What to look for
Authorized requests should reach the local executor. Withdrawing either consent or READ permission should stop the next attempt. Restoring both should let the next attempt proceed.
Relationship is the Technology
When permission changes,
the next action should change.
An assistant knows your meeting preference. Follow what happens when you withdraw consent, renew it, and change the assistant's access.
This reference application runs the Python SDK with synthetic data and a local suggestion function. No AI model is called and no meeting is booked. Each run creates a fresh relationship, executes five steps, and checks its saved history.
Run the application locally
The public page explains the experiment. Executable results require the local SDK adapter.
python -m pip install -e '.[dev]' python playground/adapter.py
Then open http://127.0.0.1:8765/evaluate.html. Start from the Quickstart if you do not have a checkout.
For the same report without a browser:
python playground/reference_experience.py --output reference-results.json
Checking for the local SDK adapter…
Same preference. Changing permission.
The cached-payload comparator deliberately reuses an initially authorized request without checking again. TRIA evaluates the current declared requirements before handing the request to the same local suggestion function.
| Step | Cached payload Intentionally ungated | Through TRIA | Inspect |
|---|
Run identity, checks, and measurement scope
What this establishes
A reproducible example of encoded governance and persistence. The cached comparator represents one failure pattern. Model reasoning quality and benefits in people's lives need separate evaluation.
Follow the evidence and reproduction path →