1. Build the ledger. Enter Table 2 of TR-2026-35: four walls, the open problems each drew across the six practitioner chapters, and the Institute holdings that sit against each. Print both columns as counts and as shares. The totals are 15 and 25.
2. Rank both columns. Sort each column and read off the last entry. Wall 2 is last on the field's side and last on ours, at 2 of 15 and 4 of 25.
3. Normalise a second way. Divide holdings by named open problems. Now wall 4 is thinnest, at 1.20, and wall 2 sits level with wall 1 at 2.00. Both readings are correct arithmetic on the same table, so an attention count is only a finding once you have said which normalisation produced it.
4. Price the last-ranked wall. Two published instruments say what wall 2 costs. RoboChallenge's temporal tag runs at 5% success against a 22% all-task mean, and the report's own explanation is that every model it tested is single-frame and therefore carries no transition function at all. EWMBench scores appearance separately from action-conditioned dynamics, and OpenSora reaches 97.7% of the best model's appearance score at 8.8% of its dynamics score. Compute both ratios before you read the next lesson.