Jorge Polanco ยท the question box
The question box on the home page is a retrieval service over the Markdown in my public repositories, measured against a checked-in question set on every push and against a judge on demand. Everything on this page is computed by the service or by a command in its repository; this page only renders it. How it was built.
The service is not answering right now, so the live numbers are not shown.
Counted from the request log the service already writes, which never contains a question. Latencies are bucket upper bounds, so a P95 is an estimate from above.
Each retrieval run checks that the right document is in the top six for every answerable question and that refusal questions stay below the gate. Each generation run asks the model every question and records whether the answer cites, whether a refusal held, and how faithful the answer is under Claude Opus 5 as judge.
| build | when | questions | recall at 6 | MRR | refusals past the gate | reranker |
|---|
| build | when | questions | cited answers | refusals answered | phrases | faithfulness | relevancy | cache reads | reranker |
|---|
The retrieval eval runs in CI on every push to main and fails the pipeline on a regression against a stored baseline. The generation eval costs money and runs on demand; its results are committed, and a small command folds them with the decision records into one document that is baked into the service image. The live numbers come from counters the service keeps in memory and rolls up per day on disk. None of this is computed in the browser. The README has the details.