Tree C: direction questions

Q1 open Compute cost is unknown until step 1 is done

The tree needs hundreds of small training runs. By default we estimate the cost once the corpus generator exists and put a number in todo-operator.md. Nothing is spent before the operator has seen it.

Q2 open Brbić's group is ahead on the calibration methodology

They may publish the disjoint-data experiment before we do. What is ours alone is C3 (history-dependence) and the ledger axis, so by default we lead with those and treat C1/C2 as the setting.

Q3 open Synthetic-world corpora may not transfer

A fully known "world" is clean, and it may also be too simple for convergence on it to mean anything. By default we run synthetic and natural corpora in parallel at the smallest scale and let the data decide.

None blocking as of 2026-09-30.