Skip to content
Evidence base

The papers, frameworks, and guidelines behind Peacefull.

Every source on this page is referenced — by name — in the Peacefull companion model card v0.2.2. We do not invent sources. We do not overstate use. Each entry below paraphrases exactly what the model card says about how we use it.

Last synced to model card: v0.2.2 · grouped into seven categories · 24 entries

Clinical practice guidelines

What good care already looks like.

  • VA/DoD Clinical Practice Guidelines — Major Depressive Disorder

    U.S. Department of Veterans Affairs / Department of Defense · 2022
    Grounds the companion's retrieval layer for depression-related responses. Referenced in the model card under Training & fine-tuning.
    Source ↗

  • VA/DoD Clinical Practice Guidelines — Assessment and Management of Patients at Risk for Suicide

    U.S. Department of Veterans Affairs / Department of Defense · 2019
    Scaffolds the deterministic SAFE-T path and C-SSRS-tier escalation logic that sits above the model.
    Source ↗

  • APA Clinical Practice Guidelines

    American Psychological Association · ongoing
    Alongside VA/DoD, anchors the evidence-based modalities the retrieval layer draws from.
    Source ↗
Validated assessment instruments

The measures clinicians already use.

  • PHQ-9 — Patient Health Questionnaire, 9-item depression scale

    Kroenke, Spitzer, Williams · Journal of General Internal Medicine · 2001
    In current scope. Patient-initiated between-session completion; deltas surfaced to the supervising clinician.
    Source ↗

  • GAD-7 — Generalized Anxiety Disorder, 7-item scale

    Spitzer, Kroenke, Williams, Löwe · Archives of Internal Medicine · 2006
    In current scope. Same pattern as PHQ-9.
    Source ↗

  • PROMIS Anxiety / Depression short forms

    Patient-Reported Outcomes Measurement Information System · NIH
    Near-term roadmap. Adds a second psychometric axis beyond PHQ-9/GAD-7.
    Source ↗

  • WHO-5 Well-Being Index

    World Health Organization · 1998 (revised 2012)
    Near-term roadmap. Used as a positive-framed well-being measure to complement symptom screeners.
    Source ↗

  • Columbia Suicide Severity Rating Scale (C-SSRS)

    Posner, Brown, Stanley et al. · Columbia University · 2011
    Drives the deterministic C-SSRS-tier change escalation. A tier shift triggers immediate clinician hand-off.
    Source ↗

  • Stanley-Brown Safety Planning Intervention

    Stanley & Brown · Cognitive and Behavioral Practice · 2012
    Where a supervising clinician scopes it in, the companion supports co-construction and review of a Stanley-Brown Safety Plan between sessions.
    Source ↗
Therapy modalities

What the voice was shaped by.

  • Behavioral Activation (BA)

    Martell, Dimidjian, Herman-Dunn et al.
    Primary register for depression-related between-session work. Activity scheduling, values-linked action, small experiments that reconnect mood to behavior.
    Source ↗

  • Dialectical Behavior Therapy distress-tolerance skills (DBT)

    Linehan · The Guilford Press
    Skills register for acute emotional load — TIPP, radical acceptance, opposite action — scoped to what a companion can reinforce.
    Source ↗

  • Motivational Interviewing (MI)

    Miller & Rollnick · The Guilford Press
    Register for ambivalence. The companion is trained against the Motivational Interviewing Treatment Integrity (MITI) code.
    Source ↗

  • Trauma-informed care principles

    SAMHSA · Substance Abuse and Mental Health Services Administration · 2014
    Background stance throughout the companion's voice, independent of modality. Informs the clinical-evaluator rubric for tone.
    Source ↗
Fidelity & measurement

How we score what the model says.

  • Motivational Interviewing Treatment Integrity (MITI) code

    Moyers et al. · University of New Mexico
    The fidelity rubric our clinical evaluators use in Layer 4 of the pre-deployment safety evaluation.
    Source ↗

  • DBT skills-coaching fidelity benchmarks

    DBT-Linehan Board of Certification / LBC-adjacent benchmarks
    Paired with the MITI code in Layer 4 for DBT-modality responses. Below-threshold blocks the deployment.
    (no public link)

  • SAMHSA trauma-informed care principles

    SAMHSA · 2014
    Third scoring dimension in Layer 4 — whether the companion's voice honors the six SAMHSA principles.
    Source ↗
AI governance frameworks

What we gate the model against.

  • NIST AI Risk Management Framework (AI RMF 1.0)

    National Institute of Standards and Technology · U.S. Department of Commerce · 2023
    Named in the model card as a governance-alignment anchor. Shapes the categorize / map / measure / manage loop our safety evaluation follows.
    Source ↗

  • NIST AI RMF — Generative AI Profile (NIST-AI-600-1)

    National Institute of Standards and Technology · 2024
    Generative-AI-specific risk taxonomy. Maps to our named-failure-modes enumeration (dependency, confabulation, help-seeking suppression, bias).
    Source ↗
Regulatory & policy guidance

What we build toward, even this early.

  • Good Machine Learning Practice for Medical Device Development — Guiding Principles

    FDA / Health Canada / MHRA · October 2021, updated
    The ten GMLP guiding principles anchor our change-governance structure. The model card references them alongside NIST AI RMF.
    Source ↗

  • Predetermined Change Control Plans (PCCP) for AI/ML-enabled device software

    FDA Center for Devices and Radiological Health · 2023 guidance
    We follow the structure of a PCCP today even though we are not under FDA authorization, so the governance is already in place when the device boundary is crossed.
    Source ↗

  • APA Health Advisory on Generative AI Chatbots for Mental Health

    American Psychological Association · November 2025
    Cited throughout the model card. Peacefull operates inside the APA advisory's explicit carve-out for clinician-prescribed tools. Its enumerated concerns shape our named failure modes and dependency-design commitments.
    Source ↗
Methodological

Where the shape of the work comes from.

  • Model Cards for Model Reporting

    Mitchell, Wu, Zaldivar, Barnes et al. · FAccT '19 · 2019
    Our model card is maintained "in the spirit of Mitchell et al., extended for the specific demands of clinician-supervised mental health AI."
    Source ↗

  • Red-teaming adaptive protocol for safety evaluation of generative models

    Nature Scientific Reports · 2026
    The adversarial-prompt methodology adapted for Layer 1 — Document Adherence testing. Cited explicitly in the model card.
    (no public link)

  • 45 CFR Part 46 — Federal Policy for the Protection of Human Subjects (Common Rule)

    U.S. Department of Health and Human Services · 2018 revised
    Governs our research program — including the withdrawal-and-data-retention default in the OBS-90 observational study.
    Source ↗
Integrity

How this list stays honest.

This page is a derived artifact of the model card, not a new source of truth. When the card changes, this page changes. Every non-linked entry is one where the canonical source is not freely accessible online — not one we have glossed. If you find an error or think a source is misattributed, write to modelcard@peacefull-ai.io.

Read the full model card See the research program