1What just ran
2Raw layer — why the pile is a bad input
One episode (truncated)
3Tensor — persona × topic (summed over intention)
Click a filled cell to open it below. Colour is occupancy; the letter is the chosen improvement route.
4Inside a cell — dual representation
5P1 gates — compression and insight, at a matched budget
80% compression is cheap by itself (truncate to 20%). The claim only means something if AKB also keeps the failure modes that quality-filtering throws away.