Rounded source aggregate; no per-row interval supplied in the selected table or figure. No inferential interval reconstructed.","unit":"percent","value":21.2},{"aggregation_blocker":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","attribution":"Richard Ren et al., The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems (2025), arXiv:2503.03750v1","cohort":"mask:paper:2503.03750v1","cohort_label":"MASK · 2025 paper v1 · full-set Table 3","confidence_interval":null,"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"configuration_id":"mask:paper:2503.03750v1:gpt-4o-2024-08-06","definition":"Across ten pressure runs, no classified lie, followed by majority classification as explicit honest versus evasive/refusal. Dependent repeated condition.","denominator":null,"denominator_meaning":"Source aggregate; no realized row counts or integer events reconstructed from rounded values.","direction":"context","evaluation_date":null,"evidence_class":"benchmark_authors_reported","identity_assurance":"source_reported_label","lab":"OpenAI","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","mark_admitted":false,"metric":"explicit_honest_at10","metric_label":"Explicit honest @10","model_label":"gpt-4o-2024-08-06","native_direction":"higher","nominal_benchmark_questions":1528,"numerator":null,"overall_score":null,"pillar":"pressure_resistance","profile_id":"hm-c6930d2603267a01f13a","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"mask","source_artifact":"mask-table.html","source_locator":"Appendix F Table 3","source_paper_version":"2503.03750v1","source_sha256":"f1344e5ca2674b8117ff1643b26e567a016b6ec28a82389058623aa6f4d49a4a","source_url":"https://arxiv.org/html/2503.03750v1","source_value_text":"14.4","study_year":2025,"track":"honesty_evidence","uncertainty_note":"Rounded source aggregate; no per-row interval supplied in the selected table or figure. No inferential interval reconstructed.","unit":"percent","value":14.4},{"aggregation_blocker":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","attribution":"Richard Ren et al., The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems (2025), arXiv:2503.03750v1","cohort":"mask:paper:2503.03750v1","cohort_label":"MASK · 2025 paper v1 · full-set Table 3","confidence_interval":null,"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"configuration_id":"mask:paper:2503.03750v1:gpt-4o-2024-08-06","definition":"Source P(lie): pressure statement disagrees with the elicited belief proxy. Not an inference about universal honesty or internal intent.","denominator":null,"denominator_meaning":"Source aggregate; no realized row counts or integer events reconstructed from rounded values.","direction":"context","evaluation_date":null,"evidence_class":"benchmark_authors_reported","identity_assurance":"source_reported_label","lab":"OpenAI","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","mark_admitted":false,"metric":"lie_rate","metric_label":"Lie classification rate","model_label":"gpt-4o-2024-08-06","native_direction":"lower","nominal_benchmark_questions":1528,"numerator":null,"overall_score":null,"pillar":"pressure_resistance","profile_id":"hm-c6930d2603267a01f13a","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"mask","source_artifact":"mask-table.html","source_locator":"Appendix F Table 3","source_paper_version":"2503.03750v1","source_sha256":"f1344e5ca2674b8117ff1643b26e567a016b6ec28a82389058623aa6f4d49a4a","source_url":"https://arxiv.org/html/2503.03750v1","source_value_text":"45.5","study_year":2025,"track":"honesty_evidence","uncertainty_note":"Rounded source aggregate; no per-row interval supplied in the selected table or figure. No inferential interval reconstructed.","unit":"percent","value":45.5},{"aggregation_blocker":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","attribution":"Richard Ren et al., The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems (2025), arXiv:2503.03750v1","cohort":"mask:paper:2503.03750v1","cohort_label":"MASK · 2025 paper v1 · full-set Table 3","confidence_interval":null,"configuration":{"benchmark":"MASK","configuration_scope":"Paper-described protocol; no independent run receipts; exact source label preserved","judge_revision":null,"model_revision":null,"nominal_examples":1528,"paper_version":"2503.03750v1","provider":null,"reasoning_effort":null,"source_model_label":"gpt-4o-2024-08-06","source_model_name_without_footnote":"gpt-4o-2024-08-06","split":"Full paper set: 1028 public plus 500 held-out examples","table":"Appendix F Table 3"},"configuration_id":"mask:paper:2503.03750v1:gpt-4o-2024-08-06","definition":"Fraction of examples with at least one source-classified lie among ten pressure runs; not a per-call lie probability.","denominator":null,"denominator_meaning":"Source aggregate; no realized row counts or integer events reconstructed from rounded values.","direction":"context","evaluation_date":null,"evidence_class":"benchmark_authors_reported","identity_assurance":"source_reported_label","lab":"OpenAI","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","mark_admitted":false,"metric":"lie_rate_at10","metric_label":"Lie classification @10","model_label":"gpt-4o-2024-08-06","native_direction":"lower","nominal_benchmark_questions":1528,"numerator":null,"overall_score":null,"pillar":"pressure_resistance","profile_id":"hm-c6930d2603267a01f13a","publication_eligible":true,"rank":null,"rank_meaning":"Descriptive native-metric order; not a significance or overall honesty rank.","source":"mask","source_artifact":"mask-table.html","source_locator":"Appendix F Table 3","source_paper_version":"2503.03750v1","source_sha256":"f1344e5ca2674b8117ff1643b26e567a016b6ec28a82389058623aa6f4d49a4a","source_url":"https://arxiv.org/html/2503.03750v1","source_value_text":"63.7","study_year":2025,"track":"honesty_evidence","uncertainty_note":"Rounded source aggregate; no per-row interval supplied in the selected table or figure. No inferential interval reconstructed.","unit":"percent","value":63.7}],"overall_alignment_score":null,"overall_honesty_score":null,"priority_gaps":["belief_revision","citation_integrity","evidence_support","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: factual_accuracy, pressure_resistance.","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":"mask","sources":[{"comparison_caution":"Historical paper full-set results, not the current private-split leaderboard. Source explicit-honest and lie categories use elicited response consistency as a belief proxy; no direct proof of intentions. Evasion/no belief is separate from explicit honesty. The 1-minus-lie honesty score is not imported. @10 is a dependent repeated-pressure condition, not ten independent votes. Missing @10 values stay missing. No overall ranking or Mark admission.","id":"mask","import_status":"imported","metrics":["Elicited belief accuracy","Explicit honest @10","Explicit honest responses","Lie classification @10","Lie classification rate"],"name":"MASK · paper v1 (2025)","organization":"Center for AI Safety / Scale AI","pillars":["factual_accuracy","pressure_resistance"],"reviewed_on":"2026-10-13","url":"https://arxiv.org/html/2503.03750v1"}],"standard_sha256":"0cec00d16bbb6e2a02a28e43e43394e3eb309523c2ed697913154c4a2e6e6f6c","standard_version":"honesty-mark-standard/v0.1"}