Leaderboard scores tell you what a model can do in a clean room. They do not tell you what it costs to run your workload, how it behaves on the long tail of inputs, or whether a cheaper model plus a pre-filter beats the expensive one.
On Horizon, the intent-classification cascade started as “send everything to a model”. The evaluation made the case for the opposite: a regex pre-filter in front of a mid-size classifier. Most trivial intents never reached a model at all. The long tail still did, and the classifier’s accuracy there was close enough that the difference never mattered to the product.
That is the useful part of evaluation. It rarely says “use the better model”. It says where reasoning earns its cost and where it is an expensive if-statement.
What I measure
- Accuracy on the cases that actually appear in production, not the balanced benchmark set.
- P50 and P95 latency for the whole path, including the pre-filter and the retries.
- Cost per successful task, not cost per token.
- Failure mode: when the model is wrong, does the system degrade gracefully or confidently invent?
The last one is the reason I care about agent architecture more than model choice. A routing layer, a deterministic pre-filter and a boring fallback will beat a better model with no boundary around it.
The rule I keep using
Algorithms → detection → structured context → AI reasoning, in that order. Push as much of the decision into deterministic code as the problem allows. Then give the model the smallest context that still explains what is happening, and measure whether the answer is worth the price.
If the evaluation does not change the architecture, I probably measured the wrong thing.