The confidence floor now scales with your corpus
· Richard Wollyce
Vesta marks an answer as a guess when nothing in your library matches well enough, and refuses when nothing matches at all. The first of those is a comparison: the best file's keyword score against a floor. Until today the floor was one number, 17.5, measured on one fleet. This post is about why one number cannot be right for every size of library, what replaced it, and what each piece of the arithmetic means, so you can argue with the threshold instead of trusting it.
The pieces, named
| Term | What it is |
|---|---|
| entry | A note the router can find: a markdown file with a Search for: line |
| N | How many entries your fleet has |
| key | One term on a note's Search for: line |
| df | Document frequency: how many entries carry a given key |
| idf | Inverse document frequency, ln(1 + N / (1 + df)). The weight of a key. A key found in one note out of a thousand weighs 6.2; a key on every note weighs almost nothing. This is the standard arithmetic for "how much does this word tell me about which file to open" |
| idf_unique(N) | idf for a key that exactly one note carries: ln(1 + N/2). The heaviest a single key can be at that size. 0.41 with one entry, 1.10 with four, 1.87 with eleven, 4.74 with 226 |
| keyword score | Each key the question matches is worth 6 × idf, and the score is the sum, plus further terms for title, file name, summary and phrase matches. It is the number kb route prints beside every file |
| floor | What the top score has to reach for the verdict to be hit. Below it the router still answers, labelled a guess; at zero it says nothing |
Why one number was wrong twice
idf grows with N, because rarity needs a corpus to be rare in. So a fixed floor means a different number of keys at every size:
| N | one unique key scores | what 17.5 meant |
|---|---|---|
| 4 | 6.6 | three unique keys |
| 11 | 11.2 | two |
| 226 | 28.4 | one |
| 1000 | 37.3 | half a key: a word in fifty files scored 18.2 and cleared it alone |
The floor got harder as the base shrank and easier as it grew, in the two directions that
hurt. Measured on the demo fleet kept in the repository, before any change: with two,
three or four entries, not one of its own gold questions reached hit. Every one was a
guess, on the right file. A library that small could not route at all.
What replaced it
The floor is not a new number. It is the old number, stated in the unit it was always
secretly in. 17.5 on a fleet of 226 entries is 17.5 / (6 × 4.74) = 0.616 of what one
unique key scores there. So:
floor(N) = 0.616 × 6 × idf_unique(N)
Read it as "a result has to score at least 62% of what a single word found in exactly one note would score in this library". At 226 entries that is 17.5 to the last decimal, by construction. At eleven it is 6.9. At four it is 4.1. At a thousand it is 23.0, which puts the word in fifty files back under it.
One more rule, because scaling fixes the ruler and does not conjure one. A fleet of one
entry never routes. With one note every key has df 1, so every key weighs the same and idf
can tell nothing apart; any shared word would clear a floor built from that. Below two
entries the verdict is at most guess. Two is the first size at which a word in both notes
weighs less than a word in one, which is the first size at which there is a ruler.
What it had to pass
Three instruments, same binary either side of the change, run on 2 September 2026.
The fleet the floor was measured on, 226 entries, 33 questions. Identical line for line. Neutral where it was calibrated, and checked rather than assumed.
The demo fleet, 11 entries, 13 questions. Routing went from 8 of 10 to 10 of 10. The two correct answers the old floor demoted to guesses are hits. All three refusals still refuse.
The blind abstention benchmark, 50 questions over the same demo. The number that matters did not move: 28 of 30 out-of-scope questions were still not answered confidently, and the two that got through are the same two medical baits, scoring above either floor. On the in-scope side, confident answers went from 6 to 12, out of the guess column.
And a fourth check, after the change only: a sweep from one entry to eleven. From two up: every gold question a hit on the right
file, no wrong file at any size, every refusal holding.
Every JSON reply carries the floor that actually applied to your fleet in gate.floor, and
every surface that says "against a floor of" reads the same number. The floor is explained
term by term, beside the two scorers it gates, in
How the Catalogue Ranks. The full record, with
every table and the revisit triggers, is ADR-0036,
decisions/0036-the-floor-scales-with-the-corpus.md,
in a repository that is public under Apache 2.0
and cloneable today.