The Notebook · the week
The week of 26 October 2026
7 ideas · 26 October 2026 to 1 November 2026
-
Monday · 26 October 2026
the answers survived the squeeze and the reasons did not
We quantise to cut serving cost and check it with an accuracy benchmark. The score holds, so the change ships. A paper from September makes me doubt that check.
-
Tuesday · 27 October 2026
the retry was safe and the side effect was not
An agent tool call timed out and the harness retried it. The first call had already succeeded, but the caller never saw the response. We ended up with two of the same record.
-
Wednesday · 28 October 2026
the retriever was fine and the pdf was not
Most retrieval failures I chase are not retrieval failures. They begin in the parser. A table in a scanned PDF becomes a flat run of numbers. Column headers detach from the cells below them.
-
Thursday · 29 October 2026
the questions looked alike and the answers did not
I keep finding semantic caches tuned against the wrong metric. The mechanism is plain enough and the economics are real. You embed the incoming query and find the nearest cached one.
-
Friday · 30 October 2026
the shortlist decided the score before the reranker did
I spent a week tuning a reranker that could not have helped. The first stage returned a shortlist of candidate documents. The reranker only reordered that shortlist.
-
Saturday · 31 October 2026
the numbers were right and the units were not
Extraction pipelines get graded on the number and not the dimension. A model reads a table and returns 4.5 for capacity. The source said megawatts. The schema expected kilowatts.
-
Sunday · 1 November 2026
the longer list looked better because guessing got easier
I have been reading about tool retrieval for agents this week. Most systems show the model a fixed number of candidate tools. We tuned that number by coverage, meaning how often the right tool appeared.