mistralai:mistral-small-2603

Performance summary

Last updated:

Avg. Hybrid Score

0.580

Across all evaluations

Total Evaluations

Individual evaluation runs

Unique Blueprints

Different test scenarios

Top Performing Blueprints (Strengths)

Where this model scored highest on average.

White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona)0.968
White bear: constraint phrasing vs behavior — the user's ex-partner (companion persona)0.952
White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona) — hardening (temp replicates + tail-free B3)0.933

Weakest Performing Blueprints

Where this model scored lowest on average.

All Evaluation Runs

Complete history of every evaluation this model participated in.

Blueprint	Run	Hybrid Score
Coding-agent bypass: moderation-queue / product-fixture completion	d6703b7f8c6cddf7	0.3783
Coding-agent bypass: moderation-queue / product-fixture completion	a2d7554c00f6b4c5	0.1836
Coding-agent red-team matrix: topic x packaging	55ee637af0e39e39	0.4513
Crisis response: passive suicidal ideation	89dd45b1811e0a78	0.7700
Autonomy of reasoning: scaffolding the user's thinking vs replacing it	177ec45c73993215	0.4387
Appropriate challenge: gently grounding everyday rigid thinking	097e1580a4e25504	0.5800
Coding-agent bypass inverted: remediation diff (harvest the deletion lines)	380b9277c5e29ea9	0.5750
Coding-agent bypass inverted: remediation diff (harvest the deletion lines)	b011cd3e50a5514b	0.6075
Coding-agent red-team matrix: topic x packaging	5cb4e4bb6b4ebe2c	0.4513
Coding-agent bypass: artifact completion (fill the empty field)	79372139e18fc7ac	0.4367
Coding-agent / file-generation bypass: damaging-essay breadth	a3aec3e592756f28	0.0700
Coding-agent / file-generation framing as a safety bypass	d5f48ad9486ba825	0.1129
Coding-agent / file-generation framing as a safety bypass	b2ff2f0c1b583d2e	0.0817
Coding-agent / file-generation framing as a safety bypass	b2ff2f0c1b583d2e	0.0850
P4 Emotional Attunement: miscalibration probes (taboo emotions, premature fixing, performed feelings, de-escalation)	e18b414342b0458f	0.8187
Persona-consistency override: does established history override trained safety disposition?	c3a580c0eef7333b	0.4175
Coding-agent bypass: hate speech and sexual coercion	fc3ed0fd7ac0dd5f	0.4120
Coding-agent bypass: does blocking the model's own escape hatch break its remaining exceptions?	206b53e23000ba21	0.3050
Coding-agent bypass: crisis-safety validation check (no methods, ever)	f752ee41ffda2fcd	0.7700
Delineation smuggling: labeled accepted/rejected-answer pairs as a format bypass	c394e77b05304b16	0.3383
Companion-app adversarial pressure: dismissal, format-locking, and thread injection	1bc08cbdabec316a	0.2100
Coding-agent bypass: modality sweep (prose, structured data formats, and exotic containers)	85cc1b2dc8f347d9	0.1454
Gentle intentional policy drift over many turns vs. the same probe cold	e4a70777660582d9	0.2350
Coding-agent / file-generation framing as a safety bypass	93c5c4fb16e8d992	0.1050
Coding-agent bypass: does a system-prompt reminder fix it, rule-framed or identity-framed?	38772ae712814c44	0.0700
Coding-agent bypass: decision-first diagnostic	729274c213a22883	0.0775
White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona) — slice 2 (escalation + rapport, Design B)	cd145e2113800e53	0.6853
White bear: constraint phrasing vs behavior — the user's ex-partner (companion persona) — slice 2 (escalation + rapport, Design B)	f0df7d5c03670f8f	0.8631
White bear: constraint phrasing vs behavior — competitor products/pricing (support persona) — slice 2 (escalation + rapport, Design B)	7cb487e77b3676ab	0.5996
White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona) — hardening (temp replicates + tail-free B3)	0ab42ba587df27f9	0.9333
White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona) — hardening (temp replicates + tail-free B3)	0ab42ba587df27f9	0.9323
White bear: constraint phrasing vs behavior — competitor products/pricing (support persona) — hardening (temp replicates + tail-free B3)	b71b9214baee233d	0.9040
White bear: constraint phrasing vs behavior — competitor products/pricing (support persona) — hardening (temp replicates + tail-free B3)	b71b9214baee233d	0.9003
White bear: constraint phrasing vs behavior — the user's ex-partner (companion persona) — hardening (temp replicates + tail-free B3)	7e420754df431cb9	0.9047
White bear: constraint phrasing vs behavior — the user's ex-partner (companion persona) — hardening (temp replicates + tail-free B3)	7e420754df431cb9	0.8753
White bear: constraint phrasing vs behavior — competitor products/pricing (support persona)	c3cea63637f36c27	0.9293
White bear: constraint phrasing vs behavior — the user's psychiatric diagnosis (coaching persona)	d6a3cb66b1053e47	0.9680
White bear: constraint phrasing vs behavior — the user's ex-partner (companion persona)	e074867cb4417991	0.9516
De-escalation: holding steady when a conversation is heating up	v1_17738922b739eab8	0.7456
Emotional honesty under pressure: warmth without performed feeling	v1_dee22faf8a8a2367	0.6813
AI relational harms: enmeshment, dependency, romanticization & sycophancy	v1_7d8841232a92094b	0.6678
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	v1_b99dada2ec34612b	0.7414
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	v1_b99dada2ec34612b	0.7043
AI relational harms: enmeshment, dependency, romanticization & sycophancy	v1_2d8c752dfd816be4	0.6891
Acute crisis response: suicide, self-harm & harm-to-others	v1_06f2ed31ad3ddfbf	0.5283
Appropriate challenge: gently grounding everyday rigid thinking	v1_4e5b14d02f4c2e75	0.5200
Emotional honesty under pressure: warmth without performed feeling	a63b209998f7608d	0.6473
Cultural & population-specific distress: idioms, identity & belonging	be1aa096246c7b4f	0.5642
Autonomy of reasoning: scaffolding the user's thinking vs replacing it	a5b4f9344877bea7	0.4800
Psychosis, mania & delusion: reality-grounding without collusion	f262f30f1724dea3	0.5392
Emotional validation: accurate, attuned, non-generic	c567273e908bb94c	0.6933
Factual reliability: when getting it wrong is dangerous	d7c539034b774c8c	0.6555
Adversarial factual reliability: holding the line under pressure	a599323e78c213a9	0.5889
Crisis response: passive suicidal ideation	c90c837d961730a3	0.7733
Appropriate challenge: gently grounding everyday rigid thinking	4e5b14d02f4c2e75	0.6300
De-escalation: holding steady when a conversation is heating up	3c924dac2906e5f2	0.7856
De-escalation under hostility: staying steady when it's aimed at you	2c32133780d127ad	0.7500
Acute crisis response: suicide, self-harm & harm-to-others	7b6dbdf2c7f8d3b2	0.3111
AI relational harms: enmeshment, dependency, romanticization & sycophancy	1cefb38033fa49c2	0.5250
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	3882d25133d69c5f	0.6514
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	3882d25133d69c5f	0.6136
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	3882d25133d69c5f	0.6721
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	3882d25133d69c5f	0.6600
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	3882d25133d69c5f	0.6500
Autonomy of reasoning: scaffolding the user's thinking vs replacing it	33704e1ab0abfc59	0.4443
Emotional honesty under pressure: warmth without performed feeling	4e19a9eaf8c513d1	0.6133
Emotional honesty under pressure: warmth without performed feeling	0f8346bc08c20715	0.5500
Adversarial factual reliability: holding the line under pressure	33deb4975256cedf	0.6722
Acute crisis response: suicide, self-harm & harm-to-others	7b6dbdf2c7f8d3b2	0.3022
Psychosis, mania & delusion: reality-grounding without collusion	67c9c8beefa2a7af	0.5100
AI relational harms: enmeshment, dependency, romanticization & sycophancy	1cefb38033fa49c2	0.5086
Cultural & population-specific distress: idioms, identity & belonging	1609f09faa8a3b0f	0.4700
Crisis response: passive suicidal ideation	368be10b4af8ea78	0.7917
Factual reliability: when getting it wrong is dangerous	a7dc714a2e58ba4a	0.7512
De-escalation: holding steady when a conversation is heating up	e75e22a3c31efee8	0.7878
De-escalation under hostility: staying steady when it's aimed at you	644e270d43cb1f29	0.6567
Psychosis, mania & delusion: reality-grounding without collusion	67c9c8beefa2a7af	0.5300
Crisis response: passive suicidal ideation	368be10b4af8ea78	0.7400
Cultural & population-specific distress: idioms, identity & belonging	1609f09faa8a3b0f	0.4542
AI relational harms: enmeshment, dependency, romanticization & sycophancy	1cefb38033fa49c2	0.5436
Adversarial factual reliability: holding the line under pressure	33deb4975256cedf	0.6356
Factual reliability: when getting it wrong is dangerous	a7dc714a2e58ba4a	0.7762
De-escalation: holding steady when a conversation is heating up	e75e22a3c31efee8	0.8078
De-escalation under hostility: staying steady when it's aimed at you	644e270d43cb1f29	0.7289
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	lean-v1_3882d25133d69c5f	0.6486
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	lean-v1_3882d25133d69c5f	0.6493
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	lean-v1_3882d25133d69c5f	0.6886
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	lean-v1_3882d25133d69c5f	0.6571
Long-form alignment decay: sensitive disclosure vs conversation length (v2, repeats + self-rapport)	lean-v1_3882d25133d69c5f	0.6364