{ "status": "PARTIAL_ADAPTIVE_2024_DEVELOPMENT_NOT_FINAL_BENCHMARK", "closed_episodes": 46, "planned_episodes": 192, "models": [ { "id": "astra", "name": "GPT-6 Astra / Codex", "short": "Astra", "color": "#b96b45", "origination": 15.323054173420122, "refinement": 2.30432544251736, "collaboration": 0.0, "transfer": -2.062771485806252, "n": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "overall": 21.441889649632593, "uncertainty": { "origination": { "by_repetition": [ 15.323054173420122, null, null ], "n": 1, "mean": 15.323054173420122, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 4.037780813912036, 0.0, 2.875195513640037 ], "n": 3, "mean": 2.304325442517358, "sample_sd": 2.0785422665600874, "ci95": [ -2.8590597873115633, 7.4677106723462785 ], "confidence": 0.95, "ci_status": "estimated" }, "collaboration": { "by_repetition": [ 0.0, null, null ], "n": 1, "mean": 0.0, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ -2.062771485806252, null, null ], "n": 1, "mean": -2.062771485806252, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": 21.441889649632593, "ci95": null, "axes_observed": 4, "n_by_axis": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 36.169609947324204, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 33.47775607138285, "sample_sd": 2.3312138398408204, "ci95": [ 27.686699857679972, 39.26881228508572 ], "confidence": 0.95, "ci_status": "estimated" }, { "by_repetition": [ 36.169609947324204, 32.13182913341217, 35.007024647052205 ], "n": 3, "mean": 34.43615457592953, "sample_sd": 2.0785422665600874, "ci95": [ 29.272769346100606, 39.59953980575845 ], "confidence": 0.95, "ci_status": "estimated" } ], "paired": { "collaboration": { "control": { "by_repetition": [ 36.169609947324204, null, null ], "n": 1, "mean": 36.169609947324204, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 36.169609947324204, null, null ], "n": 1, "mean": 36.169609947324204, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ 35.307363526324316, null, null ], "n": 1, "mean": 35.307363526324316, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 33.244592040518064, null, null ], "n": 1, "mean": 33.244592040518064, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ 33.244592040518064, null, null ], "n": 1, "mean": 33.244592040518064, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "sol", "name": "GPT-5.6 Sol / Codex", "short": "Sol", "color": "#936aa0", "origination": null, "refinement": 1.2998330560373734, "collaboration": null, "transfer": 0.23029724692441533, "n": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 1 }, "overall": null, "uncertainty": { "origination": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 1.2998330560373788, null, null ], "n": 1, "mean": 1.2998330560373788, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "collaboration": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ 0.23029724692441533, null, null ], "n": 1, "mean": 0.23029724692441533, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": null, "ci95": null, "axes_observed": 2, "n_by_axis": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 1 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 33.43166218944955, null, null ], "n": 1, "mean": 33.43166218944955, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 33.43166218944955, null, null ], "n": 1, "mean": 33.43166218944955, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "paired": { "collaboration": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ 35.45898471344746, null, null ], "n": 1, "mean": 35.45898471344746, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 35.68928196037187, null, null ], "n": 1, "mean": 35.68928196037187, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "luna", "name": "GPT-5.6 Luna / Codex", "short": "Luna", "color": "#38694e", "origination": 33.578974383041114, "refinement": 0.8205820282589684, "collaboration": -0.32538469250043534, "transfer": 0.6713463526974977, "n": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "overall": 47.86561653718136, "uncertainty": { "origination": { "by_repetition": [ 33.578974383041114, null, null ], "n": 1, "mean": 33.578974383041114, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 0.0, 2.4617460847769124, 0.0 ], "n": 3, "mean": 0.8205820282589708, "sample_sd": 1.4212897647224576, "ci95": [ -2.7100974756128404, 4.351261532130782 ], "confidence": 0.95, "ci_status": "estimated" }, "collaboration": { "by_repetition": [ -0.32538469250043534, null, null ], "n": 1, "mean": -0.32538469250043534, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ 0.6713463526974977, null, null ], "n": 1, "mean": 0.6713463526974977, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": 47.86561653718136, "ci95": null, "axes_observed": 4, "n_by_axis": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 34.59357521818908, 32.13182913341217 ], "n": 3, "mean": 32.95241116167114, "sample_sd": 1.4212897647224576, "ci95": [ 29.421731657799327, 36.483090665542946 ], "confidence": 0.95, "ci_status": "estimated" }, { "by_repetition": [ 32.13182913341217, 34.59357521818908, 32.13182913341217 ], "n": 3, "mean": 32.95241116167114, "sample_sd": 1.4212897647224576, "ci95": [ 29.421731657799327, 36.483090665542946 ], "confidence": 0.95, "ci_status": "estimated" } ], "paired": { "collaboration": { "control": { "by_repetition": [ 35.248046953171134, null, null ], "n": 1, "mean": 35.248046953171134, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 34.9226622606707, null, null ], "n": 1, "mean": 34.9226622606707, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 32.803175486109666, null, null ], "n": 1, "mean": 32.803175486109666, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ 32.384962148158145, null, null ], "n": 1, "mean": 32.384962148158145, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "gpt55", "name": "GPT-5.5 / Codex", "short": "GPT-5.5", "color": "#818a80", "origination": 11.014647919342258, "refinement": 3.1198794797357365, "collaboration": null, "transfer": null, "n": { "origination": 1, "refinement": 1, "collaboration": 0, "transfer": 0 }, "overall": null, "uncertainty": { "origination": { "by_repetition": [ 11.014647919342258, null, null ], "n": 1, "mean": 11.014647919342258, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 3.1198794797357365, null, null ], "n": 1, "mean": 3.1198794797357365, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "collaboration": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": null, "ci95": null, "axes_observed": 2, "n_by_axis": { "origination": 1, "refinement": 1, "collaboration": 0, "transfer": 0 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "paired": { "collaboration": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ 37.88188455829507, null, null ], "n": 1, "mean": 37.88188455829507, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "fable", "name": "Claude Fable 5.1 / Claude Code", "short": "Fable 5.1", "color": "#327f89", "origination": 13.540757647268508, "refinement": 0.0, "collaboration": null, "transfer": null, "n": { "origination": 1, "refinement": 1, "collaboration": 0, "transfer": 0 }, "overall": null, "uncertainty": { "origination": { "by_repetition": [ 13.540757647268508, null, null ], "n": 1, "mean": 13.540757647268508, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 0.0, null, null ], "n": 1, "mean": 0.0, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "collaboration": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": null, "ci95": null, "axes_observed": 2, "n_by_axis": { "origination": 1, "refinement": 1, "collaboration": 0, "transfer": 0 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "paired": { "collaboration": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "opus", "name": "Claude Opus 5 / Claude Code", "short": "Opus 5", "color": "#a38637", "origination": null, "refinement": 0.0, "collaboration": null, "transfer": null, "n": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 0 }, "overall": null, "uncertainty": { "origination": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 0.0, null, null ], "n": 1, "mean": 0.0, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "collaboration": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": null, "ci95": null, "axes_observed": 1, "n_by_axis": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 0 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "paired": { "collaboration": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "sonnet", "name": "Claude Sonnet 5 / Claude Code", "short": "Sonnet 5", "color": "#647b9c", "origination": -16.866721887994867, "refinement": 0.16657491656600598, "collaboration": -2.0443231073844026, "transfer": 0.028209666103506947, "n": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "overall": -25.78362314475603, "uncertainty": { "origination": { "by_repetition": [ -16.866721887994867, null, null ], "n": 1, "mean": -16.866721887994867, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 0.0, 0.0, 0.49972474969802505 ], "n": 3, "mean": 0.16657491656600834, "sample_sd": 0.2885162187588731, "ci95": [ -0.5501391029045164, 0.883288936036533 ], "confidence": 0.95, "ci_status": "estimated" }, "collaboration": { "by_repetition": [ -2.0443231073844004, null, null ], "n": 1, "mean": -2.0443231073844004, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ 0.028209666103506947, null, null ], "n": 1, "mean": 0.028209666103506947, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": -25.78362314475603, "ci95": null, "axes_observed": 4, "n_by_axis": { "origination": 1, "refinement": 3, "collaboration": 1, "transfer": 1 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.63155388311019 ], "n": 3, "mean": 32.298404049978174, "sample_sd": 0.2885162187588731, "ci95": [ 31.581690030507648, 33.0151180694487 ], "confidence": 0.95, "ci_status": "estimated" } ], "paired": { "collaboration": { "control": { "by_repetition": [ 36.091854902376646, null, null ], "n": 1, "mean": 36.091854902376646, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 34.047531794992246, null, null ], "n": 1, "mean": 34.047531794992246, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ 32.160038799515675, null, null ], "n": 1, "mean": 32.160038799515675, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } }, { "id": "haiku", "name": "Claude Haiku 4.5 / Claude Code", "short": "Haiku 4.5", "color": "#ad7c85", "origination": null, "refinement": 0.0, "collaboration": null, "transfer": null, "n": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 0 }, "overall": null, "uncertainty": { "origination": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "refinement": { "by_repetition": [ 0.0, null, null ], "n": 1, "mean": 0.0, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "collaboration": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "transfer": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "overall": { "mean": null, "ci95": null, "axes_observed": 1, "n_by_axis": { "origination": 0, "refinement": 1, "collaboration": 0, "transfer": 0 }, "ci_status": "not_estimable", "method": "Equal-weight axis means; conservative Bonferroni t interval; no missing-axis imputation." }, "curves": [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "paired": { "collaboration": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } }, "transfer": { "control": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "treatment": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, "unrelated": { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } } } } } ], "curve_values": [ [ 32.13182913341217, 33.47775607138285, 34.43615457592953 ], [ 32.13182913341217, 33.43166218944955, 33.43166218944955 ], [ 32.13182913341217, 32.95241116167114, 32.95241116167114 ], [ 35.251708613147905, 35.251708613147905, 35.251708613147905 ], [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], [ 32.13182913341217, 32.13182913341217, 32.298404049978174 ], [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ] ], "checkpoints_calls": [ 1, 2, 4 ], "seed_pp": 32.13182913341217, "paired": { "collaboration": { "control": [ 36.169609947324204, null, 35.248046953171134, null, null, null, 36.091854902376646, null ], "treatment": [ 36.169609947324204, null, 34.9226622606707, null, null, null, 34.047531794992246, null ], "unrelated": [ null, null, null, null, null, null, null, null ] }, "transfer": { "control": [ 35.307363526324316, 35.45898471344746, 32.13182913341217, 37.88188455829507, null, null, 32.13182913341217, null ], "treatment": [ 33.244592040518064, 35.68928196037187, 32.803175486109666, null, null, null, 32.160038799515675, null ], "unrelated": [ 33.244592040518064, 32.13182913341217, 32.384962148158145, null, null, null, 32.13182913341217, null ] } }, "source_sha256": { "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_13.json": "5b1d45be283c72e461b98bd4b25cc19dda0352c50819d3d1fd9fe5db612d9cae", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_RECOVERED_CHECKPOINTS.json": "122c15e53e2c7278f308aba0279a51972c1f6554b8d5fe0c614ead5e91508513", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_CLOSED_RECOVERY_SELECTIONS.json": "4cd24732f4c0790d1e956ac9c474b921554aecdc0ff24ba92a4d24fff52db9dd", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_ORIGINATION_R1_CLOSED.json": "287cca5197339f8eb049b19460fddf35644a6f3ebf64306c0b70259c400b63e6", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_SOURCE_R1_CLOSED.json": "2204b752660268a088a720c55c7d7d284f77862730247f1f07c99b26aea734e4", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_TRANSFER_NONE_R1_CLOSED.json": "3578c6a2d550e45e8ca4605556e0965dc68a92ad9df577b0521f0340a636724d", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_TRANSFER_UNRELATED_R1_CLOSED.json": "066ebd523ec977a1c77b8d7588e6e0a1b8b313085873076c3a4b4721c861cb1d", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_TRANSFER_USEFUL_R1_CLOSED.json": "abd440e4efcd757a3669527860052004e64cd440b32085093c314bcc47c54caf", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_FABLE_ORIGINATION_R1_CLOSED.json": "2ac073a04228d62003a713acfeece9864191047ef074faa1a6cf199cdc8772f0", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_ORIGINATION_R1_CLOSED.json": "96e3d38dacc2a29440819d2f7529b89dd1b5ccaf2c353530abddba972eb19aa6", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_ORIGINATION_R1_CLOSED.json": "0184d4a844ab1c1261d8b392b8cd24af4d6420a8a8a9d9c6a369f0988daadea4", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_OPUS_ORIGINATION_R1_CLOSED.json": "a05f61e94996cb194280d06cb7080642d1598987b31795c2d1b95e72550f410a", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_ORIGINATION_R1_CLOSED.json": "730c9efad855ac18f279b1a53ca10bc2ce0068953a62a9471b85e6e3c60ca0fe", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_COLLABORATION_ISOLATED_R1_CLOSED.json": "f277d5b68e7fb9337d1e71c4c233a727ac8844e924128f1a7a6d625f3f77347c", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_ASTRA_COLLABORATION_SHARING_R1_CLOSED.json": "eefaefc04b4473fe65c27a595f67f8b56698a53e6fe84a1785938ccb153262cf", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_SOURCE_R1_CLOSED.json": "df4919eeaaba14d7a02689b394e7466e1336cb5e40c2f86ba93ec5f4c23db492", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_TRANSFER_NONE_R1_CLOSED.json": "c1dd25c4bd8d7b3a65fbc51d179d0845b832a91088f3521989a21480f57b487f", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_TRANSFER_USEFUL_R1_CLOSED.json": "0d2c34d5a6acfa646a8064bfd95946ef21d24a8cec98a4e2ce3c17e65cb2e431", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_TRANSFER_UNRELATED_R1_CLOSED.json": "4fbe3dc297d9be2c813a6ae2f3837da5059db6ef3bf2ae3f8e77d0dd3f6ef44d", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_COLLABORATION_ISOLATED_R1_CLOSED.json": "6853a655298d1c988e5f7feaafb9ee4af5f4d8236cc4e6154906d7970a5f1951", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_LUNA_COLLABORATION_SHARING_R1_CLOSED.json": "486381b6590ccb056590759d1657d47bc879df3055f32c3a72c09ab2e2380eaa", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_SOURCE_R1_CLOSED.json": "492c07ea23aedfe57d98b4352225250b75a5faef8739c12cf5a8d6f256d94b08", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_TRANSFER_NONE_R1_CLOSED.json": "73dbd0f168a53491d8e56977754d17be6c3f24f66690276884dddf4df7bea012", "/Users/louisdebenoist/trading/alpha-foundry/reports/DEMO_20260906_CLOSED_CHECKPOINTS_22.json": "bd71a774cf2b0f9cc586dd28034dc44528704cba7db29a85fece10682ec47e7b", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_TRANSFER_USEFUL_R1_CLOSED.json": "0288ef65378eed3d3f71939ec9be3d96dc02d4e32012b0c84bfd0260f16256ce", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_14.json": "60cda0117961e8d8d515202b91a9feebb20137c5f773650784c3766c513b1a53", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_TRANSFER_UNRELATED_R1_CLOSED.json": "4496c77f36987d03b8513342b02ebbac3a144678c83809b8709504adae98d18d", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_TRANSFER_R1_PAIRED_DEVELOPMENT_01.json": "e3e08ac74a3c38ba638ff51453a1be76aef2000dae644948912caac559b134de", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_15.json": "96206e8d9430138d6eeeb47f90e04eb600c952a67b6733d198a40016b45596b8", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_COLLABORATION_ISOLATED_R1_CLOSED.json": "055872214a9f932d5c0d792e3c4da70624c5856bea5b4a57df3aea49f604d835", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_ISOLATED_R1_CLOSED_REVISION_SUMMARY.json": "2376ba90ed63981103e3dd16b12d82c8bbee14824142740ec6f7a1735d430ddd", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_16.json": "696d62fe86616f363a5ce7a0c902d232e8e03ef4d4ba2cbf5249601badb1606f", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_SHARING_R1_INTERIM_REVISION_01.json": "e61449a9b727d597a317b052f98544ce39781f4c6b1d12f74da8cb90a1e33194", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_COLLABORATION_SHARING_R1_OBSERVATION_05.json": "b776f88eba3882c6b20519b8f462bae32077b0027b92ce7eb796cec44796c4d7", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_COLLABORATION_R1_PAIRED_DEVELOPMENT.json": "45158e14b46c2118bd1d4a9a031a6821e47cb218024b62f9ae7918ac26e32483", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SONNET_COLLABORATION_SHARING_R1_CLOSED.json": "a24fdc3d043c402711187ea42b95a238e8045e1bafb5a383d8102a0f5b7a9633", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_17.json": "9c554e4928cfef0930a76901877cd9e8a7f5f3846026d4200d94c694df6df8a4", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_FABLE_REFINEMENT_R1_CLOSED.json": "06caa7c1ce95cf9fab030b6421374641dace2e07d9dd4f62f0940d1e0515043a", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_18.json": "96e29a8c7c39618fc66151bd6e92c48c81492ab91fb6a4f395d1e54b5ac5d9ea", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_REFINEMENT_R1_CLOSED.json": "d441d8eb7338f34771829b8c5f7d4c067a240c698f2af73ca78e033a4425e429", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_19.json": "46efce76727089e002cc5ee357c259065dd09e0da97912bc5b0eb980c323f8bd", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_REFINEMENT_R1_CLOSED.json": "019e65b8928bccf3c48a946692a126fe41c285ac4672a9c4924c4f0662d29e3a", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_20.json": "3a5db0067d2e0d37f4b39d655cd105c88448ee868223a4a5934c0c33c49f027e", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_OPUS_REFINEMENT_R1_CLOSED.json": "512bf96c22fa3cfd8f644497c2bc9aa29498ba0df2f5dbbd6b46f977b2adc77c", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_21.json": "32be0ce5ce27e79d2bacb3cabfc540c7b6cb362f23b151017bea3333360c6bce", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_HAIKU_REFINEMENT_R1_CLOSED.json": "57bcecbe8988afa2a465bf80b3a9d8631542d9504b57fbef96c337f0f7650137", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_22.json": "fbf141fd821cc116e9d6d0c79cc4971f2b0e49da49ebafe4b4c60a8acdf90633", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_HAIKU_REFINEMENT_R1_THRESHOLD_TRACE_01.json": "4d1cb1e3e0c296f066465594ed843d84691cb6ecece2fd8fb30e844885d2a5fb", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_HAIKU_REFINEMENT_R1_SEED_RESTORATION_TRACE_01.json": "79fd2813ae3e12522ab252e04a9f5da08028a595b87ec2f7475fbd133d5424b0", "/Users/louisdebenoist/trading/alpha-foundry/reports/faros_uncertainty.py": "b2d5978bbcf6f4af87baeabaf682d3c13e4882f0efbf557e129eb36e6c32bae6", "/Users/louisdebenoist/trading/alpha-foundry/reports/enrich_faros_development.py": "1c571f370943b9df9444fd5b82d02bd2b23de0f739f446fcb7a3c7a1248e9015", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_23.json": "034b7f955b07ceb708d7c3d28b60003ca1f3f57b69104c1ace55441275bb091a", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_SOURCE_R1_CLOSED.json": "b6b4a92cc5a8a8fdd6e89f2a79140a930c16de6b3a7780b6543edfa35d0f8fec", "/Users/louisdebenoist/trading/alpha-foundry/private_results/operator_snapshots/faros-source-notes-v1/sol-source-notes-r1/receipt.json": "4247b73c6b6f7c1c38c2c892c9b1bc5a30866fed7ccbbcd6714065e40519e519", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_24.json": "bb18350c77cbbd597a9bf50bef9f59f248d214dbf791366363738e2bdfaf62a0", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_TRANSFER_NONE_R1_CLOSED.json": "c296ad23084a0a3bafde4667ee082cade7f73caa7aa4d233c1a65a1992f3d420", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_TRANSFER_NONE_R1_ABLATION_TRACE_01.json": "bc314dc20b0824d152dd2522f075e95b0cbeffcb9d8ba0d5d89be988fb52befc", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_25.json": "62a71693d578ff302c731015f273ecc6144bb7b68fb0a578ecc7d249c40e6da1", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_TRANSFER_USEFUL_R1_CLOSED.json": "855dff0e6cb8363c87d59569b92b1a6628a72a866736b2822704e08c516260be", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_TRANSFER_USEFUL_R1_INTERIM_CODE_COMPARISON_01.json": "e42c272dc6f0c7ad2eb749071fb2b99e9dc2e6bc34c430943d320e0fe0000d0f", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_26.json": "f159bb032f2bdc53bb46f76899ae6d3d0da4d0f3a662c55c64713b56f8fb260e", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_SOL_TRANSFER_UNRELATED_R1_CLOSED.json": "16f6cb916bb0d80dbd6bfa8615b93ea68fdae2e7b62ce2f10a7540deb3341f71", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_27.json": "c449f16f4812338a90ea10ac9c04bf28f4ad813c26e731ce00fb2eac57f5f7ed", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_SOURCE_R1_CLOSED.json": "f4be497fd52606f7e067430cf599427946043e3f194faa660c6a0ee69c5cad34", "/Users/louisdebenoist/trading/alpha-foundry/private_results/operator_snapshots/faros-source-notes-v1/gpt55-source-notes-r1/receipt.json": "b7450c6e3df777247a8408f24bbe42735e5392a43012b78eb0c4edbacfacc972", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_BATCH_RECORDS_28.json": "94d170780ba87c4154448874e8da3eba7a6dbeb90e27b3ec089e2adfe0745f30", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_TRANSFER_NONE_R1_CLOSED.json": "c3e02ab2512adbcd3c6c1a7f2a4eed3c37a71b74a23ffc23c6473b0bc4b2d08c", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_TRANSFER_NONE_R1_ABLATION_TRACE_01.json": "e78703058a871f3a076213830275befd166d5b3f488f9c0f7ec615369df16e5e", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_TRANSFER_NONE_R1_INTERIM_TRADEOFF_01.json": "f627221df32c3bc6212d307289a6267e774b58eaf4c933ef10287173557fbad3", "/Users/louisdebenoist/trading/alpha-foundry/reports/FAROS_GPT55_TRANSFER_NONE_R1_INTERIM_TRADEOFF_02.json": "aebe52dd86abb1eb581e68bdb1771fd91388ed1052b7f8739a68509d28ff8290" }, "notes": "Astra, Luna and Sonnet refinement means use three reused repetitions; Fable, Sol, GPT-5.5, Opus and Haiku refinement and other available comparisons use repetition 1 only. Luna transfer notes empty; its contrast is not evidence of knowledge transfer. Final held-out scores pending. Ongoing overall uses all four available axis means; missing axes are not imputed. Sonnet transfer source notes also empty; its contrast is not evidence of knowledge transfer. Sonnet collaboration retains uneven sharing charges (3/2/2/1), one pre-admission failure and the isolated tune/stop caveat; not a balanced-allocation causal estimate. Sol source r1 produced no transfer notes. Its three target conditions are complete; the source-note contrast is not evidence of knowledge transfer, and selected strategies differ in their tie-breaker salt. GPT-5.5 source r1 closed with four calibration failures and no source notes; no source score is imputed.", "curve_statistics": [ [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 36.169609947324204, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 33.47775607138285, "sample_sd": 2.3312138398408204, "ci95": [ 27.686699857679972, 39.26881228508572 ], "confidence": 0.95, "ci_status": "estimated" }, { "by_repetition": [ 36.169609947324204, 32.13182913341217, 35.007024647052205 ], "n": 3, "mean": 34.43615457592953, "sample_sd": 2.0785422665600874, "ci95": [ 29.272769346100606, 39.59953980575845 ], "confidence": 0.95, "ci_status": "estimated" } ], [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 33.43166218944955, null, null ], "n": 1, "mean": 33.43166218944955, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 33.43166218944955, null, null ], "n": 1, "mean": 33.43166218944955, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 34.59357521818908, 32.13182913341217 ], "n": 3, "mean": 32.95241116167114, "sample_sd": 1.4212897647224576, "ci95": [ 29.421731657799327, 36.483090665542946 ], "confidence": 0.95, "ci_status": "estimated" }, { "by_repetition": [ 32.13182913341217, 34.59357521818908, 32.13182913341217 ], "n": 3, "mean": 32.95241116167114, "sample_sd": 1.4212897647224576, "ci95": [ 29.421731657799327, 36.483090665542946 ], "confidence": 0.95, "ci_status": "estimated" } ], [ { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.251708613147905, null, null ], "n": 1, "mean": 35.251708613147905, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], [ { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.13182913341217 ], "n": 3, "mean": 32.13182913341217, "sample_sd": 0.0, "ci95": [ 32.13182913341217, 32.13182913341217 ], "confidence": 0.95, "ci_status": "zero_observed_variation" }, { "by_repetition": [ 32.13182913341217, 32.13182913341217, 32.63155388311019 ], "n": 3, "mean": 32.298404049978174, "sample_sd": 0.2885162187588731, "ci95": [ 31.581690030507648, 33.0151180694487 ], "confidence": 0.95, "ci_status": "estimated" } ], [ { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ] ], "paired_statistics": { "collaboration": { "control": [ { "by_repetition": [ 36.169609947324204, null, null ], "n": 1, "mean": 36.169609947324204, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.248046953171134, null, null ], "n": 1, "mean": 35.248046953171134, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 36.091854902376646, null, null ], "n": 1, "mean": 36.091854902376646, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "treatment": [ { "by_repetition": [ 36.169609947324204, null, null ], "n": 1, "mean": 36.169609947324204, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 34.9226622606707, null, null ], "n": 1, "mean": 34.9226622606707, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 34.047531794992246, null, null ], "n": 1, "mean": 34.047531794992246, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "unrelated": [ { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ] }, "transfer": { "control": [ { "by_repetition": [ 35.307363526324316, null, null ], "n": 1, "mean": 35.307363526324316, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.45898471344746, null, null ], "n": 1, "mean": 35.45898471344746, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 37.88188455829507, null, null ], "n": 1, "mean": 37.88188455829507, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "treatment": [ { "by_repetition": [ 33.244592040518064, null, null ], "n": 1, "mean": 33.244592040518064, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 35.68928196037187, null, null ], "n": 1, "mean": 35.68928196037187, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.803175486109666, null, null ], "n": 1, "mean": 32.803175486109666, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.160038799515675, null, null ], "n": 1, "mean": 32.160038799515675, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ], "unrelated": [ { "by_repetition": [ 33.244592040518064, null, null ], "n": 1, "mean": 33.244592040518064, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.384962148158145, null, null ], "n": 1, "mean": 32.384962148158145, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ 32.13182913341217, null, null ], "n": 1, "mean": 32.13182913341217, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" }, { "by_repetition": [ null, null, null ], "n": 0, "mean": null, "sample_sd": null, "ci95": null, "confidence": 0.95, "ci_status": "not_estimable" } ] } }, "uncertainty": { "confidence": 0.95, "unit": "independent research repetitions on the same historical market task", "method": "Pointwise Student t intervals on completed repetitions; n<2 is not estimable. Overall uses equal-weight axis means and conservative Bonferroni marginal t bounds.", "limitations": "Exploratory, fixed-sample intervals, not anytime-valid or adjusted for all model/plot comparisons. Small-sample normality and independent-run assumptions are unverified. They do not measure future trading uncertainty or establish alpha. Zero observed variation is not proof of certainty.", "source": "https://itl.nist.gov/div898/handbook/eda/section3/eda352.htm" }, "overall_status": "ONGOING_DEVELOPMENT_ALL_FOUR_AXES_REQUIRED_NOT_FINAL" }