#Browser Harness

Browser Harness lets you evaluate models that are only accessible through a web UI — like ChatGPT, Gemini, or any custom chat interface. It captures the model's responses from the browser and imports them into eval_752 as a normal scored run.

When to use: The model you want to test has no API. You only have access through a website.

When NOT to use: You already have an API key and provider set up. Use the normal Runs workflow instead — it's faster and more reliable.

#How it works

1. You select a dataset and target adapter in eval_752
2. eval_752 generates a self-contained JavaScript script
3. You run that script in the target chat page's browser console
4. The script automates: sends prompts one at a time → waits for responses → records everything
5. When done, the script downloads a .eval752.zip file
6. You import that file back into eval_752
7. The import creates a normal run that goes through standard scoring and comparison

Nothing leaves the browser. The script runs entirely client-side and doesn't send data to any third party.

#What gets exported (and what doesn't)

The generated script contains only what's needed to run prompts:

This means the target chat page never sees your answers or scoring logic.

#Step-by-step workflow

#1. Configure the export

Open Browser Harness and set up:

• Dataset — pick the dataset to evaluate
• Section filter and Max items — optionally limit scope
• Target adapter — ChatGPT, Gemini, or Custom
• Judge provider/model — if the dataset uses LLM-as-judge scoring

Export is blocked when:

• No items match the filter
• Items contain embedded assets (text-only for now)
• Judge scoring is required but no judge is configured

#2. Choose a target adapter

ChatGPT and Gemini have built-in selector recipes and pacing defaults. You can still override the target origin, display name, and timing.

Custom requires you to provide CSS selectors for:

• New chat button
• Composer text area
• Send button
• Assistant turn container
• Response text

Optional selectors: model label, busy indicator, stop button.

#3. Generate and run the script

The page produces a raw script and a bookmarklet wrapper.

At runtime, the script:

• Shows a lightweight overlay on the target page
• Clicks "new chat" before each item
• Sends one prompt at a time
• Waits for the response to settle
• Records timing, model label, origin, and errors
• Downloads browser-harness-export.eval752.zip

#4. Import the capture

Back in eval_752, upload the .eval752.zip through the import flow. The import creates a completed run with:

• triggered_by = browser_harness
• Browser target metadata on the provider record
• All captured responses stored as normal run items
• Scoring queued through the standard worker

After import, the run appears in Runs and (once scored) in Comparison — just like any API-based run.

#Troubleshooting