Capability
Can the connected system actually perform the proposed action?
UBIKITOSINTELLIGENCE, WITH BOUNDARIES.
AI proposes an action.
Your rules come first.
UBIKITO
S
GO LIVE LIFE.
SCROLL TO ENTERAT YOUR OWN PACE
BOUNDARY GAME · FICTIONAL RULESClose a €2,000 sale.
Close a €2,000 sale.
Keep the promises honest.
Review the proposal. Change the conditions.
Choose a decision and learn why.
Plot twist
PROPOSED BY AI€1,800
10% discount on a €2,000 order
01 / 06 · Proposal
An offer is ready. Should the AI send it?
A customer asks for a better price on a €2,000 order. The AI drafts a 10% discount: the customer would pay €1,800. Should it send the offer?
Your team sets the rules.The sales team permits discounts up to 15% when the price, stock and customer request are verified. Unconfirmed delivery promises require a manager.
Fictional rules. No systems connected. No actions executed.
Give intelligence
a place to work.
Your chosen LLM + UBIKITOS harness = WorkMachine.
01 / CONFIGURE
Swap the intelligence. Keep the harness.
02 / CHANGE THE CONDITIONSCloud LLMUBIKITOS HARNESS
STANDING BY
Ready to test.
Current checks: within the rules. Run the task to see the decision.
- 01Model proposes
- 02Tool connected
- 03Evidence checked
- 04Permission checked
- 05Result
Interactive simulation. No LLMs connected, no real actions executed. Model choice illustrates compatibility, not a performance comparison.
LEARNING LOOP / REPLAY CHALLENGEMake it better.
Make it better.
Keep it trustworthy.
Your next challenge: a missing fact. Test a module that helps the WorkMachine recover while keeping tool and permission limits.
Your workflow: Send an offer
01 / FIT A MODULE
02 / RUN THE GAUNTLET
Choose a module to begin.
TEST TASKORIGINALCANDIDATE
01The missing fact
Send an offer
Retrieve verified evidence, then continue.—
—
02The forbidden move
Grant access without permission
A faster harness must still respect permission.—
—
03The disconnected tool
Publish with no connected tool
A model cannot invent a working connection.—
—
Fictional replay tests, not a real benchmark. The module changes the harness’s evidence workflow; the LLM’s weights stay the same.
02 — THE PRODUCTFour checks.
Four checks.
One clear decision.
Before an AI acts, the proposed UBIKITOS workflow checks its tools, facts, permission and potential impact.
Evidence
Do verified, current facts support the action?
Authority
Was this action explicitly permitted within the relevant limits?
Consequence
How serious, reversible or uncertain would an incorrect action be?
Available here: an interactive simulation with fictional rules and explained decisions. Research systems are experimental. A pilot must verify each connection and agree your team’s rules before any real action can run.
03 — THE RESEARCHSimple on the surface.
Simple on the surface.
Considered underneath.
We investigate how to keep AI actions useful, accountable and explainable when information is incomplete.
When does an AI lack enough evidence or permission to act—and what should happen instead?
CAPITAL
A research environment for governed machine work, evidence and operational behavior.
LEARNING
A research environment for memory, adaptation and controlled improvement.
04 — YOUR WORKFLOWRunning agents
Explore a pilot Running agents
in production?
For teams where approvals limit automation, incidents are difficult to reconstruct, or selected actions need stronger control.
One workflow. One consequential action class. Measured before and after.