ScienceDiscovery
中文 GitHub

Trace a result

A result is only as useful as your ability to check it. In this lesson you start from the figure and report of Analyze a dataset, follow them back to the code, environment and input that produced them, act on the Reviewer's findings to get a second version, and revoke the permission you granted.

The Review and provenance design page explains what is recorded and why.

1. Open the provenance of the figure

Open the Session and, in the Artifacts list, open treatment-boxplot.png. Switch from Preview to Provenance. You should see:

These are records of what ran, not a description written by the model.

2. Check where a workspace file came from

In the Workspace panel, under Files, select the information icon next to plant-growth.csv, then next to treatment-boxplot.png. The first shows User upload as its origin; the second shows the tool and execution that wrote it, the Linked Artifacts and its Revision history.

3. Read the Reviewer's findings

The Reviewer and automatic review are on by default for new installations. About a minute after growth-report.md was registered, a Reviewer card appeared in the conversation. Open it. Typical findings at the quick level name numbers in the report that are not linked to recorded evidence, with a suggestion to add that link.

Findings do not change the Artifact. They are prompts for you and the Agent; the Reviewer can also be wrong. To turn automatic review off for a Session, use Automatic review next to the approval mode under the message box.

4. Ask for a revision

Send:

The Reviewer card lists numbers in growth-report.md that are not linked to recorded evidence. Add a short "Where the numbers come from" section that names, for each table and test, the script and results file that produced it, and link the flagged values to those records. Do not rerun the analysis or change any number. Save the file and declare growth-report.md again so it becomes version 2 of the same artifact.

What you should see: the Agent reads the report and earlier records, edits the file, and declares it again. The Artifact's version selector now lists two versions, and its preview shows v2 of 2; the conversation mentions the new version. A new automatic review may follow, and it may still have findings.

5. Compare the versions

Open growth-report.md. In the version selector at the top of the Artifact, choose version 2; the preview shows v2 of 2 and the new "Where the numbers come from" section. Then choose version 1 in the same selector and look at the same tables. A Markdown report has no side-by-side diff view, so compare the two previews yourself: version 2 adds the new section and links, and every value in the tables and test results should be unchanged, because the prompt asked for links, not new calculations. Both versions remain stored; version 1 still shows the original text and its own provenance.

6. Revoke the permission you granted

In the dataset lesson you chose Allow same type for code execution, which left a standing grant. Open System configuration → Permissions. Under Standing permission grants, each grant shows its action, such as Run local code, and the resource it covers, such as workspace-code. Select Revoke on the code-execution grant. Each Session where you chose Allow same type has its own grants, including one per connector tool from the earlier lessons; revoke those you no longer need. Revoking removes a grant from the list; executions that already ran keep their records and provenance.

7. Check your result

Where to go next

You have completed the learning path. Apply it to a real research question with the domain guides, or read ScienceMemory and Reviewer to follow results through the memory graph when it is enabled.