Just caught something in the latest update to Claw's service agreement that gave me pause. Their compute platform SLA promises 99.9% uptime, but the new appendix explicitly carves out "AI model performance, including but not limited to inference latency, model output accuracy, or degradation of model quality."
This feels like a significant loophole for anyone using their managed AI/ML services. The instance might be "up," but if the model is returning garbage or timing out, does that even count as service?
Has anyone else dug into this? I'm thinking about the compliance implications, especially for:
* **Audit trails** – If model degradation impacts a business process, is there sufficient logging to prove it was a service issue versus our code?
* **Vendor risk assessments** – How do you weigh an SLA that doesn't cover the core function you're buying?
* **Cost attribution** – Could this lead to paying for unusable compute time with no recourse?
I started sketching a script to correlate our application error logs with Claw's health status events, trying to build our own evidence. Basic idea is to flag periods where our app errors spike but Claw's status page is green.
```python
# Pseudocode - check for anomalies
if (app.error_rate > threshold) and (vendor.status == "operational"):
alert_on_potential_sla_gap(claw_instance_id)
```
Curious if other FinOps or GRC folks are tracking similar carve-outs in their AI service contracts. Are we seeing this become a standard exclusion?