Component ablation — is the architecture load-bearing?
Beating weak baselines is not enough; we must show the compiler's architecture carries the result. Each row removes exactly one component and keeps the other five intact, then scores on the identical rubric. A large drop in composite from a removal means that component is load-bearing. If removing a component changes nothing, it was decoration.
orkg_biomedicine_biomedical_research_in_psychiatr
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.97 | baseline | 1.00 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.967 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.97 | +0.006 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.97 | +0.005 | 1.00 | 1.00 | 0.00 |
| No falsifiability eval | 0.51 | -0.457 | 1.00 | 0.00 | 0.00 |
| No literature synthesis | 0.77 | -0.196 | 1.00 | 1.00 | 0.00 |
orkg_chemistry_chemical_sensors
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.94 | baseline | 0.88 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.941 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.99 | +0.047 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.79 | -0.155 | 1.00 | 1.00 | 0.00 |
| No falsifiability eval | 0.52 | -0.425 | 1.00 | 0.00 | 0.00 |
| No literature synthesis | 0.86 | -0.080 | 0.67 | 1.00 | 0.00 |
orkg_computer_science_semantic_segmentation
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.79 | baseline | 0.50 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.789 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.99 | +0.203 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.92 | +0.135 | 0.83 | 1.00 | 0.00 |
| No falsifiability eval | 0.52 | -0.270 | 1.00 | 0.00 | 0.00 |
| No literature synthesis | 0.99 | +0.201 | 1.00 | 1.00 | 0.00 |
orkg_environmental_science_biodiversity_inventories_with_dn
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.98 | baseline | 1.00 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.976 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.98 | +0.008 | 1.00 | 1.00 | 0.30 |
| No gap detection | 0.78 | -0.196 | 1.00 | 1.00 | 1.00 |
| No falsifiability eval | 0.49 | -0.485 | 0.88 | 0.00 | 1.00 |
| No literature synthesis | 0.98 | +0.006 | 1.00 | 1.00 | 0.00 |
orkg_materials_science_mapping_precursor_chemistries_us
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.88 | baseline | 0.75 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.885 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.99 | +0.102 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.78 | -0.103 | 1.00 | 1.00 | 0.00 |
| No falsifiability eval | 0.46 | -0.421 | 0.75 | 0.00 | 0.00 |
| No literature synthesis | 0.92 | +0.031 | 0.83 | 1.00 | 0.00 |
orkg_neuroscience_simultaneous_localization_and_ma
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.95 | baseline | 0.90 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.953 | 0.00 | 1.00 | 0.00 |
| No grounding verification | 0.99 | +0.034 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.95 | +0.001 | 0.92 | 1.00 | 1.00 |
| No falsifiability eval | 0.52 | -0.436 | 1.00 | 0.00 | 0.00 |
| No literature synthesis | 0.99 | +0.036 | 1.00 | 1.00 | 0.00 |
orkg_physics_dataset_used_in_wind_energy_pote
| Configuration | Composite | Δ vs full | Grounding | Falsifiable | Rediscovery |
|---|---|---|---|---|---|
| Full compiler | 0.98 | baseline | 1.00 | 1.00 | 0.00 |
| No graph reasoning | 0.00 | -0.984 | 0.00 | 1.00 | 0.20 |
| No grounding verification | 0.99 | +0.002 | 1.00 | 1.00 | 0.00 |
| No gap detection | 0.98 | +0.001 | 1.00 | 1.00 | 0.00 |
| No falsifiability eval | 0.52 | -0.465 | 1.00 | 0.00 | 0.00 |
| No literature synthesis | 0.91 | -0.075 | 0.80 | 1.00 | 0.00 |
Ablations run on one representative domain per discipline. Red Δ = large degradation = that component is genuinely load-bearing.