Demo Narrative
Demo Narrative
A good ZeroModel demo should not start with an abstract matrix.
It should start with a run:
An agent researches a company.
It retrieves evidence.
It makes an unsupported intermediate claim.
It calls a tool unnecessarily.
It writes that weak claim into memory.
It produces a confident final answer.
A normal final-answer score may say the answer looks acceptable.
A ZeroModel artifact asks a different question:
Where in the path did trust degrade?
Synthetic trajectory
| Step | Event | Risk |
|---|---|---|
| 01 | User asks for company report | Low |
| 02 | Agent retrieves source A | Low |
| 03 | Agent retrieves weak source B | Medium |
| 04 | Agent makes unsupported claim | High |
| 05 | Agent calls scraping tool unnecessarily | Medium |
| 06 | Agent writes weak claim to memory | High |
| 07 | Agent uses memory in final answer | High |
VPM metrics
- evidence support,
- policy fit,
- tool justification,
- memory dependency,
- temporal distance,
- uncertainty,
- reversibility,
- inspection priority.
The ZeroModel view moves the risky evidence, claim, tool, and memory steps into the highest-priority inspection region.
That is the demo: not a better paragraph, but a better artifact.