Demo on fictional data
Model comparison scorecard
This is a demo on fictional data that shows three unnamed configurations scored on the same 25 made-up test items, and how a change that scores worse than the baseline on any measure is blocked from release.
Requests since you opened this page: 0
Works offline: not installed
The browser policy on this page forbids network requests (connect-src 'none'). The count is here so you can see it hold. How the demos work.
Scores against the baseline
Configurations A, B and C are unnamed on purpose: no model or vendor is behind them, and the scores are made up. The release rule is the one FoxGrade uses: a change that scores worse than the baseline on any measure does not ship.
| Configuration | Exact match | Reviewer edits per item | Cost per item | Release |
|---|
The items behind the scores
Each item is a fictional drafting task with a known good answer. Real test sets come from your own best past examples.
| Item | A exact | B exact | C exact | A edits | B edits | C edits |
|---|
What this shows, and what it does not
What this shows
- Exact match, reviewer edits and cost per item, per configuration.
- A release block whenever a measure regresses against the chosen baseline.
- The items behind each score.
What it does not show
- The scores are made up. No model or vendor is named, so this is not a benchmark.
- Real test sets come from your own best past examples.
- No number here is a result a client achieved.
How it would run for you
This is the shape of a Architecture Sprint: the agent drafts, a named person approves, the log stays with you.
Related services: Architecture Sprint, Production Readiness Review.