Research & evidence
The case for the work.
The limits of the evidence.
Better access, useful human–AI collaboration, and accountable decisions are related challenges—not interchangeable claims.
The current investigation.
KeyRD is the company’s independent research and development project into governed, auditable adaptive support and longitudinal dietary adherence. Its prototype and synthetic demonstrations make a proposed design inspectable; its scientific validity and effectiveness remain unproven.
Read the study agenda and collaboration questions · Explore the synthetic Maya demo
Evidence for the need. Not evidence of our success.
The selected research below motivates Statera’s direction. It is not an exhaustive systematic review, and none of these studies evaluated Statera or KeyRD. Findings, limits, and our design inferences are identified separately.
Some sources are empirical studies, others syntheses or guidance. They do not have the same evidentiary role.
Methods for the proposed adherence research.
JITAI design.
Nahum-Shani and colleagues describe the relationship among decision points, tailoring variables, intervention options, decision rules, proximal outcomes and distal outcomes. This framework helps specify what KeyRD would need to implement and test; it does not validate our choices.
Micro-randomized trials.
Klasnja and colleagues describe repeated randomization as a way to investigate proximal intervention effects and how those effects vary over time and context. A fixed synthetic trajectory cannot supply that evidence.
Dietary measurement.
The National Cancer Institute’s Dietary Assessment Primer explains measurement error in self-reported dietary data. A lunch report needs a clear definition and a validation strategy; opening a prompt is not dietary adherence, and a missing report is not nonadherence.
National Cancer Institute. Dietary Assessment Primer: Key Concepts.
These sources guide design and interpretation. None evaluates KeyRD. The candidate study still requires clinical and methodological development.
Broader background for Statera’s long-term aim.
The following sources address nutrition care, digital support and broader questions about AI and accountability. They have a different role from the JITAI and measurement methods above.
1 / Field study · 2025
AI can extend useful capability.
Finding or guidance. A staggered rollout to 5,172 customer-support agents was associated with 15% more issues resolved per hour on average. Gains were larger for less-experienced workers.
Important limit. One firm, one occupation, and one tool. Effects varied; the most experienced workers had small quality declines. This is not a clinical study or evidence about autonomous systems.
Our interpretation. AI may help useful expertise reach further. Which people benefit, and at what cost, must be measured in the setting of use.
Brynjolfsson, Li & Raymond. Generative AI at Work. The Quarterly Journal of Economics (2025).
2 / Systematic review & meta-analysis · 2024
Human involvement is not enough by itself.
Finding or guidance. Across 106 experiments, human–AI combinations performed better than humans alone on average, but worse than the better of the human-only or AI-only alternatives.
Important limit. The experiments were published in 2020–2023, differed substantially, and do not represent every task or newer model. Task performance also does not settle who has legitimate authority.
Our interpretation. Test the whole workflow against credible alternatives. Adding a human approval step is not, on its own, evidence of benefit or safety.
3 / Algorithm audit · 2019
The objective can reproduce unequal treatment.
Finding or guidance. An audit found that a widely used health-risk algorithm underestimated the needs of Black patients because it predicted spending rather than illness. Unequal access made cost a biased proxy for need.
Important limit. This finding concerns a particular algorithm and allocation setting, not all AI. Predictive accuracy on a convenient target is not the same as fair treatment.
Our interpretation. Examine the goal, the proxy, and the distribution of consequences—not only the average score. People need meaningful ways to challenge consequential decisions.
4 / WHO–World Bank global monitoring · 2025
Access remains a problem of both reach and resources.
Finding or guidance. The report estimates that 4.6 billion people were not fully covered by essential health services in 2023, and 2.1 billion faced financial hardship from out-of-pocket health spending in 2022.
Important limit. These are global health-system estimates, not estimates of demand for KeyRD or problems software alone can solve. Coverage gaps are not equivalent to receiving no care at all.
Our interpretation. Wider reach must include affordability and practical access. Software should complement care capacity and resources, not be presented as a substitute for them.
WHO & World Bank. Tracking Universal Health Coverage: 2025 global monitoring report.
5 / Clinical meta-analysis · 2021
Dietetics begins with expertise worth supporting.
Finding or guidance. Fourteen studies involving 3,338 adults compared dietitian-delivered nutrition therapy with advice from other health professionals. The pooled HbA1c difference favored dietitian care by 0.47 percentage points.
Important limit. Evidence quality was limited by heterogeneity, bias, publication bias, and indirectness. These results concern type 2 diabetes care—not software, all dietetics, or KeyRD.
Our interpretation. Begin by supporting professional care, not presuming that a software substitute can reproduce its benefits.
6 / Systematic review of randomized trials · 2020
Digital delivery is promising, but not a settled answer.
Finding or guidance. A review of five studies involving 1,056 adults found dietary-behavior improvements in four studies, while clinical results were inconsistent.
Important limit. The evidence was preliminary, with different diets, delivery methods, outcome measures, and intervention components. It does not isolate the value of automation or adaptive AI.
Our interpretation. Treat digital delivery as a hypothesis about how care can be supported. Measure adherence, meaningful outcomes, and burden rather than assume engagement equals benefit.
7 / Systematic review · 2022
A plausible app can still lack strong evidence.
Finding or guidance. Five studies with 334 participants suggested positive trends, but all three randomized trials had high risk of bias. The review could not draw firm conclusions about diet adherence or blood-pressure benefit.
Important limit. A small, specific literature on DASH apps; it is not evidence that all digital nutrition tools fail.
Our interpretation. Distinguish usability, technical success, and clinical effectiveness. A convincing demonstration cannot substitute for a well-designed evaluation.
8 / Ethical guidance · 2021
Authority and accountability are design obligations.
Finding or guidance. WHO guidance identifies autonomy, safety, transparency, accountability, inclusiveness, equity, and sustainability as central considerations for AI in health.
Important limit. This is ethical guidance, not an experiment, a certification, or an endorsement of Statera. It does not validate a particular implementation.
Our interpretation. Translate principles into enforceable boundaries and reviewable decisions. Keep patient choice and professional responsibility explicit.
World Health Organization. Ethics and governance of artificial intelligence for health (2021).
What these sources cannot tell us.
They cannot establish demand for KeyRD, safe deployment, reduced workload, improved clinical outcomes, or success across other domains. Those require product-specific evaluation.
Our 20-year aim is a value-guided direction, not a prediction derived from these studies.
The questions KeyRD must answer