1# Dataset construction (deterministic). Seeds define split membership; keys are written to data/generated/keys (evaluator-only).2synthetic:3 n_projects: 64 lengths_tokens: [2000, 8000, 16000]5 # seed ranges by split (disjoint; template families identical, entities/numbers differ)6 splits:7 train: {start: 1000, count: 40} # mapper/calibration/prompt tuning8 dev: {start: 2000, count: 30} # method selection, router fitting9 test: {start: 3000, count: 40} # locked; first evaluated at Phase 810 length_assignment: round_robin # each seed gets one length: 2000, 8000, 16000, 2000, ...11exec:12 splits:13 train: {start: 1, count: 20}14 dev: {start: 101, count: 30}15 test: {start: 201, count: 40}16 target_tokens: 200017public:18 subsets: [qasper, 2wikimqa]19 extension_subsets: [hotpotqa]20 split_rule: "sha256(document text) last hex digit: 0-3 -> dev, 4-f -> test (dev ~25%, test ~75%)"21 max_context_tokens: 1638422