Skip to content
Evidence Lab
English
EnglishРусский
Esc
↑↓navigate↵open⌘Jpreview
On this page

Evidence Lab

Set up a personal research environment and learn to use it on a real task.

Evidence Lab helps a researcher turn Codex or Claude Code into a controlled working environment for a real research process. This guide takes you from a bounded first case to a result you can inspect, repeat, and recover without relying on an old chat.

Result

You finish with five things you can use independently:

  • a dedicated project folder;
  • written project instructions and data boundaries;
  • only the skills, plugins, and connections your scenario needs;
  • one guided and one independent run with acceptance evidence;
  • a handover package with versions, ownership, support, and recovery notes.

The environment can collect, transform, compare, and package information. It does not decide the scientific question, admissible evidence, interpretation, or final conclusion for you.

Actions

  1. Read what Evidence Lab does and confirm that your goal fits the service boundary.
  2. Prepare a safe practice case whose correct result you can recognise.
  3. Complete the ChatGPT, Work, and Codex onboarding route so you can install the app, verify access, choose the right surface, and control a Codex task.
  4. Choose Codex or Claude Code for the real workflow, organise the workspace, and write the project rules.
  5. Complete the first working task using explicit inputs, output format, and acceptance criteria.
  6. Package the evidence, verify the output, then repeat the workflow without step-by-step consultant help.
  7. Preserve the accepted configuration and recovery procedure in the handover package.

Check

Without opening the original chat, you can show the task brief, source inputs, agent actions, final artifact, verification evidence, acceptance decision, and the instructions needed to repeat the work.

If it fails

Do not add more tools or write a longer prompt by default. Return to the earliest missing artifact, narrow the task to one output, and identify whether the failure came from the inputs, access, instructions, tool execution, or verification method.

Next: Engagement overview.