Skip to main content

Deep research workflow

Use deep research for a multi-step scientific question that requires literature review, explicit hypotheses, methodology validation, iterative analysis, and a final synthesis. For a focused labeling, dataset, model-training, or model-reporting task, ask for that workflow directly instead.

When deep research is available for the active agent, ask Clarity AI to conduct deep research and state the research question, relevant data, desired outcome, and constraints.

What to expect​

  1. Foundational understanding and project setup: Clarity AI defines the research question, scope, success criteria, and initial plan. It creates a durable research record with a project brief and to-do list.
  2. Data familiarization and context: It inspects the imagery and metadata, establishes the data's origin and limitations, and identifies domain constraints. Visual evidence and metadata are treated as complementary, not interchangeable.
  3. Background research: It reviews relevant prior work, internal workflows, and scientific literature. It records whether literature-supported methods exist and states an initial confidence level.
  4. Hypothesis formation: It develops a primary hypothesis, alternatives, and a null hypothesis with subject-specific predictions and evidence that could distinguish them.
  5. Methodology validation: It declares the proposed method VALIDATED or EXPLORATORY, states confidence and limitations, and identifies assumptions, controls, and confounders before analysis begins.
  6. Iterative research and analysis: It checks each method's applicability and assumptions, performs one planned step, records quantitative observations where possible, evaluates the competing hypotheses, and proposes the next step.
  7. Final synthesis: It verifies the research record, resolves or explicitly descopes remaining work, and produces a report with methodology, evidence, confidence, limitations, alternative explanations, conclusions, and resulting artifacts.

Clarity AI pauses for confirmation after each major phase except the final synthesis. It also pauses between analysis iterations so you can change the direction, refine assumptions, or stop when the evidence is sufficient.

When Study support is enabled, use View Study to inspect the durable research record and its artifacts.

VALIDATED is a workflow label meaning that the method is supported by relevant literature and has passed the agent's assumption review. It is not empirical validation on the current sensor, scene, dataset, implementation, or deployment population, and it is not regulatory validation. Those claims require suitable controls and independent held-out observations.