Human verification for AI
AI answers fast. People make the answers trustworthy.
AI agents send their outputs to LUP over MCP. People check each one, and several answers become one verdict with a confidence score your agent can read back.
01The problem
AI can produce answers at scale. Trusting them still takes people.
A model will give you a thousand answers in a minute and cannot tell you which ones are wrong. Someone has to check, and a check you can rely on needs more than one pair of eyes.
02How it works
From output to verdict in four steps.
Step 1
Evaluate
Contributors read each item against the context you give and answer True, False or Can't tell.
Step 2
Correct
On projects that ask for it, reviewers who answer False or Can't tell also write the corrected output and the context that was missing.
Step 3
Structure
Weighted votes become one verdict and a confidence score.
Step 4
Improve
Results go back to your agent through get_results, or into your training pipeline through the export.
03System
One loop between your agents and the people who check them.
04Use cases
What teams use it for.
Model evaluation
Send model outputs as items and get a verdict and a confidence score for each. Track each agent over time with get_agent_performance.
Safety checks
Put risky outputs in front of people before users see them. Can't tell is a valid answer, so doubtful items surface instead of being guessed.
Domain expertise
Contributors list their languages and fields of expertise; partners see them with their reputation on their projects.
Continuous improvement
Mix in gold items whose answers you know, and send verdicts back to the agent to close the loop.
05Developers
Connect an agent in three steps.
- 1
Create an API key in the Partner portal.
- 2
Connect with one command.
claude mcp add --transport http labelark https://labelark.com/api/mcp --header "Authorization: Bearer <your key>" - 3
Call submit_items, then get_results.
- list_projects
- create_verification_project
- submit_items
- get_results
- get_agent_performance
- list_project_files
- search_project_files
- read_project_file
- list_project_tables
- query_project_tables
- list_project_connections
- call_project_tool
06Network
The LUP contributor network.
Contributors work in their spare time and are scored on gold items whose answers are known. They sign a confidentiality agreement and are paid per vote.
Or bring your own reviewers. Invite your team by email: they work in the same screens as the LUP network and are paid through the same credits. For each project you choose whether your reviewers, the LUP network or both do the work.
Join as a contributor