I spent a day teaching my tool to make the problem look smaller

#ai #buildinpublic #observability #devtools

Every "we found $X of waste!" tool has a quiet incentive: the bigger the number, the more impressive the demo. So when I built the report that turns a trace into a waste summary, I deliberately wired in four rules whose only job is to keep that number honest:

1.Count only the re-run, never the original. If an agent does real work once and then redundantly repeats it, only the repeat is waste. Charging the legitimate first run to the "waste" column is the easiest way to double your headline number — and a lie.

2.One row per wasted span (dedupe). When one repeat pairs with several earlier runs, the naive report lists it multiple times and the waste visually balloons. I assert in a test that rows = actual wasted spans.

3."unknown" instead of a guess. No token count captured? The report says unknown. It does not invent a plausible number to fill the cell.

4.The report prints its own frozen parameters (φ, N, embedding model) in the header — so anyone reading it knows exactly which settings produced it.

None of this makes the demo flashier. That's the point. In a category where every competitor is incentivized to inflate, a report that refuses to exaggerate is the differentiator — and it's enforced in tests, not in good intentions.

Code's public: github.com/JEONSEWON/Clew-by-Custos