Rules instead of an LLM
The experiment tests understandable constraints, not language generation. An explicit rubric makes every branch inspectable and repeatable, without a model inventing new criteria.
CURIOSITY DEPARTMENT / EXPERIMENT 12
Good AI decisions need better questions.
Three fictional use cases. Five adjustable inputs. See how an explicit decision rubric changes the recommendation, and where human judgment is still needed.
This is a teaching demo, not an AI agent. It uses fixed, visible rules, fictional inputs, and no model calls. Nothing you select is sent to a server or saved. It is not a risk assessment, approval, or advice for a real project.
01 / INPUTS
Set the constraints.
Try changing just one input. A useful decision tool should tell you why the recommendation changed.
02 / EXPLAINABLE OUTPUT
Ready for your scenario.
Select inputs, then evaluate to see a transparent, illustrative recommendation.
Before going further
Identify the people who would validate the assumptions, data boundaries and acceptance criteria.
Decision rubric: sensitive information without oversight → hold; weak evidence → validate first; complex multi-system work → staged pilot; otherwise → constrained prototype. These rules are illustrative, not universal best practice.
04 / BEHIND THE BUILD
Why I designed it this way.
Recommendations, not approvals
Even the most permissive result says to consider a constrained prototype. It never authorizes access, spending, deployment, or use of sensitive information.
What real deployment would require
Validated requirements, documented data boundaries, representative and adversarial test cases, change control, failure handling, monitoring, and an accountable reviewer. None are supplied by this demo.
There are intentional limitations: rules are illustrative, the scenarios are fictional, and the tool has not been validated for any organizational decision. The learning objective is to see how changing a constraint changes a recommendation, not to outsource judgment.
03 / THE ENGINEERING POINT
Confidence isn't evidence.
Explain the recommendation
A label isn't enough. Show which constraint mattered and what evidence is missing.
Keep a human accountable
Automation can surface options, but approval and responsibility need an owner.
Test the edge cases
Change one assumption and look for a sensible outcome, not a prettier answer.