# Scientific stages

Lykeion presents research in four stages: Grounding, Hypotheses, Experimentation and Validation. A study can cover a configured part of that sequence. Start by checking its scope and assignee; do not infer that every study will execute all four stages.

## Follow a study through its scope

1. Open the study and inspect its current stage and configured range.
2. Read the agent or Team activity for that stage.
3. Open Review to inspect available checks and explanations.
4. Look for the next observed state, or the concrete reason work is waiting.

A completed hand-in is not sufficient evidence that a stage passed. A study moves to Done only after every configured stage and required review finishes. If scoring or another required service is unavailable, preserve that distinction in your own notes.

## Plan what Validation will check

For a directional hypothesis such as “the filter reduces mean error,” plan a statistical test before Experimentation and have the experiment script record the test statistic and p-value. Validation can reproduce a directional result and still mark it inconclusive when the test is missing.

For a deterministic check, specify a named outcome or numeric range in Hypotheses and have the script record the result. Validation checks that the recorded result reproduces and matches the prediction.

## Check evidence before continuing

Compare the study's outputs with its question and scope. Use [evidence and results](/docs/evidence-and-results/) to inspect files, notebook records and review. The demonstration study in these guides has Grounding hand-ins and is waiting for scoring.

If you need to choose a stage-specific assignee, read [Teams](/docs/teams/) and [studies](/docs/studies/). Check the chosen stage against the question and the deliverable you need.

## Stage walkthrough

### Grounding: Find the foundation

Find and assess sources and trace scientific claims to their evidence. Choose Grounding when you need to map existing literature, screen sources or establish what is known about a question.

- **Inputs:** A research question, boundaries and relevant literature or data.
- **Agents:** An assigned Grounding agent or Team searches and assesses sources.
- **Evidence:** Cited sources, relevance notes and reported access limitations.
- **Gate:** The server checks source existence and retraction; Jev screens sources against the question’s criteria and checks missed papers.
- **Result:** An inspectable foundation for the question; a hand-in alone is not a passed review.

### Hypotheses: Make alternatives explicit

Develop alternative explanations and predictions from the available evidence. Choose Hypotheses when your study needs testable explanations to compare before experimentation.

- **Inputs:** The question and the available grounding evidence.
- **Agents:** An assigned Hypotheses agent or Team develops rival explanations.
- **Evidence:** Proposed hypotheses and the distinctions an experiment could test.
- **Gate:** The server checks recorded citations; Jev checks falsifiability, evidence support and whether the exact prediction was already tested.
- **Result:** Explicit alternatives to investigate, rather than a confirmed explanation.

### Experimentation: Put the question to work

Carry out an analysis or computation and retain the artifacts it produces. Choose Experimentation when you have a concrete task, suitable inputs and an agent with an available machine.

- **Inputs:** The task, required data and a suitable analysis or workflow.
- **Agents:** An Experimentation agent executes the assigned work on a usable machine. Workflow studies use an individual agent.
- **Evidence:** Produced files, computation, notebook records and reported errors.
- **Gate:** The server checks saved outputs and traces reported values to execution. Jev checks how traced values are described.
- **Result:** Inspectable analysis and artifacts, with limitations to carry into review.

### Validation: Challenge the findings

Challenge findings against their supporting evidence and review checks. Choose Validation when a claim or result is ready for independent review.

- **Inputs:** Findings and their supporting evidence and artifacts.
- **Agents:** Independent reviewers test the findings; inspect the assigned roles and actual review record.
- **Evidence:** Review explanations, checks and unresolved objections.
- **Gate:** Without recorded runs, the server checks quoted passages against their sources and Jev judges claims independently of the Team’s verdict. With recorded runs, the server derives hypothesis statuses from assessments and reproduced measurements; upheld objections can lower them.
- **Result:** With recorded runs, an invalid status sends the study back to Experimentation once. Another invalid result waits for its owner. This return is separate from an ordinary item redo. Without recorded runs, claim review follows the item-review outcomes.

After hand-in, a Team’s auditor settles before required review; an individual agent goes straight to review. The server advances automatically within the configured scope when review passes, carrying unclear items when applicable. Ordinary item-review failures trigger one redo, then a block if still failing. Jev’s checks run under the study owner’s own TypeSafe key; without one, the owner approves the stage in Decisions after the server’s checks. A TypeSafe outage leaves the study waiting. At the end of its scope, completion follows required review. An individual agent is restricted to its own stage.
