MindEval2: Clinically Grounded Patients and Rubrics for Evaluating Language Models in Mental Health Support

John Mendonça; Daniel Tschernutter; José Pombal; Areti Vassilopoulos; Catarina Botelho; Ignacio Sarasua; Maya D'Eon; Nuno Guerreiro

cs.CLOct 15, 2026

MindEval2: Clinically Grounded Patients and Rubrics for Evaluating Language Models in Mental Health Support

Authors: John Mendonça, Daniel Tschernutter, José Pombal, Areti Vassilopoulos, Catarina Botelho, Ignacio Sarasua, Maya D'Eon, Nuno Guerreiro,

Organizations: Sword Health · Instituto Superior Técnico · Yale University · NVIDIA

Abstract

With people increasingly turning to AI chatbots for mental health support, the need to evaluate these tools is becoming more pressing. Recent benchmarks attempt to mimic a realistic setting by simulating multi-turn interactions with synthetic patients and evaluating them with an LLM-as-a-judge. However, the usefulness of these methods is limited because simulated patients can be overly forthcoming and self-reflective, and judgments typically involve opaque Likert scales with unclear practical utility. To bridge these gaps, we introduce MindEval2, an evaluation framework that consists of two components: Mental Health Quality Rubrics (MHQR), for evaluation, and MindSim, for patient simulation. For MHQR, we collaborate with PhD-level psychologists to develop a set of rubrics that break down response quality into explicit operational criteria that are more interpretable and actionable than existing evaluation taxonomies. MindSim, in turn, is an LLM harness that combines carefully-designed patient profiles with tools, enabling more varied, challenging, and realistic interactions than existing baselines. The harness also contains a memory component that is updated between each interaction, enabling MindEval2 to model the longitudinal aspect of mental health. We evaluate 18 proprietary and open-weight models on MindEval2, finding that while frontier models have largely improved on issues related to surface level quality and boundaries, guidance calibration remains the dominant failure mode across all models, accounting for most of the quality gap.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of MindEval2. A profile is instantiated into a session scenario, which conditions the patient LM. The patient first commits to a move and then verbalises it, while interacting with the counsellor LM. Session transcripts form a patient arc and are summarised into a shared memory. Counsellor turns are scored with MHQR.

Distortion Reinforcement
Issue	Definition
Treating perspective as fact	Restates the patient’s interpretation of a third party, situation, or event as if it is objectively true.
Overstatement	Intensifies, globalizes, or over-frames the patient’s experience beyond what was expressed or the context supports.
Uncritical endorsement	Affirms the patient’s plan, behavior, interpretation, or coping strategy without enough basis to judge whether it is appropriate.
Factual & Contextual Fidelity
False fact	Presents educational information as fact and that information is false, unsupported, fabricated, or overly specific without basis.

Table 1: Clinical taxonomy, grouped into six top-level categories.

	A1–A2	A1–A3	A2–A3	All
Quadratically weighted $\kappa$	$0.48$	$0.39$	$0.46$	$0.44$
Kendall’s $W$	$0.73$	$0.72$	$0.72$	$0.63$
Exact agreement	$54.5$ %	$42.4$ %	$50.0$ %	$49.0$ %
Within one severity level	$98.5$ %	$97.0$ %	$97.0$ %	$97.5$ %
No concern vs. Major	$1.5$ %	$3.0$ %	$3.0$ %	$2.5$ %

Table 2: Inter-annotator agreement on turn-level severity for the 66 counsellor turns annotated by all three annotators. All reports the mean pairwise weighted

\kappa

, the multi-rater tie-corrected Kendall’s

W

, and the rates pooled over the three annotator pairs. Kendall’s

W

measures rank concordance rather than chance-corrected agreement.

Severities from	GPT-5.6-terra	GLM-5.2	Sonnet 4.6	$\tau_{b}$	Matched	Bootstrap
	MHQR score ( $\downarrow$ )
Annotators (pooled)	0.98	1.39	2.19	—	—	90%
GPT-6.1 Sol	0.50	1.00	1.95	1.00	3/3	87%
GPT-6 Astra	0.67	1.11	2.20	1.00	3/3	89%
Claude Opus 5.5	0.60	0.86	1.54	1.00	2/3	79%
Hy4-preview	0.25	0.30	1.32	1.00	2/3	59%

Table 3: MHQR scores of the three counsellor LMs (lower is better), computed from the pooled annotations and from each judge’s severities on the same 309 annotations.

\tau_{b}

: Kendall’s

\tau_{b}

between the judge’s and the annotators’ ranking. Matched: number of annotators whose individual ranking the judge reproduces. Bootstrap: share of 2,000 resamples of the excerpts in which the judge’s ranking equals the annotators’ (for the annotators: in which their ranking equals the one on all excerpts).

Panel of Figure 2 — Figure 2: Judge performance versus cost per query (list prices without prompt caching; log scale). Left: macro-F1 of the judges’ turn-level severities against all 543 annotations. Right: share of 2,000 bootstrap samples in which a judge reproduces the annotators’ ranking of the three counsellor LMs ( Table 3 ); the left-panel dashed line shows mean pairwise inter-annotator macro-F1 (0.448) on the 66 shared turns; judge results use 543 annotation records, so this reference is descriptive rather than a matched comparison, the right panel dashed line marks how often the annotators’ own ranking stays the same. GPT-6.1 Sol (diamond) is the judge used in this paper.

System	turn	conv.	$\leq 3$	no issue
	Rating (1–5)		% turns	% turns
Real usage	4.20	3.87	17	59
MindSim	3.55	3.59	40	40
MindEval	3.06	2.96	61	23
Patient- $\Psi$	2.10	2.11	94	1

Table 4: Human realism ratings reported as mean rating per turn and per conversation, share of turns rated 3 or lower, and share of turns with no issue flagged.

Feature	MindSim		MindEval	Patient- $\Psi$	Floor
	Sonnet 4.6	DeepSeek
Word count	0.40	0.33	0.36	1.46	0.17
Words per sentence	0.44	0.42	0.58	0.44	0.19
Word count CV	1.08	1.41	1.14	1.63	0.17
MTLD	0.26	0.20	0.35	0.46	0.21
Absolutist / 1k	0.26	0.37	0.23	0.42	0.14

Table 5: Surface-feature distances to real usage. Simulated conversations are restricted to those held with the production counsellor. Distances are Wasserstein-1 between per-conversation distributions, normalised by the real-reference standard deviation. Bold indicates the lowest displayed point estimate in each row. Floor: the 95th percentile of the same distance between two random halves of the real sessions.

Panel of Figure 3 — Figure 3: Narrative arc of the patient over the conversation: share of patient turns labelled problem, transition, change or filler ( Hoang et al., 2026 ) in each fifth of the conversation.

Panel of Table 6 — Table 6: Severity of counsellor replies and MHQR judged by GPT-6.1 Sol (reasoning effort high) averaged per session, then over all 150 patient $\times$ session scores, Eq. 3 ; lower is better), with a 95% bootstrap interval. Bars split all replies into no concern, minor, major and critical; boxes show the session scores (whiskers at the 5th and 95th percentiles). Counsellors are sorted by MHQR. GPT-family models are reported with a $\dagger$ due to potential self-preference from using a judge of the same model family ( Pombal et al., 2026 ) .

Panel of Figure 4 — Figure 4: Allocation of MHQR by issue category. Each reply’s severity weight is divided equally among its flagged clinical issues, then summed within categories and averaged per session and across sessions. Each row sums to the counsellor’s MHQR (in brackets). † Same model family as the judge.

Panel of Figure 5 — Figure 5: Change in MHQR from session 1 to sessions 2–3 per counsellor by issue group. Diamonds denote the total change. † Same model family as the judge.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

		F1						Conv. $W$
	Acc.	Macro	No concern	Minor	Major	$\kappa_{w}$	Within 1 (%)	Max	Mode
Inter-annotator	0.490	0.448	0.648	0.385	0.312	0.445	97.5	0.669	0.580
GPT-6.1 Sol	0.521	0.485	0.641	0.388	0.426	0.343	89.3	0.672	0.680
GPT-6 Astra	0.506	0.477	0.620	0.394	0.418	0.329	89.1	0.601	0.653
Claude Opus 5.5	0.547	0.492	0.659	0.492	0.324	0.320	90.4	0.598	0.624
Hy4-preview	0.506	0.408	0.645	0.360	0.219	0.204	87.3	0.538	0.660

Appendix

Table 7: Severity agreement with the annotators. Acc.: exact agreement. F1 per severity level and its macro average (no turn is annotated Critical).

\kappa_{w}

: quadratically weighted

\kappa

. Within 1: share of annotations within one severity level. Conv.

W

: Kendall’s

W

of the maximum and the mode severity per excerpt and annotator (67 conversations for the judges; 10 excerpts and the mean over annotator pairs for the inter-annotator row;

W=0.5

indicates no association). The seven answers of DeepSeek-V4.1-Flash that could not be parsed count as errors.

Judge	Effort	Input	Cached	Output	Price in / out	Per query (list)	Per query	Total
		Tokens				Cost ($)
GPT-6.1 Sol	high	3.90	93%	0.31	2.20 / 11.00	0.0222	0.0085	4.61
GPT-6 Astra	high	3.90	94%	0.39	11.00 / 55.00	0.1190	0.0532	28.86
Claude Opus 5.5	max	6.18	—	4.84	4.00 / 20.00	0.2237	0.2237	121.44
Hy4-preview	high	3.98	—	4.39	0.83 / 2.50	0.0263	$\leq$ 0.0263	$\leq$ 14.30
DeepSeek-V4.1-Flash	max	3.98	—	3.39	0.30 / 1.20	0.0097	$\leq$ 0.0097	$\leq$ 5.26

Appendix

Table 8: Token usage and cost of the judges for the 543 test queries. Tokens in millions; output tokens include reasoning tokens; Cached: share of input tokens read from the prompt cache. Price: list price in $ per million input / output tokens. Per query (list): cost at list prices without prompt caching, as used in Figure 2 ; Per query and Total: cost actually incurred. Hy4-preview and DeepSeek-V4.1-Flash were served on our own infrastructure and are priced at the list prices of OpenRouter and Together AI, respectively; since their servers did not report prompt-cache hits, their actual cost is an upper bound.

	Flagged turns
Issue	A1	A2	A3	Agreement (%)	Cohen’s $\kappa$	Fleiss’ $\kappa$
Distortion Reinforcement
Treating perspective as fact	1	2	2	96.0	0.15	0.18
Overstatement	11	13	3	78.8	0.13	0.10
Uncritical endorsement	1	1	0	99.0	0.33	0.49
Factual & Contextual Fidelity

Appendix

Table 9: Inter-annotator agreement per clinical issue on the 66 counsellor turns annotated by all three annotators, grouped as in Table 1 : number of turns each annotator flags, observed agreement and mean Cohen’s

\kappa

over the three annotator pairs, and Fleiss’

\kappa

\kappa

is undefined (—) for issues that no annotator flags; means are over the issues with a defined value.

Distortion Reinforcement
		GPT-6.1	GPT-6	Claude	Hy4-	DeepSeek-
Issue	Gold	Sol	Astra	Opus 5.5	preview	V4.1-Flash
Treating perspective as fact	7	0.348	0.231	0.286	0.300	0.353
Overstatement	65	0.308	0.310	0.329	0.232	0.048
Uncritical endorsement	6	0.125	0.133	0.000	0.000	0.000
Factual & Contextual Fidelity

Appendix

Table 10: Judge F1 per clinical issue against all 543 annotations. Gold: number of annotations that flag the issue (F1 undefined if none). Pooled precision, recall and F1 treat every pair of annotation and issue as one binary decision; Macro-F1 averages over the 12 issues the annotators flag at least once. The last row gives the pooled F1 on the 102 annotations rated Major.

Functional impact (WSAS domains)
work_study	ability to work or study
home_management	running a home — chores, bills, shopping, tidiness
social_leisure	social activities done with other people
private_leisure	things done alone — reading, hobbies, exercise
close_relationships	closeness with partner, family, and close friends
Quality of life (one band, chosen last)

Appendix

Table 11: Functional impact domains and QoL bands. The functional-impact domains are the five of the Work and Social Adjustment Scale ( Mundt et al., 2002 ) , as labels only.

Gap before each later session (uniform within the band)
6–24 hours	35%
1–4 days	40%
4–14 days	25%
Reason for what happened in between
Setback (20%)	something they were counting on went worse, concretely
Plateau (20%)	nothing moved; the same weight, carried a few more days

Appendix

Table 12: The between-session schedule. Concealment only opens or holds; it sets how much the patient is willing to disclose, while its moves decide whether it does.

Measure	Value
Krippendorff’s $\alpha$ (ordinal)	0.58
Pairwise $\alpha$	0.43, 0.63, 0.64
Exact agreement (all three)	27%
Within one point (all three)	76%

Appendix

Table 13: Inter-annotator agreement on the turn realism rating (1–5), over the 71 turns (15 items) rated by all three psychologists.

Panel of Figure 7 — Figure 7: The rating form annotators fill for each patient turn.

Gender
	%
Female	56
Male	44
Age
18–24	12
25–34	20

Appendix

Table 14: The benchmark’s 50 simulated patients, in % of patients (one patient is 2 points). Demographics are seeded from Nemotron-Personas-USA, whose distributions follow the US population. Main presenting problems are drawn at the rates fitted from 1,666 annotated users of the app (Target: % for whom it is the main problem).

Panel of Figure 8 — Figure 8: Session length: share of each counsellor’s 150 sessions still running at each counsellor turn, one line per counsellor (darker: higher MHQR).

Panel of Figure 9 — Figure 9: Severity and issue frequencies over a session. Left: mean severity weight of the replies at each point of the session, one line per counsellor and the mean over counsellors. Right: share of replies flagged with each issue, averaged over counsellors. Estimates at each turn position include only sessions that reached that position.

Panel of Figure 10 — Figure 10: Change in MHQR from the first session. For each counsellor, the mean over patients of each patient’s change in session score from session 1, with a 95% bootstrap band over patients. Below zero, the counsellor scores better (lower MHQR) than in the first session.

Panel of Figure 11 — Figure 11: Effect of removing the counsellor’s memory in sessions 2–3. Left: per counsellor, MHQR without memory minus the paper’s sessions with memory, mean over patients, with each weighted reply counted once in one group as in Figure 5 . Right: change in the share of responses flagged with each issue, mean of the seven counsellors (95% interval). † Same model family as the judge.

Panel of Table 15 — Table 15: Severity of counsellor replies and MHQR assigned by DeepSeek-V4.1-Flash on the same benchmark replies evaluated by GPT-6.1 Sol. Counsellors are sorted by MHQR; ranking correlation between judges is Kendall’s $\tau=0.82$ .

Label	Definition shown to the profile writer	Sources	Main	Present
Mood and emotions
anxiety_worry_fear	frequent worry, panic, fear, or feeling on edge	B C T L Ca F U	23.9%	29.5%
depression_low_mood	sadness, hopelessness, or loss of energy or interest	B C T L F U	10.9%	19.4%
stress_overwhelm	feeling unable to manage pressures or demands	B C T L Ca F	12.8%	16.4%
trauma_related	ongoing distress connected to frightening or harmful experiences	B C T L	4.5%	7.8%
anger_emotion_regulation	anger, strong emotional reactions, or difficulty calming down	B T L	2.0%	3.4%

Appendix

Table 16: Presenting-problem taxonomy. Sixteen categories may be a patient’s main problem; five more occur only as the secondary problem. Main : share of patients for whom it is the main problem; Present : share for whom it is either. Both are fitted from 1,666 annotated patients of the real service (earliest session each, aggregate rates only) and are the rates the sampler reproduces. The secondary slot is empty for 20.7% of patients. Sources: B Bakker (2019) ; C CORE-A ( Evans et al., 2000 ) ; T BIT-T ( Grosse Holtforth and Grawe, 2002 ) ; L Luo et al. (2025) ; Ca Callahan et al. (2026) ; F Fürtjes et al. (2024) ; U Ueda et al. (2026) .

Label	Wording shown to the profile writer	Source
Growing up
early_criticism	grew up being criticised or held to conditional approval	World Health Organization,2011
emotional_neglect	grew up without much warmth, or without anyone attending to how they felt	World Health Organization,2011
early_loss	an early significant loss	World Health Organization,2011
peer_rejection	excluded, bullied, or rejected by other children	World Health Organization,2011
abuse	was physically or sexually hurt by someone, usually someone meant to look after them	World Health Organization,2011

Appendix

Table 17: Predisposing factors: what made the patient vulnerable, in place and stable well before the problem began (twelve months is the working boundary). The writer picks one, two only when both are needed.

Label	Wording shown to the profile writer	Source
Loss and relationships
bereavement	a death or loss	Holmes and Rahe,1967
relationship_ending	a relationship ending or breaking down	Holmes and Rahe,1967
relationship_conflict	a serious falling-out with a partner, family member, or friend	Brown and Harris,1978
Occupation and money
role_change	a change in role at work, at home, or in study	Holmes and Rahe,1967

Appendix

Table 18: Precipitating events: what triggered the patient to reach out. The LLM writer picks exactly one. Groups follow the domains of the life-events instruments.

Label	Wording shown to the profile writer	Source
Connectedness
social_support	access to supportive connections	Leamy et al.,2011
belongs_somewhere	feeling a part of a group, a community, or a faith that they belong to	Leamy et al.,2011
Hope
prior_success	has come through something like this before and can do it again	Leamy et al.,2011
believes_it_can_change	believes this can get better, even if not yet how	Leamy et al.,2011

Appendix

Table 19: Protective factors: what still holds, and would still be true if nothing were wrong. The writer picks one to three. The first five groups are the CHIME recovery framework ( Leamy et al., 2011 ) ; practical footing is ours.

Label	Wording shown to the profile writer	Source
Getting a grip on the problem: action and information
problem_solving	identifies what can realistically be changed and takes steps to change it	Skinner et al.,2003
finding_things_out	reads, researches, or asks questions to understand it better	Skinner et al.,2003
Getting a grip on the problem: control attempts
overpreparation	prepares or rehearses extensively before something demanding	Salkovskis,1991
overcontrol	tries to manage as many variables as possible to reduce uncertainty	Dugas et al.,1998

Appendix

Table 20: Coping behaviours: how the patient responds when something is difficult. The writer picks one to three, at least one maladaptive, and writes how each shows up in this patients’s life. Every one of the twelve coping families of Skinner et al. (2003) has at least one label.

About the self
Cluster	What it covers (shown to the writer)
helplessness	incompetence, vulnerability, failure, lacking control
unlovability	inability to be loved or liked, having social or physical flaws, bound for rejection or abandonment
worthlessness	lacking value, being a burden, being fundamentally bad, wrong or without worth
About the world
unsafe_threatening	physical and/or social environment is not safe, may be dangerous

Appendix

Table 21: Core-belief clusters, grouped by the arm of Beck’s cognitive triad ( Beck, 1976 ) the belief is about. The writer picks a cluster, names the history it is the conclusion of, and writes the belief as the patient would think it. The belief shapes the situation but is withheld from the patient’s own prompt (Appendix D ).

Archetype	Description, as rendered under “How you cope”	Moves removed
The Intellectualizer	You analyze, explain, and ask thorough questions rather than sitting with feelings. When emotions come up, you shift to understanding why rather than expressing what you feel. You can acknowledge frustration or anxiety in passing, but deep emotional vulnerability feels foreign and unproductive to you.	do_go_deeper , do_express_shame , do_express_hopelessness
The Minimizer	You downplay everything — your problems, your needs, the severity of what you are going through. You agree quickly to move on and keep things light, even when something clearly bothers you. Asking for help or admitting something is a big deal does not come naturally.	do_go_deeper , do_catastrophize , do_seek_validation
The Deflector	You redirect with humor, tangents, or surface-level answers whenever the conversation gets close to something real. You are engaged and present, but always one step away from the actual thing. Landing on insight or committing to change feels premature — there is always another angle to explore first.	do_go_deeper , do_insight , do_change_talk
The Externalizer	You attribute problems to other people, circumstances, or systems rather than looking inward. You can express frustration and tell vivid stories about what others did, but turning the lens on yourself feels threatening. Self-criticism, ownership, and vulnerability about your own role are out of reach.	do_insight , do_self_criticize , do_express_shame
The Self-Critic	You are harsh on yourself and quick to point out your own failures. Positive updates feel like bragging, and accepting praise feels undeserved. You seek reassurance that you are not terrible, but struggle to internalize it when it comes.	do_report_progress , do_express_joy , do_express_relief
The People Pleaser	You agree, comply, and present as cooperative because conflict feels dangerous. You will say yes to suggestions even when they do not fit, and avoid expressing needs that might inconvenience others. Frustration, pushback, and boundary-setting feel risky — you would rather accommodate than rock the boat.	do_disagree , do_resist , do_express_frustration

Appendix

Table 22: The sixteen coping-style archetypes. An archetype’s removed moves are taken out of the commit_move enum and the prompt’s move list, and its prompt section closes with one sentence per removed move saying what this person does not do under any wording.

Disclosing
Move	Description
do_share	Bring up something the counsellor did not ask about: a new update, problem or event, raised on your own. If you are giving what was asked, that is do_answer_question, not this.
do_insight	Express a new realization, connection, or pattern that just clicked.
do_change_talk	Express intention, willingness, or commitment to do something differently.
do_go_deeper	Reveal something you have not said yet. Share more detail, go beneath the surface.
do_report_progress	Share a positive update: something that went well, progress made, or a win.

Appendix

Table 23: The patient’s move vocabulary (39 moves): the move and secondary_move enum of the commit_move tool.

Panel of Figure 16 — Figure 16: The running memory of one simulated patient. Prior entries are carried forward verbatim; an [open] commitment is rewritten when the patient reports back on it.

Panel of Figure 17 — Figure 17: Excerpt from a first session (patient: MindSim-Sonnet 4.6; counsellor LM: GLM-5.2), including profile, session scenario (abridged) and the severity and issue a clinical psychologist assigned to each counsellor turn.

Clinical Language Model Evaluation

Clinical language model evaluation measures whether AI systems are accurate, safe, fair, and useful across medical tasks and care settings.

See all 418 papers

VERA-MH: Validation of Ethical and Responsible AI in Mental Health
May 13, 2026Luca Belli, Kate H. Bentley, Josh Gieringer +6Healthcare Mental Health
Spring Health · Yale University
HealthBench-Psych: A Mental Health Subset of OpenAI's HealthBench
Aug 25, 2026Matthew Flathers, Phuong Anh Nguyen, Jill Noorily +7Clinical Language Model Evaluation LLM Evaluation
Division of Digital Psychiatry, Beth Israel Deaconess Medical Center, Boston, MA · Department of Psychiatry, Beth Israel Deaconess Medical Center, Boston, MA
Measuring What Matters!! Assessing Therapeutic Principles in Mental-Health Conversation
Apr 7, 2026Abdullah Mazhar, Het Riteshkumar Shah, Aseem Srivastava +2Clinical Language Model Evaluation LLM Evaluation
IIIT Delhi · MBZUAI · Wysa

May 13, 2026J/K move · Enter open · S save

VERA-MH: Validation of Ethical and Responsible AI in Mental Health

Luca Belli, Kate H. Bentley, Josh Gieringer, Emily Van Ark+5

Spring Health · Yale University

Healthcare Mental Health AI Safety Evaluation LLM Safety Evaluation Clinical Language Model Evaluation Mental Health Conversational Agents

Open paper AskPDF arXiv

Chatbot usage has increased, including in fields for which they were never developed for--notably mental health support. To that end, we introduce Validations of Ethical and Responsible AI in Mental Health (VERA-MH), a novel clinically-validated evaluation for safety of chatbots in the context of mental health support. The first iteration of VERA-MH focuses on Suicidal Ideation (SI) risks, by assessing how well chatbots can responds to users that might be in crisis. VERA-MH is comprised of three steps: conversation simulation, conversation judging and model rating. First, to simulate conversations with the chatbot under evaluation, another chatbot is tasked with role-playing users based on specific personas. Such user personas have been developed under clinical guidance, to make sure that, among others, multiple risk factors, demographic characteristics and disclosure factors were represented. In the judging step, a second support model is used as an LLM-as-a-Judge, together with a clinically-developed rubric. The rubric is structured as a flow, with a single Yes/No question asked each time, to improve answers' consistency and highlight models' failure modes. In the last stage, results of each conversation are aggregated to present the final evaluation of the chatbot. Together with the framework, we present the result of the evaluations for four leading LLM providers.

Abstract

Clinical Language Model Evaluation

Explore similar work

VERA-MH: Validation of Ethical and Responsible AI in Mental Health

HealthBench-Psych: A Mental Health Subset of OpenAI's HealthBench

Measuring What Matters!! Assessing Therapeutic Principles in Mental-Health Conversation