STRIKE / BENCHMARKS
Evidence
Benchmarks and calibration
Figures below come from the published benchmark report for strike-1. We show gates, suites, latency and limits in plain language.
No benchmark report published yet
We publish a report only when the calibration gates pass (ECE ≤ 0.4, Brier ≤ 0.28, log loss ≤ 1.1) and a maintainer signs it off. Until then this page shows live API numbers when traffic allows, and no model scores.
Strike is in beta. Probabilities are uncalibrated in production until we finish calibration rollout.
Live API (last 24 h)
API requests
157
Last 24 h, ovh-gra
p95 latency (report)
1110 ms
Last 24 h, ovh-gra
Error rate
63.06%
Last 24 h, ovh-gra
Qualitative comparison
| Strike | Generic chat LLM | Hand-written rule | |
|---|---|---|---|
| Typed output | Yes/no, choice, score | Free text | Fixed schema |
| Confidence returned | Per-option probabilities | Not on the wire | None |
| Calibration measured | ECE, Brier, reliability bins | Not published here | Not applicable |
| Latency class | API, sub-second typical | Varies by provider | Local, instant |
| EU hosting | OVH Gravelines (France) | Depends on vendor | Your infrastructure |
| Cost model | Per API call (beta free) | Per token | Engineering time |
Live demo
Same model, same API. Ask it something and read the probabilities back.
Try the playground