human evals on small teams

[audionode8]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at speculative decoding. Turns out 'just one more test' was a lie.