REACH VET: The Honest SDVOSB Role Near the VA's Suicide-Risk Model

What is REACH VET, in plain language?

REACH VET — Recovery Engagement and Coordination for Health, Veterans Enhanced Treatment — went into national use across the VA in April 2017. Once a month, a statistical model runs against VA electronic health record data and identifies the small tier of patients at the highest modeled statistical risk — roughly the top 0.1% per facility, about 6,700 veterans a month. A flag is not a diagnosis and not a decision. It routes to a facility coordinator, who notifies the veteran’s mental-health or primary-care provider; the provider then reviews the care plan, reaches out to the veteran (typically within about two weeks), and decides whether to adjust care or build a safety plan. The human clinician is the actor throughout — the model is a triage signal that says “look here sooner,” not a verdict.

The VA has reported that REACH VET has helped identify more than 130,000 veterans at elevated risk since launch. An updated version, REACH VET 2.0, rolled out in 2025; it added military-sexual-trauma and intimate-partner-violence variables and removed race and ethnicity as inputs.

Who built it — and why that’s the boundary

REACH VET was built and is governed inside the VA: the Office of Mental Health and Suicide Prevention (OMHSP), the VHA National AI Institute (NAII), the Rocky Mountain MIRECC, the Program Evaluation and Resource Center, and academic and NIMH research partners. No public contract identifies an outside vendor that developed or validated the core algorithm. That’s the boundary every responsible firm should respect: the predictive engine is the VA’s, and changes to it move outcomes for thousands of people at the worst moment of their lives. A brand-new SDVOSB proposing to “rebuild the suicide model” is the clearest possible signal that it doesn’t understand the work. The credible question isn’t “can we build the model” — it’s “what surrounding engineering can we do well, under a prime, without ever touching the clinical reasoning layer.”

What does the evidence actually say?

Honestly, it’s mixed, and a trustworthy page says so. A 2021 JAMA Network Open evaluation found that being identified by REACH VET was associated with more outpatient care, more documented safety plans, and fewer inpatient admissions and ED visits — and a small reduction in documented suicide attempts — but it did not find a statistically significant reduction in death by suicide. A 2025 study in npj Mental Health Research went further on the model’s raw accuracy, reporting a very low positive predictive value (around 0.05%) and a high false-negative rate, with the authors concluding plainly that the underlying model “needs improvement.” Low precision is partly inherent to predicting a rare event, but it’s real, and it’s why the model is a starting signal for human review rather than an endpoint. Separately, a 2024 investigation found the original model treated some demographic factors in ways that under-identified women veterans — the critique that drove the REACH VET 2.0 changes. None of this makes REACH VET a failure; it makes it a system under active, in-house refinement — which is exactly why the surrounding monitoring and documentation work matters.

Why is this high-impact AI, and what does that trigger?

A model that stratifies access to enhanced clinical outreach on the basis of suicide risk is, under OMB Memorandum M-25-21, unambiguously high-impact AI — healthcare diagnosis and treatment applications are a presumed high-impact category. That classification triggers a defined set of minimum practices: pre-deployment testing in context, ongoing performance and impact monitoring, a documented AI impact assessment and risk-mitigation plan, real human oversight, and the ability to pause or discontinue. The VA’s published M-25-21 compliance plan folds AI review into its ATO and intake processes and ties back to the VA Trustworthy AI Framework; both REACH VET and REACH VET 2.0 appear in the VA’s 2025 AI use-case inventory. The governance layer is covered in depth in the M-25-21 spoke — and notably, producing those governance artifacts is itself a legitimate, lower-clinical-risk place for an outside firm to help.

What does the PHI bar mean for a Cloudflare-native firm?

REACH VET runs on mental-health and substance-use records — among the most protected data the VA holds, with heightened confidentiality under 38 U.S.C. § 7332 and steep penalties for knowing disclosure. Workloads at that sensitivity live inside the VA Enterprise Cloud (VAEC), a FedRAMP High environment on AWS GovCloud and Azure Government, under a VA Authority to Operate. That’s the binding constraint on what a Cloudflare-native firm can honestly touch: Cloudflare for Government is FedRAMP Moderate authorized, with FedRAMP High in process — appropriate for public-facing, non-PHI tooling, but not the boundary for clinical PHI. The honest posture is plain: anything touching protected records runs inside VAEC under a prime’s authorization; Cloudflare-native delivery fits the non-PHI front edge. (CMMC, for the record, doesn’t enter into it — that’s DoD-only.)

So where can an SDVOSB actually help?

In the surrounding engineering — and the table below is the honest version, with the clinical-risk level and where the clinician stays in control made explicit.

Who is actually contracting near this work?

The market signal confirms the shape: SDVOSB-eligible work near suicide-prevention analytics exists, but it lives under primes, in bounded scopes — not in “build the model” awards. JJR Solutions (now part of LMI), an SDVOSB, holds data-and-surveillance support work for OMHSP, including a 2024 SBIR Phase III contract with a ceiling near $19.6M running through 2029. An SDVOSB mentor-protégé joint venture, Clear Vantage Point, won the roughly $154.9M National Safeguard Initiative Against Veteran Suicide in 2023 — a broad feasibility-and-implementation effort, not a predictive-model build. ThunderCat Technology has delivered ReflexAI’s responder-training tooling for the Veterans Crisis Line under multiple awards. And the VA’s $20M Mission Daybreak challenge funded a field of prevention tools, of which only a few have confirmed follow-on VA contracts. For a brand-new firm, the realistic entry is teaming with one of these primes, not competing as a clinical-AI prime.

Working with Truvisory

Truvisory is an SBA-verified SDVOSB, founded by a combat veteran. We have no VA clinical-AI past performance, and we won’t pretend otherwise.

If you’re a prime, integrator, or VA program office working the surrounding engineering for REACH VET or related suicide-prevention analytics — data pipelines into VAEC, clinician-facing dashboard UX, M-25-21 governance documentation, or monitoring and bias-audit tooling — and you need a Cloudflare-native, FedRAMP-aware, fixed-scope SDVOSB subcontractor, that’s the work we can do. The core model belongs inside the VA, with the clinicians and researchers who built it.

Veterans Crisis Line: dial 988, then press 1 · text 838255 · chat at VeteransCrisisLine.net. Free, confidential, 24/7.

For the rest of the map — the governance layer, the ATO and VAEC mechanics, the claims-automation sibling, and how a new firm teams in — start at the pillar.