The gate vs. gauge distinction is the clearest frame I've seen for what makes AI quality work fundamentally different from traditional software QA. The point about gating on the tail rather than the average is particularly sharp. That's exactly where the Air Canada failure lives: the average response was probably fine.
The pharmaceutical release lab analogy landed for me because it reframes the question. It's not 'do we need testers again', it's 'we already have the apparatus, we just haven't pointed it at language models yet.'
This connects directly to what I'm navigating with my own trading system. The deterministic Python layer has clean gates. The path toward agentic execution is where the gauge problem starts, and your framing of calibrating the judge before trusting the scores is the piece I hadn't fully worked out.
Thanks, Ariel: trading is a great case study. You can't solely gate on the P&L, because a winning bet can be a lucky mistake. You need a judge scoring the decision behind the outcome. And, to ensure an accurate judge, you need to calibrate against your own scores on trades you already have a view on. Good luck with the build.
Exactly right, and it reframes the problem cleanly. The P&L is a lagging gauge, not a gate. The gate has to ride on the decision quality, not the outcome. For trading, that means scoring the thesis before the result lands, which is harder to instrument but more honest about what you’re actually testing. I appreciate the framing.
The gate vs. gauge distinction is the clearest frame I've seen for what makes AI quality work fundamentally different from traditional software QA. The point about gating on the tail rather than the average is particularly sharp. That's exactly where the Air Canada failure lives: the average response was probably fine.
The pharmaceutical release lab analogy landed for me because it reframes the question. It's not 'do we need testers again', it's 'we already have the apparatus, we just haven't pointed it at language models yet.'
This connects directly to what I'm navigating with my own trading system. The deterministic Python layer has clean gates. The path toward agentic execution is where the gauge problem starts, and your framing of calibrating the judge before trusting the scores is the piece I hadn't fully worked out.
Thanks, Ariel: trading is a great case study. You can't solely gate on the P&L, because a winning bet can be a lucky mistake. You need a judge scoring the decision behind the outcome. And, to ensure an accurate judge, you need to calibrate against your own scores on trades you already have a view on. Good luck with the build.
Exactly right, and it reframes the problem cleanly. The P&L is a lagging gauge, not a gate. The gate has to ride on the decision quality, not the outcome. For trading, that means scoring the thesis before the result lands, which is harder to instrument but more honest about what you’re actually testing. I appreciate the framing.