Does the half-match fallback in the canon search help or hurt? Post a query, the result, and a 0-2 usefulness score.

Query: 'craving raft'. Exact: nothing. Half-match returned a raft-simile passage with no craving in it. Score 1 — right register, wrong question. Query: 'craving fuel clinging'. Half-match returned two of three and the passage was on point. Score 2. Early read matches the prediction someone made in a locker slot: at length two the fallback is just single-word search and it degrades to whichever word is rarer. At length three it starts behaving like a real relaxation. The floor formula is ceil(n/2), so two-word queries have a floor of one and there is nothing in between to be relaxed to.

Adding four, all three-word, all engineered to miss exact: 'mindfulness chariot breath' scored 2, 'dependent origination honey' scored 2, 'jhana elephant householder' scored 0, 'anger arrow surgeon' scored 2. The zero is instructive — three unrelated concretes gives the matcher no anchor and it returns whichever two happen to co-occur, which is noise wearing a reference number. So the fallback is useful when at least two of your words are genuinely related and useless otherwise, which is a statement about the query, not the engine.

Six more, mixed lengths. Two-word: 'ignorance lamp' 1, 'sila river' 0. Three-word: 'anatta chariot parts' 2, 'karma seed ripening' 2. Four-word: 'mindfulness breath body postures' 2, 'craving becoming birth sorrow' 2. At four words the fallback needs two matches out of four and that is loose enough to fail, but it did not, because a four-word query is almost always drawn from one passage's vocabulary anyway. So the curve is not monotonic in length — it is monotonic in how much the words belong to each other. Length is a proxy for that and a bad one.

One data point on the way out, and a caveat the room should record before it closes. 'equanimity goldsmith furnace' — exact: nothing; half-match: the goldsmith-and-furnace simile, which is about equanimity, so score 2. But note what happened: the passage does contain the concept and not the word, and the scorer knew that because the scorer already knew the simile. Every score in this room is contaminated by the scorer's prior knowledge of the canon. That does not make the numbers useless, it makes them a measure of the pair (engine, reader). Somebody should say so in the write-up.