Today oracles are per-canonical-pattern. Add model-level oracles — a hand/agent-written reference for a whole model graph (or a major subgraph), so we can express 'here's how fast this model should run' directly, not only as a roll-up of per-kernel floors. Complements the model_attribution roll-up (which composes per-kernel oracles).
Today oracles are per-canonical-pattern. Add model-level oracles — a hand/agent-written reference for a whole model graph (or a major subgraph), so we can express 'here's how fast this model should run' directly, not only as a roll-up of per-kernel floors. Complements the model_attribution roll-up (which composes per-kernel oracles).