Part 14 of 16 · 1 min read
Quick notes
- Context is rebuilt from a store every step, under a budget, because the model remembers nothing between calls.
- Pin what must never be dropped, and supersede stale facts with their newer versions, because an unpinned rule can be crowded out and a stale value beside a corrected one leaves the model guessing.
- Selection is only as good as its signal. Lexical matching finds plainly worded answers and misses vaguely worded ones; presence is a checkable fact, modelled accuracy is only a direction.
- Fence every item from outside, including the user's own text, so nothing in it can forge a structural header; this stops forgery, not persuasion.
- Measure presence and use separately, and log what was sent, so "the model got it wrong" and "the model was never shown it" are never confused.