What this proves: exact recall of an arbitrary fact, from any position, at distances far beyond what the model can see at once, from text it has never seen. What it does not prove yet: the model still needs the fact written in one trained record shape and asked in that shape. Storing what matters unprompted, and answering questions in your own words, is what the next training run is for. We say so on purpose.

Why the gibberish key? A 4-letter key like qbfx cannot be guessed from the document or from general knowledge — the only way to answer is to have remembered. Why this test? It is the needle-in-a-haystack test frontier models are scored on — except they keep the whole document in view and still lose the middle; this one reads it once and lets it go. Results on external documents: 48/48 exact at 12–98KB, random positions (Sep 2026). Architecture withheld; results published. kymnera.ai