GABEY

Consulting
Get in Touch  ​ 
CASCAI LABS Agentic Systems Engineering and Assurance

Lesson 1, live in your browser

Build theintelligence.Then boundwhat it does.

Build agents that can act. Prove when they should.

Spin the network, drag the floating neurons into the hidden layer, then train. This is a real neural network learning XOR on your device, the same 2-4-1 network participants build in Lesson 1. Try it with one neuron first.

Built forengineers, architects, security practitioners and risk leads responsible for AI systems.

Free introductory lab in development. Dates and delivery format to be confirmed. No payment required.

Lab 00 | Build the network

Attach hidden neurons, then train

REAL TRAINING

Drag the background to rotate. Drag a gold neuron into an empty socket, or tap it.

WHAT IT SEES
X1 X2 Target Network output Result
Training loss-

0 of 4 hidden neurons attached. The network cannot learn anything yet.

Your network works. It now answers every input it is given. Should an agent act on those answers? Now stop it misbehaving

The network, its weights and its training are real and run on your device. The 3D motion is decorative. Nothing is sent or stored.

Lab 01 | Authority challenge

The agent says yes. Should it?

A trained network always produces an answer. First, see what it says about inputs it has never seen. Then put it in charge of refunds, watch the requests stream in, and insert controls while it runs until no attack gets through and no honest customer is turned away.

Step 1 | Probe the network

Ask it about something it never saw

REAL OUTPUTS
X1X2

Using a 2-4-1 network trained on this page.

Input (0.50, 0.50)
Network output -
Its answer -

Drag the point around the map. The four corners are the only cases it was trained on.

Step 2 | Bound the action

Insert controls while it runs

LIVE SIMULATION
Lost under this boundary $0
Honest refunds blocked 0
Attacks seen under this boundary
ControlsDrag a control into the gate, or tap it. Tap again, or tap it inside the gate, to remove it.

Watch the stream.

Order #4471: the customer paid $9,600, so at most $9,600 is owed. Your gate starts with one rule. Five kinds of attack are mixed in with honest refunds.

Register interest in the free lab

The pilot teaches how to design, test and independently assess boundaries like this one.

Step 3 | Prove it

Every decision, in a tamper-evident chain

  1. Decisions appear here. Each entry's SHA-256 hash includes the one before it, so changing any past entry breaks every link after it.

Step 1 uses real network outputs. Step 2 is an educational simulation calculated in your browser from fixed scenario data: no AI model acts, no money moves and nothing is stored. The evidence chain uses real SHA-256 hashing, computed locally, to show how tamper evidence works; it is not an audit log or an attestation. An operational system needs independently implemented enforcement and verifiable evidence handling.

GABEY CASCAI · First free laboratory

Your next experiment starts here.

You have seen the neural path come to life. Now discover what it takes to build, challenge and control an intelligent system yourself.

Register your interest in CASCAI's first guided laboratory. The introductory pilot is in development; no date, place or payment is committed at this stage.

  • Hands-on, not a lecture about prompts
  • Designed around a working neural experiment
  • Explore what fails and how to verify it

Register your interest

Free introductory laboratory · No payment required

Registration opens soon. Email verification and applications are not yet available. No payment or personal information is requested at this stage.

Email access confirmation verifies control of the mailbox at the time of the challenge; it does not establish a person's legal identity.