I've seen similar on a 100mbit line that was actually a faster line throttled to that speed, the effect was particularly upstream rather than down. Essentially some content (or more likely content from some locations) is allowed to bypass the throttle.

This might be deliberate but not malicious: through Netflix's Open Connect program ISPs can host a local cache of Netflix content to reduce their peering costs while still offering full service to many concurrent users. If the main throttling effect you are experiencing is applied topologically close to their peering lines such that it would apply to the Open Connect equipment too, rather that the Open Connect kit setting between the throttle point and the peering lines, then to make it useful in the case of new or otherwise not recently accessed data the traffic shaping rules would need to let Netflix traffic (including your speedtests) through relatively unhindered. The path between the Open Connect equipment and your line is unlikely to need throttling because that will be an internal matter with very different cost dynamics than external peering.