EXPERIMENT_MATRIX.md

Experiment Matrix

The current main experiment has no repeated seed axis. Each generated matrix row receives one fresh random rng_seed automatically.

Main Matrix

3 problems x 3 representations x 3 local searches x 3 surrogates = 81 rows

Generate it:

scripts/make_main_matrix.sh

The output is runs/main_81_matrix.jsonl.

Problems and Components

MISR

Representations:

• endpoint_sequence_pair
• triangle_free_rect
• quadratic_program_rectangles

Local searches:

• sequence_pair_pivot
• lp_dual_pivot
• program_coeff_pivot

Surrogates:

• exact_lp_gap_pressure
• triangle_free_exact_gap_pressure
• graph_conflict_proxy

Unit-Square Stabbing

Representations:

• square_direct
• line_square_incidence
• sqstab_exact_grid

Local searches:

• coord_mutation
• primal_dual_lines
• sqstab_local_hillclimb

Surrogates:

• greedy_partial_lp_bitset
• exact_stab_gap_pressure
• incidence_statistics

Guillotine

Representations:

• rect_direct_disjoint
• sequence_pair_packing
• recursive_obstruction_grammar

Local searches:

• packing_resize
• recursive_gadget_assembly
• witness_breaking

Surrogates:

• first_cut_obstruction
• depth_limited_dp
• k_subset_nonseparability

Controls

The control matrix has 9 rows:

3 problems x 3 control modes = 9 rows

Control modes:

• local_only
• model_only_weak_local
• shuffled_label

Generate it:

PYTHONPATH=src python3 -m multilevel.cli control-matrix \
  --stage controls \
  --budget-seconds 14400 \
  --problem misr \
  --problem unit_square \
  --problem guillotine \
  --out runs/control_9_matrix.jsonl

Nontrivial Size Gates

The PatternBoost training filter rejects tiny generated examples:

• misr: at least 8 rectangles
• unit_square: at least 8 squares and exact certificates should meet the current stabbing-cover pressure gate
• guillotine: at least 8 rectangles and at least 2 destroyed rectangles when that certificate field is present

The CLI flag --n can raise the minimum. For example, --n 12 means at least 12 items, not at least 8.