Configuration: Claude Mythos 5.1 in Claude Code 2.1.289, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 24% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 8.0 to 42.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Mythos 5.1 at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Mythos 5.1' tier 'mid' effort 'high', column 'Construct design % [95% CI]' Configuration: Claude Mythos 5.1 in Claude Code 2.1.289, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 16% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 4.0 to 29.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Mythos 5.1 at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Mythos 5.1' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Claude Mythos 5.1 in Claude Code 2.1.289, max reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 17.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 5.3 to 30.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 13, and 13/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Mythos 5.1 at max effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Mythos 5.1' tier 'top' effort 'max', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Claude Mythos 5.1', column 'Construct design %' Configuration: Claude Opus 5.5 in Claude Code 2.1.289, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 34.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 17.3 to 53.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Opus 5.5 at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Opus 5.5' tier 'mid' effort 'high', column 'Construct design % [95% CI]' Configuration: Claude Opus 5.5 in Claude Code 2.1.289, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 29.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 13.3 to 46.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Opus 5.5 at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Opus 5.5' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Claude Opus 5.5 in Claude Code 2.1.289, max reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 25.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 9.3 to 44.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 19, and 19/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Claude Opus 5.5 at max effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Claude Opus 5.5' tier 'top' effort 'max', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Claude Opus 5.5', column 'Construct design %' Configuration: Gemini 3.8 Flash in Gemini CLI 0.63.0, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 24% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 8.0 to 42.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 18, and 18/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 3.8 Flash at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 3.8 Flash' tier 'top' effort 'high', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Gemini 3.8 Flash', column 'Construct design %' Configuration: Gemini 3.8 Flash in Gemini CLI 0.63.0, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 22.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 5.3 to 42.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 3.8 Flash at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 3.8 Flash' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Gemini 3.8 Flash in Gemini CLI 0.63.0, medium reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 22.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 5.3 to 42.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 3.8 Flash at medium effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 3.8 Flash' tier 'mid' effort 'medium', column 'Construct design % [95% CI]' Configuration: Gemini 4 Argon in Gemini CLI 0.63.0, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 30.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 10.7 to 52.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 23, and 23/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 4 Argon at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'top' effort 'high', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Gemini 4 Argon', column 'Construct design %' Configuration: Gemini 4 Argon in Gemini CLI 0.63.0, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 31.9% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 10.7 to 52.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 4 Argon at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Gemini 4 Argon in Gemini CLI 0.63.0, medium reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 33.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 13.3 to 54.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Gemini 4 Argon at medium effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'mid' effort 'medium', column 'Construct design % [95% CI]' Configuration: GPT-6.1 Sol in Codex CLI 0.156.1, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 37.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 20.0 to 56.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6.1 Sol at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6.1 Sol' tier 'mid' effort 'high', column 'Construct design % [95% CI]' Configuration: GPT-6.1 Sol in Codex CLI 0.156.1, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 22.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 9.3 to 38.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6.1 Sol at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6.1 Sol' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: GPT-6.1 Sol in Codex CLI 0.156.1, max reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 41.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 22.7 to 61.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 31, and 31/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6.1 Sol at max effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6.1 Sol' tier 'top' effort 'max', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'GPT-6.1 Sol', column 'Construct design %' Configuration: GPT-6 Astra in Codex CLI 0.156.1, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 44% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 22.7 to 66.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6 Astra at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6 Astra' tier 'mid' effort 'high', column 'Construct design % [95% CI]' Configuration: GPT-6 Astra in Codex CLI 0.156.1, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 28% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 13.3 to 45.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6 Astra at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6 Astra' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: GPT-6 Astra in Codex CLI 0.156.1, max reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 42.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 22.7 to 64.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 32, and 32/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source GPT-6 Astra at max effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'GPT-6 Astra' tier 'top' effort 'max', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'GPT-6 Astra', column 'Construct design %' Configuration: Grok 4.7 in Grok Build 1.0.46, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 22.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 6.7 to 41.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Grok 4.7 at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Grok 4.7' tier 'mid' effort 'high', column 'Construct design % [95% CI]' Configuration: Grok 4.7 in Grok Build 1.0.46, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 10.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 2.7 to 21.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Grok 4.7 at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Grok 4.7' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Grok 4.7 in Grok Build 1.0.46, xhigh reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 26.7% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 9.3 to 46.7. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 20, and 20/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Grok 4.7 at xhigh effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Grok 4.7' tier 'top' effort 'xhigh', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Grok 4.7', column 'Construct design %' Configuration: Kimi K3 in Pi 1.0.3, high reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 20% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; top reasoning tier, the highest of this system's three settings)percent · higher
Uncertainty: 95% CI 6.7 to 36.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts. The Table G2 per-task solve counts for this system sum to 15, and 15/75 reproduces the printed rate.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Kimi K3 at high effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Kimi K3' tier 'top' effort 'high', column 'Construct design % [95% CI]'; the same cell is printed in Table 2, row 'Kimi K3', column 'Construct design %' Configuration: Kimi K3 in Pi 1.0.3, low reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 16% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; low reasoning tier, the lowest of this system's three settings)percent · higher
Uncertainty: 95% CI 2.7 to 33.3. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Kimi K3 at low effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Kimi K3' tier 'low' effort 'low', column 'Construct design % [95% CI]' Configuration: Kimi K3 in Pi 1.0.3, medium reasoning effort Benchmark: BenchBench-Cloning construct-design set Dataset subset: BenchBench-Cloning construct-design tasks (15) 29.3% Success rate (construct-design pass rate at a single attempt, pass at a final score of at least 0.9, deterministic grading on five tasks and deterministic multiplied by a GPT-6 Astra rubric score on ten; mid reasoning tier, the middle of this system's three settings)percent · higher
Uncertainty: 95% CI 12.0 to 48.0. 95% confidence intervals from 20,000 bootstrap resamples of the tasks, retaining all attempts of each sampled task (Section 3.4).
Coverage: 75/75 attempts scored (15 tasks x 5 attempts.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Table D2's construct-design rate for Gemini 4 Argon at low effort, 31.9%, cannot be a mean of binary outcomes over the stated 15 tasks times 5 attempts: no integer numerator over 75 falls in its rounding window [23.8875, 23.9625], while all 23 other printed construct rates do. Its printed 95% interval [10.7, 52.0] is also byte-identical to the same system's top-effort interval. The two signals together indicate a transcription fault in the paper rather than a measured value, and the paper prints nothing that resolves it: the lane's aggregate is 15.8 under either 31.9 or 32.0, and Table G2's per-task solves are top-effort only. The printed value is kept as published and this entry withholds it from automatic comparison; the lane's other results are not covered.BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Gemini 4 Argon' tier 'low' effort 'low', column 'Construct design % [95% CI]'
Methods, coverage and source Kimi K3 at medium effort on BenchBench-Cloning construct-design pass rate
benchbench-cloning-20261106-protocol-construct-design-pass-rate
Aggregation: Not reported
BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development (arXiv:2610.33355v1) · Table D2, row 'Kimi K3' tier 'mid' effort 'medium', column 'Construct design % [95% CI]'