Data plane
Checks cohort integrity, missingness, documentation inequity, leakage, and spurious signals.
Distributed Open Justice Oversight
Clinical AI should be tested like a system, not celebrated like a score.
DOJO is an open, community-driven platform for continuous adversarial evaluation of health AI—from the data that shapes a model to the workflow in which people rely on it.
DATA
+
MODEL
+
WORKFLOW
One system. Continuous evidence.
01 / What is DOJO
MIT Critical DataA shared evaluation layer for clinical AI.
Distributed Open Justice Oversight is an open, community-driven platform for adversarial evaluation and remediation of health AI systems.
It helps researchers, clinicians, and evaluation teams examine the full system—not only a model score—before and during clinical use.
01Open
Shared tools, challenge sets, and evidence that can be reviewed and improved.
02Local-first
Designed to work near the data and within institutional security requirements.
03Community-driven
Clinical, technical, and lived expertise shape what gets tested.
02 / Architecture
From data to care, evaluated as one system.
Three connected planes make failures visible at the point where they enter the system.
Model plane
Challenges robustness, calibration, rare cases, shifted inputs, unsafe confidence, and abstention.
Clinical workflow
Tests interoperability, infrastructure, human-AI interaction, and readiness in context.
01Define riskFreeze the intended use and threat model.
02ChallengeRun controlled stress tests across the planes.
03ObserveTrace findings to evidence and provenance.
04AssureReport readiness, limits, and next action.
Evidence objectIdentity · Challenge · Observation · Assurance · Lineage
Every consequential claim stays linked to the model, data, evaluator, environment, and re-test that produced it.
Selected publications from MIT Critical Data / DOJO
Research that informs the DOJO evaluation layer.
Selected work from the MIT Critical Data research program and DOJO team, spanning data quality, model evaluation, bias, safety, and clinical deployment.
01 / DOJO article
Distributed Open Justice Oversight (DOJO): A Community-Driven, Modality-Agnostic Platform for Adversarial Evaluation of Health AI
The primary DOJO paper describes the community-driven evaluation layer, its three planes, and its system-level approach to health AI.
Read the DOJO paper02 / Multi-agent oversight
Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in Clinical Multi-Agent Systems
Examines how shortcuts spread through clinical multi-agent systems and why independent oversight is needed to detect benchmark gaming and socially plausible errors.
03 / Imaging audit
ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs
Introduces paired image-swap audits to test whether report-conditioned medical vision-language models actually respond to images.
04 / Clinical calculation
Towards a Deterministic Math Solver for Clinical Language Models
Tests a program-based solver for clinical calculators, separating tool use, formula access, and execution reliability from model arithmetic.
05Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review
Maps how bias enters AI systems and perpetuates healthcare disparities across data, design, deployment, and evaluation.
Open paper06Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare
Frames clinical AI as a living system that needs continual monitoring, updating, and quality improvement after deployment.
Open paper07The myth of generalisability in clinical research and machine learning in health care
Shows why performance in one setting may not transfer to another and why external validation is central to clinical readiness.
Open paper08Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study
Audits whether a general-purpose medical model reproduces racial and gender biases in healthcare responses.
Open paper09Ethics of large language models in medicine and medical research
Examines the ethical risks of large language models in medicine, including accountability, safety, and evaluation.
Open paper10Leveraging electronic health records for data science: common pitfalls and how to avoid them
Identifies common pitfalls across the electronic health record data lifecycle that can undermine reliable clinical AI.
Open paper11Equity in essence: a call for operationalising fairness in machine learning for healthcare
Calls for operational fairness metrics and practices that make equity measurable in machine learning for healthcare.
Open paper12The reproducibility crisis in the age of digital medicine
Connects reproducibility, transparent reporting, and trustworthy evidence in digital medicine.
Open paper13The “inconvenient truth” about AI in healthcare
Explains why technical accuracy alone is not enough to make AI useful, safe, or equitable in healthcare.
Open paper14An embedded ethics approach for AI development
Advocates embedding ethical reflection throughout AI development rather than treating governance as a final checkpoint.
Open paper03 / Team
Research Team
DOJO is developed by an interdisciplinary team working across clinical research, machine learning, engineering, and public accountability.
Principal Investigator
Multidisciplinary Team Researchers
Isaac Gavilanes
Julie Chasseriaud
Mohammad Shahin
Sebastian Cajas
Yehudhah Rodriguez
Contact
Help build the evaluation layer clinical AI needs.
For collaboration, contribution, or institutional questions, start a conversation through the DOJO repository.
Leo Celi
lceli@mit.edu
Sebastian Cajas
sebasmos@mit.edu
Yehudhah Rodriguez
yehudhah.rodriguez@yachaytech.edu.ec