The boring way is also a feature for users. Knowing a tool will still work the same in three years matters more than most product comparisons acknowledge.
This is what Graeber called "box-ticking" in Bullshit Jobs: work that exists so the org can prove it's doing something, not because the doing matters. The leaderboard isn't measuring productivity, it's producing proof of AI adoption. Once an exec says "we're an AI-first company," the rest of the org needs to show that's actually happening. Token counts are the easiest thing to put on a dashboard, regardless of whether anyone got anything done.
Rejections are usually conditional on the world at the time: a constraint, a dependency, a workaround that exists today. When those conditions change the rejection is stale but the log still reads "we tried this and it failed." How do you think about surfacing stale entries for revisit? Is it on the agent to spot them on its own or is there a manual deprecation step?
Have you seen the same chain pattern outside finance yet? Wonder whether investment scams are the most conspicuous because the payout per convert is high or whether it's seeded the widest on YouTube specifically.
This is the kind of porting work I always hope for when I see a CUDA-only release. Have you thought about publishing the gather-scatter sparse 3D convolution and SDPA attention swaps as a standalone toolkit or writeup? A lot of folks running models locally on Apple Silicon hit the same wall with flash_attn, nvdiffrast, and custom sparse kernels and end up redoing the same work.