Qwen3.6 SciCode MoE Expert Saliency Analysis

Model: unsloth/Qwen3.6-35B-A3B-GGUF UD-Q4_K_XL. Dataset slice: SciCode MVP, prompt-prefill traces. Generated from mvp_expert_heatmap.parquet.

Interpretation boundary: this is observational saliency/proxy importance, not gradient saliency or causal ablation. Importance here combines router-weight mass and contribution L2 proxy from traced llama.cpp tensors. It is strong evidence for which experts are hot and carrying mass on these SciCode prompts, but it is not yet proof that removing/cold-loading the remaining experts is lossless.
Trace rows
76,604,880
Questions
291
Experts observed
256
Layer-expert pairs
10,061
Effective experts, importance
122.0 / 256
Global expert Gini, importance
0.508
Top 20 experts importance
29.7%
Effective layer-expert pairs
274 / 10,061

Executive Takeaways

How Many Experts Cover the Mass?

Rows with mass_threshold answer: how many experts are needed to account for 20%, 50%, 80%, 90%, or 95% of the measured mass. Rows with top_entity_fraction answer the inverse question: how much mass is captured by the top 20%, 50%, or 80% of experts/pairs.

metric entity mass_threshold count_needed share_of_entities top_entity_fraction count_used mass_captured
expert importance expert_id 20.0% 12 4.7%
expert importance expert_id 50.0% 45 17.6%
expert importance expert_id 80.0% 111 43.4%
expert importance expert_id 90.0% 155 60.5%
expert importance expert_id 95.0% 189 73.8%
expert importance expert_id 20.0% 52 54.9%
expert importance expert_id 50.0% 128 84.6%
expert importance expert_id 80.0% 205 96.8%
expert activation frequency expert_id 20.0% 31 12.1%
expert activation frequency expert_id 50.0% 95 37.1%
expert activation frequency expert_id 80.0% 178 69.5%
expert activation frequency expert_id 90.0% 213 83.2%
expert activation frequency expert_id 95.0% 233 91.0%
expert activation frequency expert_id 20.0% 52 31.1%
expert activation frequency expert_id 50.0% 128 63.2%
expert activation frequency expert_id 80.0% 205 88.1%
layer-expert importance layer_expert 20.0% 22 0.2%
layer-expert importance layer_expert 50.0% 110 1.1%
layer-expert importance layer_expert 80.0% 472 4.7%
layer-expert importance layer_expert 90.0% 949 9.4%
layer-expert importance layer_expert 95.0% 1,594 15.8%
layer-expert importance layer_expert 20.0% 2,013 96.6%
layer-expert importance layer_expert 50.0% 5,031 99.7%
layer-expert importance layer_expert 80.0% 8,049 100.0%
layer-expert activation frequency layer_expert 20.0% 223 2.2%
layer-expert activation frequency layer_expert 50.0% 1,105 11.0%
layer-expert activation frequency layer_expert 80.0% 3,395 33.7%
layer-expert activation frequency layer_expert 90.0% 4,939 49.1%
layer-expert activation frequency layer_expert 95.0% 6,152 61.1%
layer-expert activation frequency layer_expert 20.0% 2,013 65.4%
layer-expert activation frequency layer_expert 50.0% 5,031 90.5%
layer-expert activation frequency layer_expert 80.0% 8,049 99.1%

Distribution Tightness by Layer

Lower effective expert count and higher top-10 share mean a layer is routing importance through a tighter expert subset.

layer effective_n_hhi gini top1_share top5_share top10_share top20_share
20 6.729 0.946 31.6% 72.4% 83.6% 90.0%
12 7.598 0.9298 26.5% 68.0% 77.6% 86.8%
1 8.033 0.7783 33.5% 52.0% 59.9% 67.3%
8 8.805 0.9414 19.4% 68.6% 82.2% 91.3%
29 8.893 0.9009 30.2% 53.4% 66.7% 82.1%
5 9.053 0.9137 29.0% 55.2% 71.5% 86.2%
35 9.137 0.9304 21.5% 65.0% 81.5% 88.7%
37 9.183 0.9284 24.8% 61.6% 76.7% 87.4%
39 9.26 0.8415 19.0% 68.1% 80.0% 89.8%
14 9.267 0.9394 22.4% 62.8% 78.3% 90.4%
38 9.295 0.9301 25.7% 61.7% 79.9% 89.7%
7 9.599 0.9314 22.6% 59.0% 76.7% 90.3%
28 9.733 0.9221 23.4% 60.7% 76.9% 87.4%
36 9.894 0.9219 22.4% 60.1% 75.9% 87.8%
15 9.913 0.923 26.3% 54.7% 73.5% 88.0%
21 10.2 0.9393 19.1% 64.6% 78.7% 90.6%
18 10.75 0.9291 19.5% 59.7% 75.8% 88.2%
9 11.19 0.9283 19.8% 59.7% 76.6% 88.3%
24 11.27 0.913 20.0% 54.9% 69.0% 82.8%
11 11.46 0.9012 20.3% 55.7% 70.9% 82.5%
23 11.68 0.9182 16.6% 59.4% 76.4% 87.1%
27 11.81 0.9052 20.4% 54.1% 70.1% 83.2%
4 11.95 0.8947 17.5% 59.6% 70.3% 80.0%
19 12.1 0.9212 21.5% 52.5% 71.3% 87.0%
26 12.18 0.9235 16.8% 58.3% 73.2% 88.1%
13 12.73 0.9051 21.0% 49.1% 65.1% 82.1%
25 13.29 0.9041 21.0% 49.2% 63.9% 79.9%
22 13.77 0.923 13.2% 54.4% 74.7% 86.9%
3 14.49 0.8631 16.0% 50.9% 67.1% 78.5%
33 14.94 0.8827 18.4% 48.6% 64.8% 78.6%
31 14.98 0.891 20.2% 44.0% 62.0% 80.7%
34 15.13 0.8977 14.7% 51.0% 67.6% 80.8%
32 15.72 0.8842 19.6% 43.2% 61.2% 78.9%
6 16.19 0.8853 17.0% 46.3% 64.3% 78.2%
30 16.29 0.8872 14.0% 49.5% 65.9% 79.4%
16 17.55 0.9126 11.0% 45.9% 67.4% 85.5%
17 18.47 0.9081 13.0% 43.2% 63.8% 84.8%
10 18.54 0.8887 12.2% 43.5% 64.0% 79.4%
0 43.68 0.683 10.1% 25.5% 35.5% 49.3%
2 44.44 0.6539 8.8% 27.1% 36.9% 47.9%

Layer x Expert Saliency

Top Global Experts

This table ranks experts after summing across all layers and SciCode questions. domain is the highest-share inferred task domain for that expert; lift_vs_corpus shows whether that domain is overrepresented for the expert versus the full SciCode MVP mix.

expert_id importance_score importance_score_share activation_count activation_count_share router_weight_mass contribution_l2_sum active_questions domain expert_domain_share lift_vs_corpus top_questions
209 1.044e+07 2.3% 519,228 0.7% 6.335e+04 1.451e+05 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-77-77.12 (Berendsen_thermostat)
31 9.549e+06 2.1% 470,490 0.6% 6.237e+04 1.908e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-55-55.1 (Swift_Hohenberg); scicode-77-77.12 (Berendsen_thermostat)
221 9.497e+06 2.1% 472,993 0.6% 7.132e+04 9.855e+04 291 ODE / time integration 100.0% 5.809 scicode-54-54.2 (SUPG); scicode-54-54.1 (SUPG); scicode-54-54.3 (SUPG)
230 8.608e+06 1.9% 309,188 0.4% 4.134e+04 1.399e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-79-79.4 (Nose_Hoover_chain_thermostat); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array)
130 8.485e+06 1.9% 697,441 0.9% 9.812e+04 1.467e+05 291 ODE / time integration 100.0% 5.809 scicode-77-77.12 (Berendsen_thermostat); scicode-77-77.4 (Berendsen_thermostat); scicode-33-33.1 (phase_diagram_chern_haldane_model_v1)
64 7.205e+06 1.6% 608,325 0.8% 7.835e+04 1.71e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-77-77.12 (Berendsen_thermostat); scicode-13-13.5 (Maxwell_Equation_Solver)
52 6.99e+06 1.5% 292,328 0.4% 4.05e+04 1.344e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-20-20.2 (phonon_angular_momentum); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
118 6.92e+06 1.5% 544,294 0.7% 7.176e+04 1.57e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-55-55.1 (Swift_Hohenberg); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
90 6.458e+06 1.4% 575,565 0.8% 7.927e+04 1.225e+05 291 ODE / time integration 100.0% 5.809 scicode-20-20.2 (phonon_angular_momentum); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-33-33.1 (phase_diagram_chern_haldane_model_v1)
169 6.425e+06 1.4% 293,399 0.4% 3.74e+04 1.077e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-20-20.2 (phonon_angular_momentum)
201 5.95e+06 1.3% 480,424 0.6% 6.961e+04 1.37e+05 291 ODE / time integration 100.0% 5.809 scicode-20-20.2 (phonon_angular_momentum); scicode-18-18.2 (NURBS); scicode-15-15.1 (Crank_Nicolson_for_time_dependent_Schrodinger)
121 5.891e+06 1.3% 541,786 0.7% 6.955e+04 1.231e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-12-12.2 (Schrodinger_DFT_with_SCF)
140 5.841e+06 1.3% 372,033 0.5% 5.209e+04 1.017e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
73 5.786e+06 1.3% 406,447 0.5% 5.3e+04 1.117e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-17-17.1 (linear_tetrahedron_method); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
134 5.346e+06 1.2% 584,087 0.8% 7.233e+04 1.427e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-20-20.2 (phonon_angular_momentum)
146 5.332e+06 1.2% 265,222 0.3% 3.423e+04 1.149e+05 291 ODE / time integration 100.0% 5.809 scicode-77-77.12 (Berendsen_thermostat); scicode-60-60.5 (Widom_particle_insertion); scicode-77-77.4 (Berendsen_thermostat)
96 5.291e+06 1.2% 288,969 0.4% 3.789e+04 1.108e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-22-22.1 (Beam_translation_reexpansion); scicode-79-79.4 (Nose_Hoover_chain_thermostat)
251 5.049e+06 1.1% 515,632 0.7% 6.975e+04 1.52e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-12-12.2 (Schrodinger_DFT_with_SCF)
226 4.963e+06 1.1% 547,163 0.7% 6.733e+04 1.451e+05 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-55-55.1 (Swift_Hohenberg); scicode-55-55.2 (Swift_Hohenberg)
117 4.888e+06 1.1% 596,647 0.8% 7.393e+04 1.24e+05 291 ODE / time integration 100.0% 5.809 scicode-55-55.1 (Swift_Hohenberg); scicode-55-55.3 (Swift_Hohenberg); scicode-55-55.2 (Swift_Hohenberg)
102 4.77e+06 1.0% 360,305 0.5% 4.982e+04 1.02e+05 291 ODE / time integration 100.0% 5.809 scicode-59-59.4 (VQE); scicode-65-65.4 (GHZ_protocol_fidelity); scicode-65-65.2 (GHZ_protocol_fidelity)
103 4.698e+06 1.0% 318,659 0.4% 4.287e+04 1.635e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-20-20.2 (phonon_angular_momentum)
162 4.607e+06 1.0% 413,880 0.5% 5.279e+04 8.564e+04 291 ODE / time integration 100.0% 5.809 scicode-54-54.2 (SUPG); scicode-54-54.3 (SUPG); scicode-13-13.5 (Maxwell_Equation_Solver)
78 4.557e+06 1.0% 368,636 0.5% 4.885e+04 1.082e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-12-12.2 (Schrodinger_DFT_with_SCF)
227 4.425e+06 1.0% 284,427 0.4% 3.925e+04 6.268e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-43-43.1 (two_end_fiber_laser_generator); scicode-37-37.1 (ray_optics_spherical_aberration)
1 4.271e+06 0.9% 546,892 0.7% 7.471e+04 1.238e+05 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-43-43.1 (two_end_fiber_laser_generator); scicode-28-28.2 (Gaussian_Beam_Intensity)
89 4.15e+06 0.9% 299,631 0.4% 4.037e+04 1.196e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-22-22.1 (Beam_translation_reexpansion); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
199 4.145e+06 0.9% 228,845 0.3% 2.768e+04 1.098e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-79-79.4 (Nose_Hoover_chain_thermostat); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array)
53 4.108e+06 0.9% 502,642 0.7% 6.335e+04 1.005e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array)
166 4.091e+06 0.9% 477,474 0.6% 6.296e+04 9.258e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-9-9.1 (Weighted_Jacobi); scicode-5-5.1 (Lanczos)
57 3.947e+06 0.9% 267,375 0.3% 3.318e+04 1.201e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-79-79.4 (Nose_Hoover_chain_thermostat); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array)
12 3.732e+06 0.8% 419,716 0.5% 5.628e+04 1.027e+05 291 ODE / time integration 100.0% 5.809 scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-33-33.1 (phase_diagram_chern_haldane_model_v1); scicode-13-13.9 (Maxwell_Equation_Solver)
75 3.555e+06 0.8% 465,357 0.6% 5.823e+04 1.066e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
76 3.554e+06 0.8% 327,518 0.4% 4.346e+04 8.387e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-33-33.1 (phase_diagram_chern_haldane_model_v1); scicode-36-36.2 (Quasi_Fermi_levels_of_photo_resistor_out_of_equilibrium)
72 3.501e+06 0.8% 487,065 0.6% 6.381e+04 1.347e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-22-22.1 (Beam_translation_reexpansion)
150 3.3e+06 0.7% 246,344 0.3% 3.537e+04 7.47e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-20-20.2 (phonon_angular_momentum)
160 3.264e+06 0.7% 431,146 0.6% 6.24e+04 9.554e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-77-77.12 (Berendsen_thermostat)
126 3.254e+06 0.7% 338,628 0.4% 4.397e+04 8.702e+04 291 ODE / time integration 100.0% 5.809 scicode-73-73.3 (Xray_conversion_II); scicode-73-73.8 (Xray_conversion_II); scicode-73-73.5 (Xray_conversion_II)
124 3.233e+06 0.7% 466,685 0.6% 5.747e+04 7.909e+04 291 ODE / time integration 100.0% 5.809 scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-55-55.1 (Swift_Hohenberg); scicode-55-55.3 (Swift_Hohenberg)
33 3.231e+06 0.7% 215,746 0.3% 3.242e+04 8.855e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-20-20.2 (phonon_angular_momentum)
176 3.222e+06 0.7% 409,482 0.5% 5.049e+04 8.432e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-17-17.1 (linear_tetrahedron_method); scicode-20-20.2 (phonon_angular_momentum)
58 3.208e+06 0.7% 434,967 0.6% 5.611e+04 8.66e+04 291 ODE / time integration 100.0% 5.809 scicode-55-55.1 (Swift_Hohenberg); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-55-55.4 (Swift_Hohenberg)
15 3.199e+06 0.7% 468,579 0.6% 6.165e+04 1.151e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-63-63.1 (Estimating_Stock_Option_Price)
175 3.155e+06 0.7% 351,948 0.5% 4.534e+04 9.671e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-43-43.1 (two_end_fiber_laser_generator); scicode-43-43.3 (two_end_fiber_laser_generator)
50 3.088e+06 0.7% 489,946 0.6% 5.728e+04 1.279e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-13-13.5 (Maxwell_Equation_Solver)
198 3.079e+06 0.7% 456,616 0.6% 5.74e+04 9.783e+04 291 ODE / time integration 100.0% 5.809 scicode-12-12.2 (Schrodinger_DFT_with_SCF); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-17-17.1 (linear_tetrahedron_method)
16 2.926e+06 0.6% 184,399 0.2% 3.094e+04 6.496e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-20-20.2 (phonon_angular_momentum)
210 2.901e+06 0.6% 413,436 0.5% 5.687e+04 8.225e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array)
91 2.893e+06 0.6% 268,537 0.4% 3.344e+04 8.734e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-79-79.4 (Nose_Hoover_chain_thermostat); scicode-22-22.1 (Beam_translation_reexpansion)
187 2.88e+06 0.6% 353,306 0.5% 4.379e+04 7.733e+04 291 ODE / time integration 100.0% 5.809 scicode-73-73.3 (Xray_conversion_II); scicode-73-73.8 (Xray_conversion_II); scicode-2-2.1 (Gaussian_Beam_Focus)
35 2.869e+06 0.6% 360,828 0.5% 4.32e+04 9.327e+04 291 ODE / time integration 100.0% 5.809 scicode-55-55.1 (Swift_Hohenberg); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-13-13.11 (Maxwell_Equation_Solver)
179 2.832e+06 0.6% 382,373 0.5% 4.798e+04 8.504e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-43-43.1 (two_end_fiber_laser_generator); scicode-28-28.1 (Gaussian_Beam_Intensity)
133 2.815e+06 0.6% 385,341 0.5% 5.328e+04 1.012e+05 291 ODE / time integration 100.0% 5.809 scicode-20-20.2 (phonon_angular_momentum); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-45-45.1 (finite_difference_heat_equation)
74 2.814e+06 0.6% 351,016 0.5% 4.955e+04 1.016e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-20-20.2 (phonon_angular_momentum); scicode-2-2.1 (Gaussian_Beam_Focus)
45 2.787e+06 0.6% 417,156 0.5% 5.275e+04 9.504e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-20-20.2 (phonon_angular_momentum); scicode-17-17.1 (linear_tetrahedron_method)
151 2.719e+06 0.6% 522,223 0.7% 6.522e+04 9.693e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-64-64.6 (GCMC)
29 2.715e+06 0.6% 183,210 0.2% 2.37e+04 7.616e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-54-54.2 (SUPG); scicode-22-22.1 (Beam_translation_reexpansion)
14 2.689e+06 0.6% 284,454 0.4% 3.689e+04 8.111e+04 291 ODE / time integration 100.0% 5.809 scicode-20-20.2 (phonon_angular_momentum); scicode-55-55.3 (Swift_Hohenberg); scicode-55-55.1 (Swift_Hohenberg)
83 2.668e+06 0.6% 500,301 0.7% 6.171e+04 1.05e+05 291 ODE / time integration 100.0% 5.809 scicode-20-20.2 (phonon_angular_momentum); scicode-55-55.1 (Swift_Hohenberg); scicode-55-55.3 (Swift_Hohenberg)
248 2.644e+06 0.6% 392,592 0.5% 4.973e+04 8.329e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-32-32.1 (Multiparticle_dynamics_in_the_optical_tweezer_array); scicode-43-43.1 (two_end_fiber_laser_generator)
183 2.538e+06 0.6% 400,470 0.5% 4.889e+04 8.361e+04 291 ODE / time integration 100.0% 5.809 scicode-45-45.1 (finite_difference_heat_equation); scicode-77-77.12 (Berendsen_thermostat); scicode-20-20.2 (phonon_angular_momentum)
157 2.538e+06 0.6% 365,166 0.5% 5.033e+04 9.33e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-17-17.1 (linear_tetrahedron_method)
85 2.449e+06 0.5% 465,572 0.6% 5.849e+04 1.187e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-12-12.8 (Schrodinger_DFT_with_SCF); scicode-20-20.2 (phonon_angular_momentum)
202 2.443e+06 0.5% 406,883 0.5% 4.853e+04 9.327e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-22-22.1 (Beam_translation_reexpansion); scicode-69-69.7 (LEG_Dyson_equation_semi_infinite)
0 2.384e+06 0.5% 413,818 0.5% 5.311e+04 1.122e+05 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-12-12.2 (Schrodinger_DFT_with_SCF)
186 2.382e+06 0.5% 328,666 0.4% 4.189e+04 6.453e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-54-54.2 (SUPG); scicode-20-20.2 (phonon_angular_momentum)
47 2.327e+06 0.5% 350,049 0.5% 4.231e+04 6.819e+04 291 ODE / time integration 100.0% 5.809 scicode-13-13.9 (Maxwell_Equation_Solver); scicode-13-13.8 (Maxwell_Equation_Solver); scicode-13-13.11 (Maxwell_Equation_Solver)
3 2.324e+06 0.5% 356,205 0.5% 4.282e+04 9.572e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-2-2.1 (Gaussian_Beam_Focus); scicode-12-12.8 (Schrodinger_DFT_with_SCF)
173 2.264e+06 0.5% 268,051 0.3% 3.289e+04 7.044e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-77-77.12 (Berendsen_thermostat); scicode-55-55.1 (Swift_Hohenberg)
207 2.252e+06 0.5% 258,108 0.3% 3.325e+04 7.325e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-22-22.1 (Beam_translation_reexpansion); scicode-54-54.2 (SUPG)
158 2.223e+06 0.5% 280,771 0.4% 3.211e+04 6.326e+04 291 ODE / time integration 100.0% 5.809 scicode-9-9.1 (Weighted_Jacobi); scicode-55-55.3 (Swift_Hohenberg); scicode-55-55.1 (Swift_Hohenberg)
95 2.197e+06 0.5% 412,676 0.5% 5.025e+04 8.773e+04 291 ODE / time integration 100.0% 5.809 scicode-63-63.1 (Estimating_Stock_Option_Price); scicode-63-63.2 (Estimating_Stock_Option_Price); scicode-63-63.3 (Estimating_Stock_Option_Price)
56 2.193e+06 0.5% 316,618 0.4% 3.874e+04 9.474e+04 291 ODE / time integration 100.0% 5.809 scicode-55-55.3 (Swift_Hohenberg); scicode-55-55.1 (Swift_Hohenberg); scicode-55-55.2 (Swift_Hohenberg)
105 2.184e+06 0.5% 346,364 0.5% 4.231e+04 6.789e+04 291 ODE / time integration 100.0% 5.809 scicode-77-77.12 (Berendsen_thermostat); scicode-77-77.4 (Berendsen_thermostat); scicode-77-77.10 (Berendsen_thermostat)
23 2.111e+06 0.5% 450,371 0.6% 5.727e+04 9.894e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-43-43.1 (two_end_fiber_laser_generator); scicode-79-79.2 (Nose_Hoover_chain_thermostat)
135 2.068e+06 0.5% 394,154 0.5% 4.804e+04 7.609e+04 291 ODE / time integration 100.0% 5.809 scicode-58-58.4 (Tolman_Oppenheimer_Volkoff_star); scicode-22-22.2 (Beam_translation_reexpansion); scicode-13-13.9 (Maxwell_Equation_Solver)
163 2.05e+06 0.5% 360,773 0.5% 5.265e+04 8.067e+04 291 ODE / time integration 100.0% 5.809 scicode-77-77.12 (Berendsen_thermostat); scicode-73-73.8 (Xray_conversion_II); scicode-73-73.5 (Xray_conversion_II)
155 2.024e+06 0.4% 234,971 0.3% 2.875e+04 6.527e+04 291 ODE / time integration 100.0% 5.809 scicode-2-2.1 (Gaussian_Beam_Focus); scicode-55-55.1 (Swift_Hohenberg); scicode-43-43.1 (two_end_fiber_laser_generator)
81 2.022e+06 0.4% 317,937 0.4% 4.31e+04 5.025e+04 291 ODE / time integration 100.0% 5.809 scicode-66-66.6 (kolmogorov_crespi_potential); scicode-20-20.2 (phonon_angular_momentum); scicode-15-15.1 (Crank_Nicolson_for_time_dependent_Schrodinger)
30 1.983e+06 0.4% 365,569 0.5% 4.742e+04 5.892e+04 291 ODE / time integration 100.0% 5.809 scicode-79-79.2 (Nose_Hoover_chain_thermostat); scicode-53-53.1 (Stochastic_Lotka_Volterra); scicode-50-50.4 (Replica_symmetry_breaking)

Top Layer-Expert Pairs

These are the most actionable residency candidates if the runtime can pin at layer-expert granularity rather than only by expert ID globally.

layer expert_id importance_score importance_share activation_count router_weight_mass contribution_l2_sum active_questions
38 209 9.201e+06 2.0% 188,190 2.529e+04 9.407e+04 291
37 230 7.749e+06 1.7% 45,295 8,501 9.564e+04 257
35 169 6.284e+06 1.4% 102,937 1.497e+04 8.021e+04 291
20 221 6.25e+06 1.4% 132,152 2.274e+04 3.531e+04 291
36 52 5.838e+06 1.3% 105,104 1.648e+04 7.863e+04 291
35 31 5.487e+06 1.2% 107,094 1.524e+04 8.114e+04 291
37 146 5.002e+06 1.1% 93,562 1.353e+04 7.669e+04 291
36 96 4.46e+06 1.0% 42,722 7,796 6.712e+04 287
38 57 3.41e+06 0.7% 37,389 5,565 6.904e+04 264
38 199 3.343e+06 0.7% 39,488 6,248 6.082e+04 255
38 103 3.251e+06 0.7% 26,183 5,704 9.411e+04 291
29 150 3.107e+06 0.7% 45,195 1.196e+04 4.914e+04 291
38 89 2.86e+06 0.6% 39,265 5,998 6.767e+04 288
15 73 2.851e+06 0.6% 111,074 1.675e+04 2.964e+04 291
14 121 2.828e+06 0.6% 104,859 1.587e+04 2.785e+04 291
28 130 2.827e+06 0.6% 152,755 2.172e+04 3.487e+04 291
34 16 2.802e+06 0.6% 29,972 1.121e+04 4.396e+04 291
33 31 2.784e+06 0.6% 85,427 1.101e+04 4.904e+04 290
21 64 2.765e+06 0.6% 99,908 1.488e+04 2.905e+04 291
19 140 2.732e+06 0.6% 94,000 1.539e+04 3.134e+04 291
32 33 2.713e+06 0.6% 31,850 9,768 4.87e+04 291
37 29 2.641e+06 0.6% 58,816 8,606 6.216e+04 291
20 90 2.6e+06 0.6% 94,650 1.503e+04 2.821e+04 285
12 130 2.593e+06 0.6% 77,857 1.557e+04 2.356e+04 288
34 126 2.55e+06 0.6% 61,542 8,325 3.534e+04 272
31 140 2.529e+06 0.6% 77,873 1.219e+04 3.738e+04 288
18 118 2.455e+06 0.5% 102,366 1.575e+04 2.803e+04 291
35 14 2.432e+06 0.5% 81,285 1.277e+04 5.07e+04 291
35 251 2.401e+06 0.5% 79,029 1.031e+04 4.798e+04 291
35 91 2.386e+06 0.5% 40,224 6,439 4.38e+04 263
21 166 2.356e+06 0.5% 122,150 1.759e+04 2.568e+04 291
37 35 2.285e+06 0.5% 68,902 8,994 5.171e+04 291
8 221 2.273e+06 0.5% 106,773 1.581e+04 1.668e+04 291
14 201 2.229e+06 0.5% 109,619 1.769e+04 2.473e+04 291
20 102 2.217e+06 0.5% 74,960 1.194e+04 2.303e+04 274
20 227 2.21e+06 0.5% 49,228 8,055 1.463e+04 254
25 47 2.098e+06 0.5% 136,593 1.684e+04 3.191e+04 291
36 158 2.073e+06 0.5% 101,871 1.277e+04 3.968e+04 291
8 90 2.011e+06 0.4% 95,404 1.54e+04 2.144e+04 288
27 73 2e+06 0.4% 90,299 1.257e+04 2.849e+04 291
24 162 1.96e+06 0.4% 69,491 1.113e+04 1.872e+04 291
12 162 1.952e+06 0.4% 81,633 1.244e+04 1.665e+04 288
38 173 1.926e+06 0.4% 81,659 1.13e+04 3.723e+04 291
13 12 1.873e+06 0.4% 91,569 1.378e+04 2.208e+04 287
7 81 1.847e+06 0.4% 141,121 2.295e+04 2.13e+04 291
35 207 1.838e+06 0.4% 45,341 6,872 3.779e+04 291
18 78 1.82e+06 0.4% 80,736 1.161e+04 2.376e+04 291
26 121 1.804e+06 0.4% 87,390 1.237e+04 2.524e+04 291
28 176 1.801e+06 0.4% 102,743 1.376e+04 3.027e+04 291
9 64 1.766e+06 0.4% 95,732 1.396e+04 2.034e+04 289
24 130 1.746e+06 0.4% 61,790 1.102e+04 2.215e+04 283
18 117 1.743e+06 0.4% 116,607 1.583e+04 2.37e+04 291
34 134 1.743e+06 0.4% 45,613 6,793 3.785e+04 290
36 155 1.711e+06 0.4% 70,064 9,192 3.682e+04 291
38 175 1.66e+06 0.4% 72,486 9,986 4.502e+04 291
8 227 1.655e+06 0.4% 50,283 8,548 1.101e+04 265
21 1 1.651e+06 0.4% 100,031 1.676e+04 2.482e+04 291
8 102 1.63e+06 0.4% 74,760 1.243e+04 1.565e+04 269
23 53 1.623e+06 0.4% 94,602 1.259e+04 2.416e+04 291
36 72 1.615e+06 0.4% 40,197 5,712 3.909e+04 290
26 201 1.597e+06 0.4% 85,152 1.336e+04 2.463e+04 291
33 78 1.564e+06 0.3% 44,666 7,287 3.068e+04 290
37 201 1.542e+06 0.3% 42,356 5,937 3.829e+04 289
38 56 1.532e+06 0.3% 59,495 7,683 4.186e+04 291
27 218 1.52e+06 0.3% 100,079 1.301e+04 2.709e+04 291
23 134 1.497e+06 0.3% 103,499 1.415e+04 2.471e+04 291
7 124 1.474e+06 0.3% 103,309 1.474e+04 1.815e+04 291
19 133 1.449e+06 0.3% 82,662 1.246e+04 2.629e+04 291
21 226 1.448e+06 0.3% 86,361 1.108e+04 1.759e+04 291
34 50 1.439e+06 0.3% 69,010 8,009 3.713e+04 291
22 160 1.43e+06 0.3% 44,169 9,775 2.816e+04 291
14 58 1.425e+06 0.3% 95,491 1.294e+04 1.658e+04 291
22 124 1.397e+06 0.3% 125,105 1.565e+04 2.364e+04 291
5 210 1.388e+06 0.3% 114,126 1.905e+04 1.633e+04 291
30 117 1.374e+06 0.3% 95,439 1.21e+04 2.574e+04 291
37 64 1.357e+06 0.3% 74,281 8,826 3.664e+04 291
28 74 1.325e+06 0.3% 41,537 1.066e+04 2.5e+04 291
33 76 1.3e+06 0.3% 37,658 5,876 2.771e+04 287
11 53 1.275e+06 0.3% 90,712 1.246e+04 1.979e+04 291
17 75 1.236e+06 0.3% 89,181 1.148e+04 1.96e+04 291
22 186 1.232e+06 0.3% 75,382 9,868 2.067e+04 288
13 187 1.209e+06 0.3% 78,597 1.059e+04 1.408e+04 291
36 118 1.192e+06 0.3% 63,124 7,149 3.143e+04 291
34 118 1.186e+06 0.3% 44,555 5,821 3.194e+04 291
9 166 1.182e+06 0.3% 125,383 1.664e+04 1.53e+04 291
36 110 1.163e+06 0.3% 28,569 4,732 3.955e+04 291
26 163 1.154e+06 0.3% 68,267 1.221e+04 2.336e+04 291
25 12 1.148e+06 0.3% 69,283 9,899 1.873e+04 290
21 76 1.133e+06 0.2% 63,936 9,690 1.587e+04 276
16 183 1.108e+06 0.2% 90,927 1.265e+04 2.157e+04 291
30 118 1.087e+06 0.2% 59,775 8,642 2.326e+04 291
15 145 1.06e+06 0.2% 81,772 1.126e+04 1.843e+04 291
22 45 1.05e+06 0.2% 77,924 1.014e+04 2.233e+04 291
30 248 1.033e+06 0.2% 66,140 8,938 2.002e+04 281
23 80 1.03e+06 0.2% 44,996 9,590 2.197e+04 291
20 198 1.026e+06 0.2% 81,354 1.049e+04 1.822e+04 291
37 55 1.024e+06 0.2% 86,061 1.031e+04 2.452e+04 291
16 135 1.01e+06 0.2% 54,869 7,856 1.112e+04 285
26 58 9.976e+05 0.2% 73,381 1.021e+04 1.591e+04 291
11 134 9.974e+05 0.2% 99,941 1.338e+04 1.733e+04 291

Inferred Domain Associations

For the top 40 global experts, this heat map shows what fraction of each expert's importance proxy came from each inferred task domain.

Residency Implication

For this SciCode cluster, a conservative residency policy should start from layer-expert pairs, not just global expert IDs. Global expert rankings tell us which expert IDs repeatedly matter, but the strongest concentrations occur at specific layers. A practical next experiment is to build candidate sets at 50%, 80%, 90%, and 95% importance mass, then run ablation or forced CPU-residency comparisons against exact-match/code-eval quality and throughput.