Worse than a fool

I promised a note on the failure where a memory surfaces too much. The fix was a posture, not a trick. When it cannot tell whether a memory is relevant to right now, it says nothing.

I ended the recall note with a debt: surfacing too much is its own failure, the fix is salience, and I was still tuning it. This is that note. It did not go where I expected.

I had built Sapix to speak up when the conversation touched something it already held. The trouble is that touching is cheap. A name appears, a link exists, and it pipes up, whether or not the linked thing has anything to do with what I am actually doing. Presence is not relevance. A memory that surfaces because a word matched is barely better than one that never surfaces at all; it just fails louder.

So I went to measure how bad it was, and the first thing I found was that the log I would have measured from was poisoned, by my own tests, writing their fixtures into the same file the real signal lived in. I had been about to tune the system against numbers that were half noise of my own making. Not a flattering lesson: the instrument was lying, and I almost trusted it.

The clean data said the obvious thing. The fix was not a cleverer score. It was a posture. Before surfacing anything, ask the warm cortex whether this note is actually about the current moment, and if the answer cannot be had, for any reason, say nothing.

That last clause is the whole point, and it is where I landed on a rule I now like more than the feature it governs. Worse than a fool is a motivated fool. A system that stays quiet when it is unsure is merely unhelpful. One that speaks up confidently while being wrong about what you need is actively worse, because it spends your trust to do it. So when the relevance check cannot run, the answer is not to guess and surface anyway. It is to fail closed. Quiet is the safe direction; confident-and-wrong is not.

The part I have not finished is the one that makes this honest at scale. I cannot tell, by hand, whether the things it now chooses to surface are any good. Nobody can, past a certain volume. So the real work is not the threshold I picked today; it is the mechanism that watches whether a surfaced memory got used, and tunes itself from that. The judgment was never going to be mine to give note by note. The most I can build is the thing that learns it, and then glance at it now and then.