How the repo handles metrics
The operational UI already stores benchmark history next to run identity and source confidence. That is exactly the kind of architecture a serious research program should expose.
Back to atlas
Evidence
What counts as proof
Benchmark Evidence
The admin dashboard already distinguishes benchmark history, live metrics, and source confidence. The public site should expose that maturity in a cleaner, less operator-centric form.
Key points
Benchmarks are not just numbers; they are provenance objects.
A futuristic site can still be disciplined about what is measured and what is target-state.
The best public story here is one that makes the evaluation standard visible.
Metric focus
TPS + latency
The webapp surfaces both throughput and p95 latency.
Confidence field
High / Medium / Low
Evidence quality is already represented in the system.
Public stance
Grounded claims only
Performance copy should stay linked to source quality.
Public presentation strategy
Public-facing benchmark sections should distinguish reference-envelope metrics from measured project outputs, so the site remains ambitious without becoming slippery.
Call out reference numbers explicitly.
Separate measured run outputs from architectural targets.
Keep a short explanation of confidence and provenance near the charts.