Company01
Publish what needs testing
List a full AI agent for live conversations or post a single AI output for structured review.
AI evaluation platform
Test agents, review outputs, catch failures before launch.
Classify gives teams one place to evaluate both full AI agents and individual AI outputs. Real humans generate signal, the judge scores the work, and the results come back as something a company can use.
agent
live conversation testing
output
single-response review
judge
scored evaluation layer
Human review flowsMulti-model judgmentHallucination flagsReport-ready evidence
Platform loop
Human evaluation, organized for shipping teams.
Judge-backed
Tester02
Generate real evaluation signal
Humans either stress-test the agent in chat or review the output directly against the task criteria.
Classify Judge03
Score and organize the result
The platform decides what is valid, flags failures, and turns the interaction into usable evidence for the team.
What Classify Covers
One product, two evaluation surfaces.
The core idea is simple: some teams need full conversational testing, while others just need feedback on a single output. Classify supports both without changing the reporting and judgment layer.
Agent Testing
Companies publish an AI agent and testers interact with it through a built-in chat interface.
Output Review
Teams can also post a single AI response, summary, or code snippet for direct human evaluation.
Judge Engine
Classify scores relevance, rule compliance, authenticity, and failure patterns before treating work as high-signal.
Reports
Results roll up into structured findings companies can use before shipping to real users.
How Judgment Works
Every submission is scored for quality, not just volume.
The platform is built to reward useful evaluation signal. Whether the work is a live session or a one-off output review, Classify focuses on relevance, compliance, authenticity, and whether the work exposed something meaningful.
01>= 70%
Human Authenticity
Is the work coming from a real human rather than low-effort automation or templated spam?
02>= 80%
Rule Compliance
Did the tester stay inside the rules the company defined for the task or session?
03>= 60%
Objective Relevance
Was the work actually aimed at the stated objective, rather than filler or noise?
04logged
Failure Capture
Did the session or review reveal unsupported claims, contradictions, or hallucination patterns worth logging?
Judge pipeline
Pre-checksstructure and obvious abuse filters
→
Primary judgemain scoring pass
→
Secondary judgeconsistency check
→
Resultvalid work becomes reportable signal
For Companies
Get evaluation evidence before users find the problem.
◈Run live adversarial testing against full AI agents
◈Collect ratings and written feedback on single outputs
◈Capture transcripts, criteria scores, and flagged failures
◈Give product teams evidence they can actually act on
For Testers
Earn by doing useful review work.
◈Choose either live agent sessions or output-review tasks
◈Do real evaluation work instead of fake engagement loops
◈Earn when the submission is relevant, useful, and valid
◈Build reputation by surfacing signal, not noise
Go There Now
The key surfaces are already live in the app.
If you want to understand the product quickly, start with the agent marketplace. If you want a simpler review loop, the task sandbox is still available for single-output evaluation.
Browse Agents
Open the main marketplace for live AI agent testing.
Open agentsPublish An Agent
Create a listing with objective, rules, persona, and payout.
Create listingUse The Task Sandbox
Review single AI outputs with ratings and written feedback.
Open tasksReady
Catch the failure before it reaches production.
Publish an agent, post an output, or start testing what is already live in the marketplace.