◈
Performance requirements
- ≥95% accuracy retention vs baseline
- Batch-1 mean and p95 latency wins
- ≥1.5× speed or ≥25% power reduction
- p95 bound ≤1.3× mean
4–8× smaller with ≥95% retention. Batch-1 latency mean and p95 wins on declared hardware.
Updated Sep 2025 • 20× compression on roadmap
Our Assurance Spec (ECL-95) requires ≥95% accuracy retention vs baseline, batch-1 mean and p95 latency improvement, a 2-hour stability soak with drift ≤5% and no sustained throttling, and deterministic rebuilds with signed artifacts.
If a deliverable fails on our harness, we rework once or waive the POC fee.
Every metric, hash, and proof point — metrics.csv, latency and stability logs, signed manifest, SBOM, and a one-command repro script. Client packs include models only when permitted.
View a sample verification. No data uploads.
model: sonar_v2.onnx · 3.7× smaller p95_latency: 18.3 ms (−42%) retention: 97.1% stability: 2hr soak PASS sha256: 19a3f2b… ✓ signed manifest
Ranges on specified hardware. Not independent certifications.
| Capability | TensorRT/ONNX | Cloud services | SAYKAI MAP |
|---|---|---|---|
| Compression ratio | 2–4× | 3–5× | 4–8× (20× roadmap) |
| Physics-based failure prediction | No | No | Yes |
| Mission certification | No | No | MAP/ECL-95 |
| Submarine ready | No | No | Designed for |
| Radiation simulation | No | No | Modeled |
| Evidence Pack | Limited | Basic metrics | Complete |
Capability comparisons are illustrative. No third-party endorsements claimed.
Three ICPs for mission-critical compression.
Limited pilot slots. 8–12 week engagement. High-touch service, not self-serve.