# Results tables

Generated by `code/analyze.py`. Flag rates use cuts that flag at most 6 of 124 length-matched human texts.

## Human controls

| set | n | desklib | fakespot | Binoculars | Fast-DetectGPT | flagged by any |
|---|---|---|---|---|---|---|
| internal | 70 | 0.0% | 1.4% | 4.3% | 5.7% | 10.0% |
| federal | 54 | 11.1% | 9.3% | 5.6% | 3.7% | 24.1% |
| all | 124 | 4.8% | 4.8% | 4.8% | 4.8% | 16.1% |

Cuts (oriented so higher = more AI; Binoculars is negated): desklib 0.9962, fakespot 0.9980, Binoculars -0.9118, Fast-DetectGPT 1.4871

## Set 2026-11 (2026-11)

| tool | n | desklib flagged | fakespot flagged | Binoculars flagged | Fast-DetectGPT flagged | passes all four | judge mean | judge >= 4 | NLI fwd | NLI bwd | facts changed | facts changed, no quotes | passthrough |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| reright (run 5, v15) | 33 | 36% | 27% | 15% | 6% | 55% | 4.33 | 76% | 98% | 98% | 5% of 21 | 5% of 21 | 3% |
| reright (run 1) | 33 | 45% | 42% | 15% | 21% | 39% (-33..+0) | 4.33 (-0.42..+0.42) | 73% | 96% | 96% | 14% of 21 (-10..+29) | 10% of 21 | 15% |
| Rephrasy | 33 | 24% | 15% | 0% | 0% | 70% (-12..+39) | 2.48 (-2.21..-1.48) | 12% | 65% | 68% | 43% of 21 (+19..+62) | 38% of 21 | - |
| SuperHumanizer | 33 | 36% | 12% | 9% | 3% | 58% (-18..+24) | 3.36 (-1.48..-0.42) | 39% | 89% | 89% | 33% of 21 (+5..+52) | 29% of 21 | - |
| UndetectedGPT | 33 | 39% | 18% | 24% | 27% | 48% (-27..+15) | 4.12 (-0.78..+0.34) | 69% | 92% | 95% | 38% of 21 (+14..+52) | 29% of 21 | - |
| CleverHumanizer | 33 | 24% | 6% | 42% | 21% | 45% (-30..+9) | 3.12 (-1.67..-0.73) | 36% | 76% | 96% | 38% of 21 (+14..+52) | 29% of 21 | - |
| SmartHumanizer | 33 | 36% | 33% | 12% | 15% | 45% (-27..+6) | 4.12 (-0.76..+0.33) | 70% | 94% | 96% | 38% of 21 (+14..+52) | 33% of 21 | - |
| reright (run 2) | 33 | 48% | 39% | 18% | 12% | 42% (-27..+0) | 4.70 (+0.03..+0.73) | 88% | 98% | 97% | 5% of 21 (-14..+14) | 0% of 21 | 21% |
| HIX Bypass | 33 | 58% | 39% | 21% | 15% | 30% (-42..-6) | 4.39 (-0.48..+0.61) | 76% | 96% | 94% | 38% of 21 (+14..+57) | 29% of 21 | - |
| GPTinf | 33 | 39% | 45% | 27% | 18% | 27% (-45..-9) | 2.45 (-2.24..-1.48) | 3% | 77% | 94% | 43% of 21 (+19..+57) | 38% of 21 | - |
| ReHumanize | 33 | 79% | 73% | 55% | 42% | 9% (-64..-27) | 4.67 (-0.06..+0.76) | 88% | 92% | 95% | 38% of 21 (+14..+52) | 33% of 21 | - |
| StealthGPT | 33 | 85% | 55% | 70% | 61% | 3% (-67..-33) | 3.00 (-1.79..-0.88) | 27% | 70% | 81% | 43% of 21 (+14..+62) | 33% of 21 | - |
| Undetectable AI | 33 | 97% | 79% | 88% | 76% | 3% (-70..-33) | 4.09 (-0.70..+0.21) | 73% | 90% | 83% | 38% of 21 (+14..+52) | 38% of 21 | - |
| reright (run 3, v14) | 33 | 39% | 39% | 12% | 12% | 48% (-21..+6) | 4.30 (-0.48..+0.42) | 73% | 98% | 97% | 0% of 21 (-14..+0) | 0% of 21 | 3% |
| reright (run 4, v14) | 33 | 24% | 39% | 9% | 9% | 48% (-24..+9) | 4.16 (-0.56..+0.28) | 69% | 97% | 97% | 5% of 21 (-14..+14) | 5% of 21 | 6% |
| reright (run 6, v15) | 33 | 36% | 30% | 15% | 6% | 52% (-15..+9) | 4.12 (-0.61..+0.12) | 67% | 96% | 97% | 5% of 21 (+0..+0) | 5% of 21 | 3% |
| Unmodified AI input | 33 | 94% | 85% | 45% | 30% | 3% (-70..-33) | - | - | - | - | 0% of 21 (-14..+0) | 0% of 21 | - |

Intervals: paired bootstrap 95% of (row minus reright (run 5, v15)), in points for shares and score units for the judge.

## Set 2026-10 (2026-10)

| tool | n | desklib flagged | fakespot flagged | Binoculars flagged | Fast-DetectGPT flagged | passes all four | judge mean | judge >= 4 | NLI fwd | NLI bwd | facts changed | facts changed, no quotes | passthrough |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| reright (run 5, v15) | 33 | 27% | 39% | 21% | 12% | 48% | 4.48 | 82% | 97% | 98% | 0% of 20 | 0% of 19 | 6% |
| reright (run 1) | 33 | 42% | 61% | 24% | 15% | 27% (-39..-3) | 4.48 (-0.36..+0.33) | 79% | 98% | 97% | 15% of 20 (+0..+30) | 5% of 19 | 24% |
| HIX Bypass | 33 | 9% | 3% | 0% | 0% | 88% (+21..+58) | 3.79 (-1.15..-0.24) | 58% | 99% | 92% | 35% of 20 (+15..+55) | 11% of 19 | - |
| AI Humanize | 33 | 12% | 18% | 18% | 15% | 61% (-9..+33) | 2.42 (-2.42..-1.67) | 9% | 71% | 92% | 65% of 20 (+45..+85) | 47% of 19 | - |
| SuperHumanizer | 33 | 33% | 15% | 3% | 0% | 58% (-12..+30) | 3.64 (-1.33..-0.36) | 48% | 88% | 88% | 35% of 20 (+15..+55) | 21% of 19 | - |
| WriteHuman | 33 | 21% | 18% | 39% | 42% | 42% (-27..+15) | 2.09 (-2.67..-2.09) | 6% | 45% | 65% | 65% of 20 (+45..+85) | 58% of 19 | - |
| UndetectedGPT | 33 | 42% | 9% | 27% | 9% | 39% (-33..+15) | 3.91 (-1.03..-0.12) | 61% | 88% | 92% | 35% of 20 (+15..+55) | 21% of 19 | - |
| reright (run 2) | 33 | 48% | 48% | 30% | 18% | 33% (-36..+6) | 4.48 (-0.33..+0.33) | 79% | 96% | 96% | 20% of 20 (+5..+40) | 5% of 19 | 15% |
| SmartHumanizer | 33 | 52% | 42% | 18% | 6% | 27% (-42..+3) | 3.88 (-1.03..-0.15) | 55% | 92% | 97% | 40% of 20 (+20..+60) | 26% of 19 | - |
| GPTinf | 33 | 42% | 55% | 24% | 9% | 21% (-48..-6) | 2.36 (-2.48..-1.73) | 9% | 70% | 91% | 45% of 20 (+25..+65) | 26% of 19 | - |
| CleverHumanizer | 33 | 52% | 42% | 70% | 67% | 15% (-52..-15) | 2.12 (-2.67..-2.03) | 3% | 56% | 65% | 50% of 20 (+25..+70) | 37% of 19 | - |
| ReHumanize | 33 | 79% | 79% | 42% | 24% | 6% (-61..-24) | 4.73 (+0.00..+0.48) | 91% | 92% | 95% | 45% of 20 (+25..+65) | 21% of 19 | - |
| StealthGPT | 33 | 73% | 48% | 79% | 70% | 6% (-61..-24) | 3.24 (-1.76..-0.76) | 39% | 69% | 81% | 55% of 20 (+30..+75) | 53% of 19 | - |
| Undetectable AI | 33 | 94% | 79% | 82% | 79% | 0% (-67..-30) | 4.15 (-0.76..+0.09) | 70% | 91% | 82% | 55% of 20 (+35..+75) | 32% of 19 | - |
| reright (run 3, v14) | 33 | 36% | 42% | 18% | 12% | 45% (-15..+9) | 4.24 (-0.67..+0.15) | 76% | 98% | 97% | 10% of 20 (+0..+25) | 11% of 19 | 6% |
| reright (run 4, v14) | 33 | 39% | 45% | 18% | 12% | 36% (-30..+6) | 4.45 (-0.39..+0.33) | 79% | 97% | 98% | 5% of 20 (+0..+15) | 5% of 19 | 9% |
| reright (run 6, v15) | 33 | 45% | 30% | 21% | 12% | 48% (-15..+15) | 4.09 (-0.79..+0.00) | 73% | 96% | 96% | 0% of 20 (+0..+0) | 0% of 19 | 6% |
| Unmodified AI input | 33 | 91% | 85% | 64% | 36% | 0% (-67..-30) | - | - | - | - | 0% of 20 (+0..+0) | 0% of 19 | - |

Intervals: paired bootstrap 95% of (row minus reright (run 5, v15)), in points for shares and score units for the judge.

## Set 2026-09 (2026-09)

| tool | n | desklib flagged | fakespot flagged | Binoculars flagged | Fast-DetectGPT flagged | passes all four | judge mean | judge >= 4 | NLI fwd | NLI bwd | facts changed | facts changed, no quotes | passthrough |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| reright (run 5, v15) | 33 | 45% | 27% | 12% | 9% | 48% | 4.42 | 79% | 97% | 97% | 0% of 26 | 0% of 25 | 3% |
| reright (run 1) | 33 | 64% | 39% | 33% | 12% | 24% (-42..-9) | 4.64 (-0.12..+0.55) | 91% | 97% | 96% | 15% of 26 (+4..+31) | 0% of 25 | 24% |
| WriteHuman | 33 | 9% | 6% | 6% | 9% | 76% (+6..+48) | 2.48 (-2.36..-1.45) | 15% | 70% | 82% | 46% of 26 (+27..+65) | 32% of 25 | - |
| HIX Bypass | 33 | 21% | 6% | 0% | 0% | 73% (+3..+45) | 3.42 (-1.27..-0.73) | 36% | 97% | 92% | 35% of 26 (+15..+54) | 16% of 25 | - |
| AI Humanize | 33 | 24% | 15% | 21% | 6% | 52% (-18..+24) | 2.45 (-2.39..-1.52) | 12% | 74% | 91% | 42% of 26 (+23..+62) | 40% of 25 | - |
| SuperHumanizer | 33 | 45% | 15% | 12% | 6% | 52% (-21..+27) | 3.45 (-1.42..-0.48) | 42% | 86% | 82% | 35% of 26 (+15..+54) | 16% of 25 | - |
| SmartHumanizer | 33 | 55% | 3% | 3% | 0% | 39% (-33..+15) | 4.27 (-0.67..+0.36) | 73% | 97% | 96% | 27% of 26 (+12..+46) | 16% of 25 | - |
| UndetectedGPT | 33 | 52% | 12% | 3% | 3% | 36% (-30..+9) | 3.91 (-1.06..+0.00) | 58% | 93% | 96% | 19% of 26 (+4..+35) | 8% of 25 | - |
| CleverHumanizer | 33 | 42% | 27% | 67% | 52% | 27% (-42..+3) | 2.42 (-2.36..-1.61) | 6% | 57% | 63% | 50% of 26 (+31..+69) | 40% of 25 | - |
| GPTinf | 33 | 58% | 36% | 30% | 9% | 27% (-42..+0) | 2.58 (-2.21..-1.48) | 15% | 72% | 93% | 38% of 26 (+19..+58) | 24% of 25 | - |
| StealthGPT | 33 | 67% | 36% | 48% | 36% | 21% (-45..-12) | 3.00 (-1.91..-0.97) | 27% | 65% | 76% | 31% of 26 (+15..+50) | 24% of 25 | - |
| reright (run 2) | 33 | 73% | 48% | 30% | 15% | 18% (-48..-12) | 4.61 (-0.18..+0.58) | 88% | 98% | 96% | 19% of 26 (+4..+35) | 0% of 25 | 24% |
| Undetectable AI | 33 | 97% | 79% | 70% | 52% | 3% (-61..-27) | 4.36 (-0.48..+0.36) | 76% | 96% | 87% | 23% of 26 (+8..+38) | 8% of 25 | - |
| ReHumanize | 33 | 91% | 82% | 67% | 70% | 0% (-64..-30) | 4.85 (+0.12..+0.76) | 97% | 92% | 95% | 38% of 26 (+19..+58) | 24% of 25 | - |
| reright (run 3, v14) | 33 | 45% | 18% | 12% | 9% | 55% (-9..+21) | 4.15 (-0.64..+0.12) | 67% | 98% | 98% | 0% of 26 (+0..+0) | 0% of 25 | 3% |
| reright (run 4, v14) | 33 | 30% | 18% | 12% | 12% | 67% (+3..+36) | 4.45 (-0.36..+0.42) | 79% | 98% | 97% | 0% of 26 (+0..+0) | 0% of 25 | 9% |
| reright (run 6, v15) | 33 | 36% | 21% | 18% | 12% | 61% (-3..+27) | 4.48 (-0.33..+0.45) | 82% | 98% | 98% | 0% of 26 (+0..+0) | 0% of 25 | 3% |
| Unmodified AI input | 33 | 100% | 85% | 67% | 30% | 0% (-64..-30) | - | - | - | - | 0% of 26 (+0..+0) | 0% of 25 | - |

Intervals: paired bootstrap 95% of (row minus reright (run 5, v15)), in points for shares and score units for the judge.

## Set pooled (2026-09, 2026-10, 2026-11)

| tool | n | desklib flagged | fakespot flagged | Binoculars flagged | Fast-DetectGPT flagged | passes all four | judge mean | judge >= 4 | NLI fwd | NLI bwd | facts changed | facts changed, no quotes | passthrough |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| reright (run 5, v15) | 99 | 36% | 31% | 16% | 9% | 51% | 4.41 | 79% | 97% | 98% | 1% of 67 | 2% of 65 | 4% |
| reright (run 1) | 99 | 51% | 47% | 24% | 16% | 30% (-31..-10) | 4.48 (-0.15..+0.29) | 81% | 97% | 96% | 15% of 67 (+4..+22) | 5% of 65 | 21% |
| HIX Bypass | 99 | 29% | 16% | 7% | 5% | 64% (+1..+25) | 3.87 (-0.80..-0.27) | 57% | 97% | 93% | 36% of 67 (+24..+46) | 18% of 65 | - |
| SuperHumanizer | 99 | 38% | 14% | 8% | 3% | 56% (-7..+18) | 3.48 (-1.21..-0.66) | 43% | 88% | 86% | 34% of 67 (+21..+45) | 22% of 65 | - |
| UndetectedGPT | 99 | 44% | 13% | 18% | 13% | 41% (-21..+3) | 3.98 (-0.72..-0.13) | 62% | 91% | 94% | 30% of 67 (+18..+39) | 18% of 65 | - |
| SmartHumanizer | 99 | 47% | 26% | 11% | 7% | 37% (-26..-1) | 4.09 (-0.62..-0.03) | 66% | 94% | 97% | 34% of 67 (+22..+43) | 25% of 65 | - |
| reright (run 2) | 99 | 57% | 45% | 26% | 15% | 31% (-29..-9) | 4.60 (-0.03..+0.39) | 85% | 97% | 96% | 15% of 67 (+4..+22) | 2% of 65 | 20% |
| CleverHumanizer | 99 | 39% | 25% | 60% | 46% | 29% (-33..-9) | 2.56 (-2.09..-1.61) | 15% | 63% | 75% | 46% of 67 (+33..+57) | 35% of 65 | - |
| GPTinf | 99 | 46% | 45% | 27% | 12% | 25% (-38..-13) | 2.46 (-2.16..-1.74) | 9% | 73% | 92% | 42% of 67 (+30..+52) | 29% of 65 | - |
| StealthGPT | 99 | 75% | 46% | 66% | 56% | 10% (-51..-30) | 3.08 (-1.62..-1.07) | 31% | 68% | 80% | 42% of 67 (+28..+52) | 35% of 65 | - |
| ReHumanize | 99 | 83% | 78% | 55% | 45% | 5% (-56..-35) | 4.75 (+0.13..+0.54) | 92% | 92% | 95% | 40% of 67 (+27..+51) | 26% of 65 | - |
| Undetectable AI | 99 | 96% | 79% | 80% | 69% | 2% (-59..-38) | 4.20 (-0.46..+0.04) | 73% | 92% | 84% | 37% of 67 (+25..+48) | 25% of 65 | - |
| reright (run 3, v14) | 99 | 40% | 33% | 14% | 11% | 49% (-9..+7) | 4.23 (-0.43..+0.05) | 72% | 98% | 97% | 3% of 67 (-3..+6) | 3% of 65 | 4% |
| reright (run 4, v14) | 99 | 31% | 34% | 13% | 11% | 51% (-10..+10) | 4.36 (-0.28..+0.18) | 76% | 98% | 98% | 3% of 67 (-3..+6) | 3% of 65 | 8% |
| reright (run 6, v15) | 99 | 39% | 27% | 18% | 10% | 54% (-6..+12) | 4.23 (-0.41..+0.04) | 74% | 97% | 97% | 1% of 67 (+0..+0) | 2% of 65 | 4% |
| Unmodified AI input | 99 | 95% | 85% | 59% | 32% | 1% (-60..-39) | - | - | - | - | 0% of 67 (-4..+0) | 0% of 65 | - |

Intervals: paired bootstrap 95% of (row minus reright (run 5, v15)), in points for shares and score units for the judge.