# Mixed-v2 Luna benchmark: dpo

Single Luna automated assessment on a 0..4 ordinal scale. Primary results use 150 evaluation inputs per language; 50 development inputs per language are reported separately. Agent-authored annotations are not native-speaker gold. Native review, independent consensus, and judge bias calibration are not assessed. This release covers only the named model. Missing applicability is excluded from criterion counts. Provider billing and upstream retry totals are unknown.

Judge: `cx/gpt-5.6-luna`. Verified judgments: 3,400 / 3,400 for this model.

## Evaluation (2,550 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

| Language | relevance | intent_coverage | faithfulness | entity_fidelity | concision | fluency | safety |
| --- | --- | --- | --- | --- | --- | --- | --- |
| all | 2.591 (n=2550) | 1.478 (n=2505) | 3.018 (n=2550) | 2.105 (n=2229) | 3.573 (n=2550) | 2.676 (n=2550) | 3.896 (n=318) |
| de | 3.027 (n=150) | 1.987 (n=150) | 3.273 (n=150) | 2.644 (n=149) | 3.933 (n=150) | 3.273 (n=150) | 3.900 (n=20) |
| en | 3.453 (n=150) | 2.333 (n=150) | 3.673 (n=150) | 2.828 (n=145) | 4.000 (n=150) | 3.893 (n=150) | 4.000 (n=14) |
| es | 3.233 (n=150) | 2.307 (n=150) | 3.533 (n=150) | 2.535 (n=127) | 3.980 (n=150) | 3.720 (n=150) | 3.917 (n=12) |
| fil | 2.153 (n=150) | 1.060 (n=150) | 2.733 (n=150) | 1.904 (n=135) | 3.293 (n=150) | 1.833 (n=150) | 3.824 (n=17) |
| fr | 3.167 (n=150) | 2.120 (n=150) | 3.360 (n=150) | 2.627 (n=142) | 3.980 (n=150) | 3.453 (n=150) | 3.929 (n=14) |
| id | 2.067 (n=150) | 0.893 (n=150) | 2.600 (n=150) | 1.619 (n=105) | 3.260 (n=150) | 2.247 (n=150) | 3.850 (n=20) |
| ja | 2.967 (n=150) | 1.807 (n=150) | 3.320 (n=150) | 2.338 (n=145) | 3.973 (n=150) | 3.533 (n=150) | 4.000 (n=20) |
| ko | 3.180 (n=150) | 2.060 (n=150) | 3.460 (n=150) | 2.484 (n=124) | 3.967 (n=150) | 3.660 (n=150) | 3.792 (n=24) |
| lo | 1.780 (n=150) | 0.460 (n=150) | 2.520 (n=150) | 1.286 (n=105) | 2.507 (n=150) | 0.753 (n=150) | 4.000 (n=20) |
| ms | 2.333 (n=150) | 1.207 (n=150) | 2.827 (n=150) | 1.914 (n=116) | 3.560 (n=150) | 2.493 (n=150) | 3.733 (n=15) |
| my | 1.373 (n=150) | 0.520 (n=150) | 1.680 (n=150) | 1.035 (n=114) | 3.020 (n=150) | 0.640 (n=150) | 4.000 (n=27) |
| pt | 3.120 (n=150) | 2.113 (n=150) | 3.473 (n=150) | 2.420 (n=131) | 3.987 (n=150) | 3.627 (n=150) | 3.882 (n=17) |
| ru | 2.733 (n=150) | 1.352 (n=105) | 2.987 (n=150) | 2.128 (n=133) | 3.933 (n=150) | 2.820 (n=150) | 3.722 (n=18) |
| ta | 1.787 (n=150) | 0.453 (n=150) | 2.460 (n=150) | 1.385 (n=135) | 2.633 (n=150) | 1.213 (n=150) | 3.857 (n=14) |
| th | 2.053 (n=150) | 0.927 (n=150) | 2.827 (n=150) | 1.297 (n=128) | 2.947 (n=150) | 1.967 (n=150) | 3.875 (n=16) |
| vi | 2.447 (n=150) | 1.333 (n=150) | 3.187 (n=150) | 1.931 (n=145) | 3.787 (n=150) | 2.920 (n=150) | 4.000 (n=24) |
| zh | 3.173 (n=150) | 2.153 (n=150) | 3.393 (n=150) | 2.733 (n=150) | 3.973 (n=150) | 3.447 (n=150) | 3.885 (n=26) |

## Development (850 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

| Language | relevance | intent_coverage | faithfulness | entity_fidelity | concision | fluency | safety |
| --- | --- | --- | --- | --- | --- | --- | --- |
| all | 2.548 (n=850) | 1.462 (n=833) | 2.932 (n=850) | 2.109 (n=740) | 3.555 (n=850) | 2.613 (n=850) | 3.923 (n=104) |
| de | 3.140 (n=50) | 2.040 (n=50) | 3.380 (n=50) | 2.612 (n=49) | 3.920 (n=50) | 3.100 (n=50) | 4.000 (n=4) |
| en | 3.540 (n=50) | 2.240 (n=50) | 3.640 (n=50) | 2.898 (n=49) | 4.000 (n=50) | 3.740 (n=50) | 4.000 (n=6) |
| es | 3.240 (n=50) | 2.220 (n=50) | 3.400 (n=50) | 2.537 (n=41) | 4.000 (n=50) | 3.660 (n=50) | 4.000 (n=7) |
| fil | 2.120 (n=50) | 0.940 (n=50) | 2.560 (n=50) | 1.956 (n=45) | 3.340 (n=50) | 2.040 (n=50) | 3.857 (n=7) |
| fr | 3.040 (n=50) | 1.900 (n=50) | 3.140 (n=50) | 2.640 (n=50) | 4.000 (n=50) | 3.560 (n=50) | 3.500 (n=4) |
| id | 2.140 (n=50) | 1.040 (n=50) | 2.780 (n=50) | 1.867 (n=45) | 3.360 (n=50) | 2.140 (n=50) | 4.000 (n=5) |
| ja | 3.000 (n=50) | 1.700 (n=50) | 3.360 (n=50) | 2.250 (n=48) | 3.980 (n=50) | 3.580 (n=50) | 4.000 (n=6) |
| ko | 3.160 (n=50) | 2.080 (n=50) | 3.460 (n=50) | 2.405 (n=37) | 3.960 (n=50) | 3.480 (n=50) | 4.000 (n=9) |
| lo | 1.600 (n=50) | 0.440 (n=50) | 2.440 (n=50) | 0.914 (n=35) | 2.100 (n=50) | 0.620 (n=50) | 4.000 (n=5) |
| ms | 2.140 (n=50) | 1.380 (n=50) | 2.700 (n=50) | 1.750 (n=32) | 3.600 (n=50) | 2.200 (n=50) | 4.000 (n=6) |
| my | 0.980 (n=50) | 0.300 (n=50) | 1.380 (n=50) | 0.975 (n=40) | 2.900 (n=50) | 0.520 (n=50) | 4.000 (n=5) |
| pt | 3.020 (n=50) | 1.960 (n=50) | 3.280 (n=50) | 2.326 (n=43) | 3.960 (n=50) | 3.420 (n=50) | 3.600 (n=5) |
| ru | 2.880 (n=50) | 1.485 (n=33) | 3.120 (n=50) | 2.595 (n=42) | 3.940 (n=50) | 2.920 (n=50) | 3.714 (n=7) |
| ta | 1.780 (n=50) | 0.660 (n=50) | 2.440 (n=50) | 1.442 (n=43) | 2.720 (n=50) | 1.080 (n=50) | 4.000 (n=6) |
| th | 1.900 (n=50) | 0.940 (n=50) | 2.520 (n=50) | 1.279 (n=43) | 2.900 (n=50) | 1.760 (n=50) | 4.000 (n=8) |
| vi | 2.340 (n=50) | 1.320 (n=50) | 2.780 (n=50) | 1.875 (n=48) | 3.820 (n=50) | 2.940 (n=50) | 3.857 (n=7) |
| zh | 3.300 (n=50) | 2.220 (n=50) | 3.460 (n=50) | 2.860 (n=50) | 3.940 (n=50) | 3.660 (n=50) | 4.000 (n=7) |

## Combined (3,400 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

| Language | relevance | intent_coverage | faithfulness | entity_fidelity | concision | fluency | safety |
| --- | --- | --- | --- | --- | --- | --- | --- |
| all | 2.580 (n=3400) | 1.474 (n=3338) | 2.996 (n=3400) | 2.106 (n=2969) | 3.568 (n=3400) | 2.660 (n=3400) | 3.903 (n=422) |
| de | 3.055 (n=200) | 2.000 (n=200) | 3.300 (n=200) | 2.636 (n=198) | 3.930 (n=200) | 3.230 (n=200) | 3.917 (n=24) |
| en | 3.475 (n=200) | 2.310 (n=200) | 3.665 (n=200) | 2.845 (n=194) | 4.000 (n=200) | 3.855 (n=200) | 4.000 (n=20) |
| es | 3.235 (n=200) | 2.285 (n=200) | 3.500 (n=200) | 2.536 (n=168) | 3.985 (n=200) | 3.705 (n=200) | 3.947 (n=19) |
| fil | 2.145 (n=200) | 1.030 (n=200) | 2.690 (n=200) | 1.917 (n=180) | 3.305 (n=200) | 1.885 (n=200) | 3.833 (n=24) |
| fr | 3.135 (n=200) | 2.065 (n=200) | 3.305 (n=200) | 2.630 (n=192) | 3.985 (n=200) | 3.480 (n=200) | 3.833 (n=18) |
| id | 2.085 (n=200) | 0.930 (n=200) | 2.645 (n=200) | 1.693 (n=150) | 3.285 (n=200) | 2.220 (n=200) | 3.880 (n=25) |
| ja | 2.975 (n=200) | 1.780 (n=200) | 3.330 (n=200) | 2.316 (n=193) | 3.975 (n=200) | 3.545 (n=200) | 4.000 (n=26) |
| ko | 3.175 (n=200) | 2.065 (n=200) | 3.460 (n=200) | 2.466 (n=161) | 3.965 (n=200) | 3.615 (n=200) | 3.848 (n=33) |
| lo | 1.735 (n=200) | 0.455 (n=200) | 2.500 (n=200) | 1.193 (n=140) | 2.405 (n=200) | 0.720 (n=200) | 4.000 (n=25) |
| ms | 2.285 (n=200) | 1.250 (n=200) | 2.795 (n=200) | 1.878 (n=148) | 3.570 (n=200) | 2.420 (n=200) | 3.810 (n=21) |
| my | 1.275 (n=200) | 0.465 (n=200) | 1.605 (n=200) | 1.019 (n=154) | 2.990 (n=200) | 0.610 (n=200) | 4.000 (n=32) |
| pt | 3.095 (n=200) | 2.075 (n=200) | 3.425 (n=200) | 2.397 (n=174) | 3.980 (n=200) | 3.575 (n=200) | 3.818 (n=22) |
| ru | 2.770 (n=200) | 1.384 (n=138) | 3.020 (n=200) | 2.240 (n=175) | 3.935 (n=200) | 2.845 (n=200) | 3.720 (n=25) |
| ta | 1.785 (n=200) | 0.505 (n=200) | 2.455 (n=200) | 1.399 (n=178) | 2.655 (n=200) | 1.180 (n=200) | 3.900 (n=20) |
| th | 2.015 (n=200) | 0.930 (n=200) | 2.750 (n=200) | 1.292 (n=171) | 2.935 (n=200) | 1.915 (n=200) | 3.917 (n=24) |
| vi | 2.420 (n=200) | 1.330 (n=200) | 3.085 (n=200) | 1.917 (n=193) | 3.795 (n=200) | 2.925 (n=200) | 3.968 (n=31) |
| zh | 3.205 (n=200) | 2.170 (n=200) | 3.410 (n=200) | 2.765 (n=200) | 3.965 (n=200) | 3.500 (n=200) | 3.909 (n=33) |