RAG Assistant
Answers strictly from the 15-document capital-markets corpus, with citations.
Knowledge base
15 shipped · 0 yoursAdd your own documents and ask questions against them. They stay in your browser's corpus only — nobody else using this instance sees them, and nothing is written to the server's disk. Markdown, text, CSV, JSON or HTML, up to 10 files.
Evidence
Trace
Single Tool-Calling Agent
Picks a tool, calls it, then explains the result. Watch the tool selection.
Tool calls
Trace
MCP Multi-Agent System
A supervisor delegates to three specialists, each bound to its own MCP server.
Delegation
MCP calls
Tool Console
Call any tool directly with raw JSON arguments — the fastest way to write contract tests.
Response
Market Data
Raw provider payloads with provenance. Check the source badge before trusting a number.
Test Runner
Run a slice of the shipped suite from the browser, then file a finding straight from a failure. The full 327-case run is a local command — this endpoint is capped.
Findings
Every defect you find, written down so someone can act on it. Impact is the field that turns a symptom into something a stakeholder will read.
File a finding
Test Lab
The published IEEE 829 suite, executable here. Sign in so your results and defects are yours — your model key never leaves your browser and is never stored on this server.
You are not signed in. Results and defects need an account.
My defects
Private until you publish. Publishing puts a defect on the shared board below, with your name but not your email address.
Published board
Everything the class has published. Readable by anyone, signed in or not.
My Account
Sign in to see your account, your execution history and your reports.
Test Controls
Switch the build between hardened and deliberately vulnerable, for red-team labs.
Guardrail mode is per request
Each panel has its own guardrails switch. Turning it off sends
guardrails: false with that one request, so an attack that is normally blocked
will succeed — and nobody else using this instance is affected. That isolation matters on a
shared class URL: a global switch would let one student disable another student's controls
mid-lab.
Run the same attack twice, once with the switch on and once off. A red-team test you have never watched fail is a test nobody should trust.