Trace health
Is the system healthy right now? Scope filters on the right — open Traces to search and slice the list.
Admin: scoring smoke job
Runs a tiny deterministic scoring task on the server (async job, polled from the browser). Use the same secret as ADMIN_API_KEY on the API. Only share this with operators you trust; it grants access to admin routes.
Data pipeline— fetch latency from eval traces, run tests from the dashboard.
Reliability over time
Drift and regressions in the selected window.
Risk distribution
Count of traces by hallucination risk band.
Success vs failure
GOOD+RISKY vs FAIL classification.
No high-severity anomalies in the last 5 minutes.
By agent
Model comparison
Same evaluation slice — compare grounding, risk, latency, and consistency. One color per model in charts. Adjust filters in the panel on the right.
gpt-4o-mini (score by balanced)
Weighted by the metric selector in the filter panel.
gpt-4o-mini (1293 ms)
Lowest average latency in this slice.
gpt-4o-mini (risk 0.54)
Lowest mean hallucination risk.
gpt-4o (σ rel 0.015)
Lowest reliability σ (needs 2+ traces).
Reliability trend by model
Each line is one model’s reliability over time.
Quality vs latency
Upper-left is best — high reliability, low latency.
| Model | n | Avg G | Avg rel | Avg risk | Avg ms | σ rel | Tradeoff |
|---|---|---|---|---|---|---|---|
| claude-sonnet-4-6 | 15 | 0.325 | 0.388 | 0.675 | 3311 | 0.0466 | 0.117 |
| gpt-4o-mini | 10 | 0.462 | 0.543 | 0.538 | 1293 | 0.1929 | 0.420 |
| claude-opus-4-6 | 5 | 0.317 | 0.380 | 0.683 | 3346 | 0.0721 | 0.114 |
| gpt-5.4 | 5 | 0.355 | 0.383 | 0.645 | 1351 | 0.2136 | 0.284 |
| gpt-4o | 5 | 0.191 | 0.286 | 0.809 | 1705 | 0.0146 | 0.168 |
Advanced: side-by-side trace diff
Open two trace detail pages in separate tabs and compare prompt, evidence, grounding, and verdict. A future release can link traces that share the same prompt hash or experiment case id.