Update 20 — Saying What You Meant

Metadata

  • Update number: 20
  • Version: 1.9.1 → 1.9.2
  • Theme: the coverage audit gets the one thing it was missing, and both of its lists stop pretending to be complete

The short version

Update 19 shipped the coverage audit: ask any routine is this finding what I meant? and it shows the results sitting furthest from what the routine is for, and papers already in your corpus that the routine never returned.

It compares against an intent — a sentence describing what you are actually looking for. There was nowhere to write one. Every audit therefore fell back to the routine’s keywords, which means comparing a query against the results that query produced. The panel said so, in those words, and the reading was still circular.

There is now a box in the routine editor. What this routine is really looking for — optional, one sentence, and the audit compares against it from the next time you open the panel.

And both audit lists now say how long they really are. They show 25 papers. If there are 312, the list says Showing 25 of 312 instead of stopping quietly at 25 and letting you read that as the whole answer.

Why the field was worth its own release

The audit’s value is entirely in the gap between what you told resmon to search for and what you actually wanted. Keywords are the first; the intent is the second. Without a place to write the second, the audit could only compare the first against itself — and an honest label on a circular reading is still a circular reading.

Writing one takes a sentence. Methods for irregular time series in astronomy is a different statement from time series AND astronomy, and the difference is exactly what the audit is looking for.

Nothing is filled in for you. resmon will not copy your keywords into the field: a routine whose intent was written by resmon would claim you had stated one, and the whole point of the distinction is that the app can tell you which of the two facts it used.

The counts

A list capped at 25 with nothing said about the cap is a small lie of omission — it reads as “there are 25”, which is a number resmon never measured. Both lists now carry their total.

One of the two totals is honest in a further way. The off-target count is exact: it is drawn from the routine’s own results, all of which are ranked. The missed in corpus count comes from a bounded query over the index rather than a scan of it, so when that query fills up with papers still inside the range, resmon knows there are at least that many and does not know how many more. It says at least in that case, rather than printing a precise number it did not measure.

Also in this release

Editing a routine now discards the cached audit for it, so writing an intent and reopening the panel shows the new comparison rather than the previous answer.

For harnesses driving resmon over MCP, create_routine takes an optional intent, and get_routine’s coverage counts are now the totals rather than the length of the page they came from. Contract v1.3, additive; nothing existing changes shape.

What has not changed

The audit still only compares against papers resmon already holds. It has no idea what exists in the literature and is not in your corpus, “missed” still means missed by this routine and found by something else, and that sentence is still printed under the lists rather than in a footnote.