[ { "id": "IF01", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the briefing body, headings, and chart labels use English. Verification method: check whether the title page, body, chart titles, and table labels are all in English. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 3: Create an eight-slide English research briefing.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G01" ], "benchmark_source": "vas" }, { "id": "IF02", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: results and discussion are based on the supplied BERT paper v2, without substituting later BERT variants or online benchmark scores. Verification method: check that the cited paper version, dataset results, and numbers all come from the supplied v2 materials, with no later variant names or new online leaderboard scores substituting. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 5: Work offline; do not substitute later BERT variants or online benchmark scores..", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G02" ], "benchmark_source": "vas" }, { "id": "IF03", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: explain the motivation for bidirectional pretraining. Verification method: check whether the body has an explicit statement explaining why bidirectional pretraining is needed, rather than merely listing the model structure. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 11: The motivation for bidirectional pretraining.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G03" ], "benchmark_source": "vas" }, { "id": "IF04", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: explain how pretraining connects to downstream fine-tuning. Verification method: check whether the body explains the connection via the process in which pretraining yields representations/parameters, which are then adapted to downstream tasks through fine-tuning. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 11: how pretraining connects to downstream fine-tuning.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G03" ], "benchmark_source": "vas" }, { "id": "IF05", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: state which parameters are updated during fine-tuning, and point out that all pretrained parameters are updated. Verification method: check whether the fine-tuning explanation states that all pretrained parameters are updated end-to-end, rather than training only the newly added output layer. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 11: including which parameters are updated.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G03" ], "benchmark_source": "vas" }, { "id": "IF06", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: report the BERT_BASE configuration: L=12, H=768, A=12, total parameters 110M. Verification method: check whether all four items of the Base architecture—number of layers, hidden dimension, attention heads, and total parameter count—appear correctly in the slide. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 333: BERTBASE (L=12, H=768, A=12, Total Param- eters=110M).", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G04" ], "benchmark_source": "vas" }, { "id": "IF07", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: report the BERT_LARGE configuration: L=24, H=1024, A=16, total parameters 340M. Verification method: check whether all four items of the Large architecture—number of layers, hidden dimension, attention heads, and total parameter count—appear correctly in the slide. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 334: BERTLARGE (L=24, H=1024, A=16, Total Parameters=340M).", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G05" ], "benchmark_source": "vas" }, { "id": "IF08", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the MLM diagram must be a TeX-editable graphic, not an embedded image or a paper screenshot. Verification method: check whether the source of the MLM diagram is TikZ or another TeX graphic structure, and check that no raster image carries this diagram. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 13: draw an editable diagram.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G06" ], "benchmark_source": "vas" }, { "id": "IF09", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the MLM diagram distinguishes the selection proportion of all token positions from the replacement probability conditional on selection, and does not draw the replacement probability as a proportion of all tokens. Verification method: check whether the diagram clearly marks the selection rate of all positions and labels the three items MASK, random token, and keep original token as conditional probabilities \"after selection,\" rather than shares of all tokens. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 13: distinguishes the fraction of all token positions selected for prediction from replacement probabilities conditional on selection.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G06" ], "benchmark_source": "vas" }, { "id": "IF10", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the MLM diagram explains that the prediction target in all replacement branches is the original token. Verification method: check whether the [MASK] branch, the random token branch, and the keep original token branch all indicate prediction of the original token. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 13: Explain the prediction target in all branches..", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G06" ], "benchmark_source": "vas" }, { "id": "IF11", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: any invented token sequence is labeled as a teaching illustration, not a recorded experiment. Verification method: check each location whether there is an illustration label or equivalent nearby any invented token sequence, and that it does not imply it is a recorded experiment from the paper. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 13: Label any invented token sequence as an illustration, not a recorded experiment..", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G06" ], "benchmark_source": "vas" }, { "id": "IF12", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: explain the sampling method and labels of NSP. Verification method: check whether it explains the sampling proportion of B being the actual next sentence/random sentence, and the IsNext/NotNext labels. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 14: NSP sampling and labels.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G07" ], "benchmark_source": "vas" }, { "id": "IF13", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: NSP must not be presented as a downstream question-answering metric. Verification method: check whether NSP is discussed only as a pretraining task and is not listed among SQuAD metrics, question-answering evaluation, or question-answering result explanations. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 14: without presenting NSP as a downstream question-answering metric.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G07" ], "benchmark_source": "vas" }, { "id": "IF14", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table is a native editable TeX table, not an image. Verification method: check whether the results page source uses a TeX table environment, rather than an embedded table screenshot or external image. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 15: A native editable table.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G08" ], "benchmark_source": "vas" }, { "id": "IF15", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table includes SQuAD 1.1 BERT_BASE Single Dev with EM=80.8, F1=88.5. Verification method: check the Base Single row under the SQuAD 1.1 group, where the Dev EM and F1 values are 80.8 and 88.5 respectively. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 782: BERTBASE (Single) 80.8 88.5.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G09" ], "benchmark_source": "vas" }, { "id": "IF16", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table includes SQuAD 1.1 BERT_LARGE Single Dev with EM=84.1, F1=90.9. Verification method: check the Large Single row under the SQuAD 1.1 group, where the Dev EM and F1 values are 84.1 and 90.9 respectively. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 786: BERTLARGE (Single) 84.1 90.9.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G10" ], "benchmark_source": "vas" }, { "id": "IF17", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table includes SQuAD 1.1 BERT_LARGE Ens.+TriviaQA Test with EM=87.4, F1=93.2. Verification method: check the Large Ens.+TriviaQA row under the SQuAD 1.1 group, where the Test EM and F1 values are 87.4 and 93.2 respectively. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 796: BERTLARGE (Ens.+TriviaQA) 86.2 92.2 87.4 93.2.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G11" ], "benchmark_source": "vas" }, { "id": "IF18", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table includes SQuAD 2.0 BERT_LARGE Single Dev with EM=78.7, F1=81.9. Verification method: check the Large Single row under the SQuAD 2.0 group, where the Dev EM and F1 values are 78.7 and 81.9 respectively. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 828: BERTLARGE (Single) 78.7 81.9 80.0 83.1.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G12" ], "benchmark_source": "vas" }, { "id": "IF19", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table includes SQuAD 2.0 BERT_LARGE Single Test with EM=80.0, F1=83.1. Verification method: check the Large Single row under the SQuAD 2.0 group, where the Test EM and F1 values are 80.0 and 83.1 respectively. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Material basis (relative to inputs/): workspace/sources/paper.txt original line 828: BERTLARGE (Single) 78.7 81.9 80.0 83.1.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G12" ], "benchmark_source": "vas" }, { "id": "IF20", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: the results table explicitly identifies dataset version, data split, and model/training setting. Verification method: check whether table groups or column labels separately indicate SQuAD 1.1/2.0, Dev/Test, Base/Large/Single/Ens.+TriviaQA, and other settings. Pass criterion: this requirement is fully met; missing, incorrect, or partially met is No. Evaluation tool failure cannot be the basis for No. Verification boundary: verify against the frozen task statement and necessary original files/ground truth; the author's own statement does not substitute for artifact evidence. Task basis: prompt.md original line 15: Explicitly identify dataset version, split, model/training setting, EM and F1..", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G08" ], "benchmark_source": "vas" }, { "id": "IF21", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The results table explicitly identifies the EM and F1 metrics. Verification method: Check whether the header contains EM and F1 as two columns or equivalent explicit labels. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 15: Explicitly identify dataset version, split, model/training setting, EM and F1.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G08" ], "benchmark_source": "vas" }, { "id": "IF22", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Blank SQuAD results in the source paper remain blank; do not fill in values on your own. Verification method: Check whether the Test result cells for SQuAD 1.1 Base Single and Large Single are blank or explicitly N/A, rather than filled with speculative values. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 15: Do not fill blank source results.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G08" ], "benchmark_source": "vas" }, { "id": "IF23", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Explain why these result rows cannot serve as a controlled comparison of model size alone. Verification method: Check whether the results explanation points out confounding factors such as Single/Ensemble, additional TriviaQA data, dataset version, or split. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 15: Explain why these rows are not a controlled comparison of model size alone.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G08" ], "benchmark_source": "vas" }, { "id": "IF24", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Explain what the reported question-answering evidence supports. Verification method: Check whether there is an explicit evidence interpretation of the SQuAD results stating the scope of what they support, rather than merely restating the numbers. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 3: what the reported question-answering evidence supports.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G13" ], "benchmark_source": "vas" }, { "id": "IF25", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Include an explicit concise takeaway. Verification method: Check whether the conclusion section or final page contains a recognizable one-sentence core conclusion. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 16: A concise takeaway and at least one limitation of interpreting these selected results.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G14" ], "benchmark_source": "vas" }, { "id": "IF26", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Give at least one limitation in interpreting the selected results. Verification method: Check whether there is at least one explicit limitation, and whether that limitation is relevant to interpreting the selected SQuAD results. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 16: A concise takeaway and at least one limitation of interpreting these selected results.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G14" ], "benchmark_source": "vas" }, { "id": "IF27", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Retain the title treatment of the supplied Metropolis. Verification method: Compare against the supplied Metropolis snapshot and check whether the presentation of the title page and frame titles follows its title treatment. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Retain the supplied Metropolis title treatment, palette and spacing conventions.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G15" ], "benchmark_source": "vas" }, { "id": "IF28", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Retain the palette of the supplied Metropolis. Verification method: Compare against the snapshot and check whether the primary colors, accent colors, and background palette remain consistent. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Retain the supplied Metropolis title treatment, palette and spacing conventions.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G15" ], "benchmark_source": "vas" }, { "id": "IF29", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Retain the spacing conventions of the supplied Metropolis. Verification method: Compare against the snapshot and check whether page margins, block spacing, and title-to-body spacing do not clearly deviate from the theme conventions. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Retain the supplied Metropolis title treatment, palette and spacing conventions.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G15" ], "benchmark_source": "vas" }, { "id": "IF30", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Deliver artifact/main.tex. Verification method: Check whether artifact/main.tex exists in the delivery directory. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: Deliver `artifact/main.tex`, all locally required supporting files, `artifact/briefing.pdf`, and a short `artifact/BUILD.md` with the XeLaTeX command and font requirements.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G16" ], "benchmark_source": "vas" }, { "id": "IF31", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Deliver artifact/briefing.pdf. Verification method: Check whether artifact/briefing.pdf exists in the delivery directory. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: Deliver `artifact/main.tex`, all locally required supporting files, `artifact/briefing.pdf`, and a short `artifact/BUILD.md` with the XeLaTeX command and font requirements.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G17" ], "benchmark_source": "vas" }, { "id": "IF32", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Deliver artifact/BUILD.md. Verification method: Check whether artifact/BUILD.md exists in the delivery directory. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: Deliver `artifact/main.tex`, all locally required supporting files, `artifact/briefing.pdf`, and a short `artifact/BUILD.md` with the XeLaTeX command and font requirements.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G16" ], "benchmark_source": "vas" }, { "id": "IF33", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The delivery directory contains all local supporting files needed for the build. Verification method: Check whether the .sty files, images, or other local resources referenced by main.tex are all located in the delivery directory and do not depend on undelivered files. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: all locally required supporting files.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G18" ], "benchmark_source": "vas" }, { "id": "IF34", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Copy the seven .sty files of the supplied Metropolis into the output directory. Verification method: Check whether the output directory contains the seven .sty files of the supplied theme. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/README.md original line 5: Copy the seven .sty files into your artifact/ directory before building main.tex.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G19" ], "benchmark_source": "vas" }, { "id": "IF35", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Preserve attribution and source license notices when copying or modifying theme files. Verification method: Check whether the copied or modified theme files retain the copyright headers and the CC BY-SA 4.0 and other source license notices. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/README.md original line 5: Preserve attribution and source license notices when copying or changing theme files.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G20" ], "benchmark_source": "vas" }, { "id": "IF36", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Preserve the BERT paper attribution and do not imply that the theme license covers the paper. Verification method: Check whether the title page/source citations preserve the paper attribution, and whether the license statement does not place the paper under the theme license. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/source.json original line 10: Paper redistribution not cleared; preserve attribution. Theme license does not cover paper.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G21" ], "benchmark_source": "vas" }, { "id": "IF37", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The PDF has exactly eight pages. Verification method: Use a PDF tool to verify that the page count equals 8. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: The PDF must have exactly eight landscape 16:9 pages including the title.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G22" ], "benchmark_source": "vas" }, { "id": "IF38", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The PDF pages are landscape 16:9. Verification method: Check that the page aspect ratio is 16:9 and the orientation is landscape. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: The PDF must have exactly eight landscape 16:9 pages including the title.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G22" ], "benchmark_source": "vas" }, { "id": "IF39", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The eight pages include a title page. Verification method: Check that page 1 is a title page and the total page count is still 8. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: including the title.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G18" ], "benchmark_source": "vas" }, { "id": "IF40", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Do not use overlays that increase the page count. Verification method: Check whether the source uses \\pause, overlay specifications, or similar constructs that create multi-page overlays, and confirm that the page count was not increased by them. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: Avoid overlays and automatic section pages that increase the page count.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G22" ], "benchmark_source": "vas" }, { "id": "IF41", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Do not use automatic section pages that increase the page count. Verification method: Check whether the source sets \\AtBeginSection or similar automatic section pages, and confirm that the page count was not increased by them. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: Avoid overlays and automatic section pages that increase the page count.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G22" ], "benchmark_source": "vas" }, { "id": "IF42", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: BUILD.md gives the supplied XeLaTeX command and notes to repeat compilation when necessary. Verification method: Check whether BUILD.md records that xelatex command and a note to repeat compilation. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/README.md original line 5: xelatex -interaction=nonstopmode -halt-on-error main.tex (repeat if references require it).", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G19" ], "benchmark_source": "vas" }, { "id": "IF43", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: BUILD.md gives the font requirements. Verification method: Check whether BUILD.md lists the fonts required for the build or a font substitution explanation. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 7: with the XeLaTeX command and font requirements.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G18" ], "benchmark_source": "vas" }, { "id": "IF44", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: If font substitution is used, it must be documented. Verification method: Check the actual font settings; if the theme's default Fira-style fonts are not used, check whether BUILD.md or main.tex contains an explicit substitution explanation. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Font substitution is allowed if documented.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G23" ], "benchmark_source": "vas" }, { "id": "IF45", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The source citation identifies the paper version arXiv:1810.04805v2. Verification method: Check whether the page or footer citation contains arXiv:1810.04805v2 or an equivalent v2 version identifier. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: brief source references (paper version and relevant section, figure or table).", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G20" ], "benchmark_source": "vas" }, { "id": "IF46", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The source citation includes the relevant section, figure, or table number. Verification method: Check whether the corresponding Section/Figure/Table citation is given near the facts page, the MLM figure, and the results table. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: brief source references (paper version and relevant section, figure or table).", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G20" ], "benchmark_source": "vas" }, { "id": "IF47", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The main explanations remain editable in TeX and are not replaced by paper screenshots. Verification method: Check whether the main explanations are TeX text/formulas/structured content and are not replaced by full-page paper screenshots or images. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Main explanations, the MLM diagram and results table must remain editable in TeX; do not turn paper screenshots into slides.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G24" ], "benchmark_source": "vas" }, { "id": "IF48", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Required assets are embedded locally and do not depend on external paths or the network. Verification method: Check whether image, style, and other resource references point to files within the delivery directory, rather than absolute paths, external workspace/sources/ paths, or network URLs. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: Embed required assets locally.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G15" ], "benchmark_source": "vas" }, { "id": "IF49", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The supplied build command runs successfully from the delivery directory. Verification method: Create a working copy of the delivery in evaluation/work/ that preserves relative resource relationships, execute the given XeLaTeX build steps according to BUILD.md, check for absence of fatal errors and verify that the build result corresponds to the submitted briefing.pdf; do not modify artifact/, do not repair source code, and do not add missing files. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 18: ensure the supplied build command works from the delivered directory.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G15" ], "benchmark_source": "vas" }, { "id": "IF50", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: When text extraction loses table structure, use workspace/sources/paper.pdf as the authoritative reference for checking results. Verification method: Compare the SQuAD rows in the results table against Table 2/Table 3 of the supplied workspace/sources/paper.pdf, confirming that values and blank cells were not misled by text extraction. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Task basis: prompt.md original line 5: Use the PDF as authoritative when extraction loses table structure.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G02" ], "benchmark_source": "vas" }, { "id": "IF51", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: Input files remain read-only and must not be rewritten, deleted, or replaced. Verification method: Verify final integrity against the actually provided inputs and the frozen originals or trusted content hashes; do not infer violations from modification times or the absence of a production-process log, and do not require retention of historical working files. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/README.md original line 7 excerpt: Inputs are read-only.[…].", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G25" ], "benchmark_source": "vas" }, { "id": "IF52", "dimension": "Task Fulfillment / Compliance", "criterion": "Requirement: The build does not assume Fira Sans is installed on the system. Verification method: Check whether the font settings and BUILD.md avoid unconditional dependence on system Fira Sans, or provide a usable substitute/local font scheme when it is missing. Pass criterion: Fully satisfies this requirement; missing, incorrect, or partially satisfied is No. Evaluation tool failure cannot be grounds for No. Verification boundary: Verify based on the frozen prompt and necessary source/ground truth files; the author's self-report does not replace artifact evidence. Material basis (relative to inputs/): workspace/README.md original line 5: Font substitution is allowed; do not assume Fira Sans is installed.", "answers": { "Yes": "Complete all necessary checks for this criterion, with actual evidence supporting satisfaction of the requirement; equivalent implementations within the task's scope are allowed.", "No": "At least one necessary requirement is not satisfied; identify evidence of actual omissions, errors, or partial satisfaction." }, "compliance_groups": [ "G19" ], "benchmark_source": "vas" }, { "id": "C1", "dimension": "Content", "criterion": "Terminology consistency and audience comprehensibility.\nWhat is measured: The briefing should use terms such as pretraining, fine-tuning, bidirectional encoder, MLM, NSP, Base/Large, Single/Ensemble, and +TriviaQA consistently, and explain them on first occurrence in wording understandable to an audience who has not read the paper; it is not required to list every definition one by one.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check whether MLM, NSP, fine-tuning, Single, Ensemble, and +TriviaQA have consistent definitions or context at first occurrence. Check whether the terminology for Base/Large is written consistently across pages. Check whether the connection between bidirectional pretraining and fine-tuning is explained in wording understandable to an audience who has not read the paper.\nScoring boundary: Whether the terminology, concepts, and wording allow the intended audience to understand accurately; IF checks factual implementation, and C1 does not award points for the number of terms.\nOriginal full refinement inspection basis: The most easily confused concepts and qualifications remain concise and accurate, and there is no drift in meaning across summaries, details, and actual examples. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: Core terms are conflated so that readers cannot determine the training stage or model setting, for example describing MLM as a downstream task with no correction throughout. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Multiple key terms are repeatedly used without definition or are referred to inconsistently, clearly hindering understanding for an audience who has not read the paper. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: One term is imprecise or a key term is only implicitly explained later, causing a localized comprehension obstacle. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: Terminology is used correctly, brief explanations are given where necessary, and the audience can follow the narrative. The necessary terms, objects, units, and references in this item are accurate, and the intended audience can understand the required meanings. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: terminology is consistent throughout, and a concise comparison is given for Base/Large and Single/Ensemble. Quality improvement must be manifested as: easily confused terms, quantities, or references are briefly clarified where they appear, reducing the need for readers to guess. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: terms, symbols, and diagram labels form a unified terminology correspondence within the work, so readers do not need to look back when something recurs across pages. Definitions, qualifications, and actual labels support one another, and complex meanings do not require stitching together across sections. the terminology system is accurate and fits the pacing of a reading-group presentation, concepts are naturally anchored on first occurrence, and they are precisely reused with the same meaning later. The most easily confused concepts and qualifications remain concise and accurate, and there is no drift in meaning across summaries, details, and actual examples. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The briefing should use terms such as pretraining, fine-tuning, bidirectional encoder, MLM, NSP, Base/Large, Single/Ensemble, and +TriviaQA consistently, and explain them on first occurrence in wording understandable to an audience who has not read the paper; it is not required to list every definition one by one. Difficult-case and inspection basis: The most easily confused concepts and qualifications remain concise and accurate, and there is no drift in meaning across summaries, details, and actual examples. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. An unusually economical conceptual formulation makes a genuinely confusable distinction directly understandable, with less interpretation than an equally accurate, concise conventional explanation. Identify the distinction and the wording/representation that produces this extra understanding; merely correct terminology or fluent prose is not exceptional. Apply the exceptional benefit specifically to: The briefing should use terms such as pretraining, fine-tuning, bidirectional encoder, MLM, NSP, Base/Large, Single/Ensemble, and +TriviaQA consistently, and explain them on first occurrence in wording understandable to an audience who has not read the paper; it is not required to list every definition one by one. In-task verification cases for that benefit: The most easily confused concepts and qualifications remain concise and accurate, and there is no drift in meaning across summaries, details, and actual examples. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 9, "benchmark_source": "vas" }, { "id": "C2", "dimension": "Content", "criterion": "Paper mechanism and numerical accuracy.\nWhat is measured: The pretraining/fine-tuning mechanisms, Base/Large configurations, MLM ratio, NSP labels, and the EM/F1 in the results table must be consistent with arXiv:1810.04805v2; do not fill blank values and do not alter settings.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check cell by cell the EM/F1 for SQuAD 1.1 Base Single/Large Single Dev, Large Ens.+TriviaQA Test, and SQuAD 2.0 Large Single Dev/Test. Check the MLM 15% and 80/10/10 and the NSP 50/50 IsNext/NotNext. Check the number of layers, hidden dimension, attention heads, and parameter count for Base/Large. Check that all pretrained parameters are updated during fine-tuning and that there is a task-specific output layer.\nScoring boundary: The support relationship between task claims and evidence/data/actual process; do not re-evaluate graphic encoding itself or whether source documents are easy to find.\nOriginal full refinement inspection basis: The most complex key comparison or state chain can still be checked item by item; approximations, limitations, and counterexamples are not concealed by the conclusion. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nScenario reference · content quality (no additional threshold): pretraining/fine-tuning parameter updates, Base/Large configurations, MLM selection and replacement, NSP, the specified SQuAD 1.1/2.0 rows and their comparability limitations, and conclusions and limitations.\nApplicable boundary for this task (relevant items all apply; no additional scoring items): It is not required to put all facts from the supplied raw material on screen; only the mechanisms and result rows explicitly required by the prompt are verified.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: Core values or multiple mechanisms conflict with the paper, for example fine-tuning updates only some parameters or SQuAD values are mismatched. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: At least one specified result is wrong, attributed to the wrong item, or a blank Test cell is filled in on the author's own initiative. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: There is one deviation in a numerical value or mechanism statement that does not affect the overall narrative but is noticeable to readers and causes confusion. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: The specified facts and values are correct, with no obvious misquotation. Key claims, values, and the actual process are supported by the prompt and source files, and there is no substantive conflict regarding comparison conditions and conclusion strength. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: all specified values are accurate, and necessary qualifications are given for easily confused settings. Quality improvement must be manifested as: key claims are matched with nearby values, conditions, or process evidence used to support them, making verification more direct. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: facts, tables, and text cross-check one another, and blank items and dataset versions are explicitly handled. Quality improvement must be manifested as: evidence selection, comparison scope, and conclusion strength work together to make clear what can and cannot be inferred. on the basis of accuracy, semantic calibration among facts is achieved, for example clearly stating that 93.2 corresponds to Large Ens.+TriviaQA Test rather than ordinary Large Single. The most complex key comparison or state chain can still be checked item by item; approximations, limitations, and counterexamples are not concealed by the conclusion. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The pretraining/fine-tuning mechanisms, Base/Large configurations, MLM ratio, NSP labels, and the EM/F1 in the results table must be consistent with arXiv:1810.04805v2; do not fill blank values and do not alter settings. Difficult-case and inspection basis: The most complex key comparison or state chain can still be checked item by item; approximations, limitations, and counterexamples are not concealed by the conclusion. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A non-obvious presentation of the evidence exposes an important conditional relationship that a flawless fact-and-citation account would leave the reader to reconstruct. Show the additional understanding and its support, without inventing a claim, requiring more data, or rewarding the count of sources. Apply the exceptional benefit specifically to: The pretraining/fine-tuning mechanisms, Base/Large configurations, MLM ratio, NSP labels, and the EM/F1 in the results table must be consistent with arXiv:1810.04805v2; do not fill blank values and do not alter settings. In-task verification cases for that benefit: The most complex key comparison or state chain can still be checked item by item; approximations, limitations, and counterexamples are not concealed by the conclusion. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 9, "benchmark_source": "vas" }, { "id": "C3", "dimension": "Content", "criterion": "Information selection and the pretraining–fine-tuning connection.\nWhat is measured: Content is organized around the reading group's needs, highlighting the connection from bidirectional pretraining to downstream fine-tuning and avoiding piling up irrelevant paper content; evidence boundaries and attribution limitations are handled by O2.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check whether the eight pages cover the motivation, architecture, MLM, NSP, SQuAD results, conclusions, and limitations listed in the prompt. Check whether irrelevant paper content crowds out the content required by the prompt. Check whether the connection from pretraining to fine-tuning is explicitly stated.\nScoring boundary: Content selection, density, tone, and necessary qualifications around the scenario purpose; not font size or page visual hierarchy.\nOriginal full refinement inspection basis: The most information-dense places still have appropriate selection, the key points and exceptions are understandable, and there is no redundant expansion or oversimplification. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nTask scenario: An English research briefing for a machine learning reading group: the audience understands neural networks but has not read BERT, and needs to produce a compilable 16:9 eight-page Beamer presentation in an offline workspace using the given Metropolis theme, accurately explaining pretraining, fine-tuning, and the SQuAD evidence.\nScenario reference · purpose (no additional threshold): An academic research briefing to be presented orally at a group meeting and subsequently circulated.\nScenario reference · content selection (no additional threshold): Use BERT arXiv:1810.04805v2 as the sole paper basis, with accurate concepts, clearly bounded evidence, editable figures and tables, and brief traceable sources.\nApplicable boundary for this task (relevant items all apply; no additional scoring items): Use arXiv:1810.04805v2 as the standard; do not mix in later BERT variants or online leaderboard scores.\nApplicable boundary for this task (relevant items all apply; no additional scoring items): Eight landscape 16:9 pages including the title page; do not use overlays or automatic section pages to increase the page count.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: The content selection is unrelated to the reading group's questions, or pretraining is not connected to fine-tuning at all. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Key mechanism details are missing or a large amount of irrelevant content crowds the eight pages, diluting the main line. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: There is one inappropriate selection choice or the statement of the pretraining-to-fine-tuning connection is too weak. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: The content volume is suitable for an introductory briefing, and the relationship between pretraining and fine-tuning is explained. The required content and necessary qualifications are complete, and the depth of explanation, tone, and information volume suit the intended audience for this task. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: there is deliberate selection around motivation, mechanism, and evidence, without piling up irrelevant paper content. Quality improvement must be manifested as: removing restatements that do not aid understanding or placing necessary qualifications where they are needed, making the key points more direct. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: selection serves the narrative, and necessary mechanism details and evidence are retained without duplicating one another. Quality improvement must be manifested as: hierarchy, depth of explanation, and tone suit the audience; streamlined content still retains necessary conditions. selection is mature, retaining necessary mechanism details while naturally integrating the pretraining–fine-tuning connection into the narrative. The most information-dense places still have appropriate selection, the key points and exceptions are understandable, and there is no redundant expansion or oversimplification. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: Content is organized around the reading group's needs, highlighting the connection from bidirectional pretraining to downstream fine-tuning and avoiding piling up irrelevant paper content; evidence boundaries and attribution limitations are handled by O2. Difficult-case and inspection basis: The most information-dense places still have appropriate selection, the key points and exceptions are understandable, and there is no redundant expansion or oversimplification. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. An unusually incisive selection resolves the tension between necessary qualifications and the audience’s main purpose: it conveys more useful understanding with no loss of required substance. Explain why a conventionally fully refined summary would still impose more interpretation; brevity, completeness or promotional impact alone is insufficient. Apply the exceptional benefit specifically to: Content is organized around the reading group's needs, highlighting the connection from bidirectional pretraining to downstream fine-tuning and avoiding piling up irrelevant paper content; evidence boundaries and attribution limitations are handled by O2. In-task verification cases for that benefit: The most information-dense places still have appropriate selection, the key points and exceptions are understandable, and there is no redundant expansion or oversimplification. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 7, "benchmark_source": "vas" }, { "id": "O1", "dimension": "Organization", "criterion": "The main narrative path across eight pages.\nWhat is measured: The eight pages should form a coherent path from motivation, pretraining mechanism, and fine-tuning to evidence and limitations, with order and length allocation serving an audience who has not read the paper.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Record the claim of each page in page order and check whether adjacent pages have a carrying relationship. Confirm that there is a logical bridge between the fine-tuning parameter update page and the SQuAD results page.\nScoring boundary: The logic of the overall argument/decision/operation/concept path; do not score by a fixed structural template.\nOriginal full refinement inspection basis: The most complex branches or comparisons can still be incorporated into the main line, the closure is grounded, and there are no dangling steps or repeated detours. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: There is no traceable main line between pages, and readers cannot proceed from motivation to evidence conclusions. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The narrative is broken or key links are missing, for example there is no connection from pretraining to fine-tuning parameter updates. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: There is one jump or repetition that makes the path less smooth, but it can still be filled in by the reader. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: The eight pages have a clear beginning, middle mechanism and results, and closure, and the path holds. The main premises, process, and conclusions can be connected into a complete and understandable path. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: transition pages and connecting sentences make the main path smoother, and the page load is reasonably distributed. Quality improvement must be manifested as: one key transition clearly connects the premise to the next step, making the main line easier to follow. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: each page has a clear narrative function, and motivation—mechanism—evidence—limitations progress layer by layer and support one another. Quality improvement must be manifested as: the premises, development, and conclusions of each part refer back to one another, and the order and length serve the explanation or decision in this task. the overall narrative rhythm is mature, balancing mechanism explanation, evidence qualification, and entry points for reading-group discussion within eight pages. The most complex branches or comparisons can still be incorporated into the main line, the closure is grounded, and there are no dangling steps or repeated detours. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The eight pages should form a coherent path from motivation, pretraining mechanism, and fine-tuning to evidence and limitations, with order and length allocation serving an audience who has not read the paper. Difficult-case and inspection basis: The most complex branches or comparisons can still be incorporated into the main line, the closure is grounded, and there are no dangling steps or repeated detours. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A non-obvious organization of the existing argument or user path makes separate necessary parts illuminate one another and materially reduces reasoning or navigation steps. Compare the benefit with an equally complete, logically sound conventional order; no additional section or prescribed order is required. Apply the exceptional benefit specifically to: The eight pages should form a coherent path from motivation, pretraining mechanism, and fine-tuning to evidence and limitations, with order and length allocation serving an audience who has not read the paper. In-task verification cases for that benefit: The most complex branches or comparisons can still be incorporated into the main line, the closure is grounded, and there are no dangling steps or repeated detours. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 6, "benchmark_source": "vas" }, { "id": "O2", "dimension": "Organization", "criterion": "Conclusion boundaries and restrained attribution.\nWhat is measured: The conclusion should state the incomparability caused by different dataset versions, Dev/Test splits, Single/Ensemble, and the additional training data of +TriviaQA, should not attribute the specified rows to model size alone, and should not present the paper's then-current results as the current best.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check whether the conclusion page explicitly states the differences in Single/Ensemble, Dev/Test, +TriviaQA, and dataset versions. Check whether the conclusion distinguishes the results reported in the paper from the current best results. Check whether it avoids attributing improvements across versions or settings.\nScoring boundary: The semantic order and relationships of explanations within a page, stage, node, or comparison unit.\nOriginal full refinement inspection basis: The most complex unit can also be understood in the required order, and local explanations are concise, self-contained, and consistent with the overall main line. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: The specified SQuAD rows are treated as a unified ranking of model size or method quality, and the core conclusion is misleading. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Key qualifications about dataset version, split, ensembling, or additional training data differences are missing. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: One statement of evidence scope is too broad, for example calling 93.2 a Large Single result. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: It can explain that these rows cannot serve as a controlled comparison, and the strength of the conclusion matches the evidence. Within each necessary unit, correct relationships can be established among the object, explanation, conditions, and results. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: a brief qualification is given for each type of result difference, and the paper's reported results are not written as the current best. Quality improvement must be manifested as: the arrangement of objects, explanations, and results in local units reduces one necessary look-back. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the qualifications and the conclusion page echo each other, so readers can judge what the evidence can and cannot support. Quality improvement must be manifested as: the figures, text, conditions, and conclusions of each major unit form a locally understandable relationship. the conclusion is specific and restrained, the main confounders are explicitly pointed out, and the strength of the conclusion fully matches the selected evidence. The most complex unit can also be understood in the required order, and local explanations are concise, self-contained, and consistent with the overall main line. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The conclusion should state the incomparability caused by different dataset versions, Dev/Test splits, Single/Ensemble, and the additional training data of +TriviaQA, should not attribute the specified rows to model size alone, and should not present the paper's then-current results as the current best. Difficult-case and inspection basis: The most complex unit can also be understood in the required order, and local explanations are concise, self-contained, and consistent with the overall main line. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. An economical local organization makes a difficult relationship among conditions, evidence and objects understandable as one unit rather than assembled mentally from carefully grouped parts. Identify the semantic reconstruction it eliminates; proximity and tidy grouping alone do not qualify. Apply the exceptional benefit specifically to: The conclusion should state the incomparability caused by different dataset versions, Dev/Test splits, Single/Ensemble, and the additional training data of +TriviaQA, should not attribute the specified rows to model size alone, and should not present the paper's then-current results as the current best. In-task verification cases for that benefit: The most complex unit can also be understood in the required order, and local explanations are concise, self-contained, and consistent with the overall main line. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 5, "benchmark_source": "vas" }, { "id": "O3", "dimension": "Organization", "criterion": "Cross-page identity and visual encoding inheritance.\nWhat is measured: Terminology, color semantics, diagram symbols, and results-table labels remain consistent when reproduced across pages, so readers do not need to relearn the same encoding; whether semantics carry over between pages is handled by O3, and static visual scanning rhythm is handled by V8.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Compare whether the colors of the selection/replacement branches in the MLM diagram are consistent with the subsequent text. Check whether Base/Large and Single/Ensemble are written consistently across pages. Check whether terms and symbols have consistent meanings when reproduced across pages.\nScoring boundary: Semantic inheritance across pages/states, or the closure of branch paths and identity/action within a single page; do not force an added ending.\nOriginal full refinement inspection basis: Complex branches, final states, or resets still maintain identity and causal continuity, and readers do not need later information to patch the current meaning. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nEvaluate meaning carried across the actual slides, not invented software branches or reset states.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: The same symbol or color has opposite meanings on different pages, and cross-page semantic inheritance fails. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Multiple cross-page labels or color meanings drift, interfering with understanding. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: There is one cross-page inconsistency in terminology or visual encoding, but it can still be inferred. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: Cross-page labels and color semantics are consistent. Necessary cross-section, cross-page, or cross-state relationships are traceable, and the semantics of existing endings or resets are clear. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: diagram symbols, table headers, and terminology remain consistent when reproduced across pages. Quality improvement must be manifested as: one cross-page, cross-section, or cross-state transition clearly preserves object identity and existing meaning. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: cross-page identity forms a stable encoding system, for example the MLM branch colors correspond to the subsequent discussion. Quality improvement must be manifested as: the main transitions, branches, and endings form continuous semantics; the positions of repeated objects and necessary exceptions are traceable. cross-page semantic inheritance is mature, and readers can naturally carry information from the mechanism pages into the evidence pages without relying on off-page explanations. Complex branches, final states, or resets still maintain identity and causal continuity, and readers do not need later information to patch the current meaning. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: Terminology, color semantics, diagram symbols, and results-table labels remain consistent when reproduced across pages, so readers do not need to relearn the same encoding; whether semantics carry over between pages is handled by O3, and static visual scanning rhythm is handled by V8. Difficult-case and inspection basis: Complex branches, final states, or resets still maintain identity and causal continuity, and readers do not need later information to patch the current meaning. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A purposeful carry-over lets the audience preserve and develop the right meaning through the task’s actual transitions, materially reducing reconstruction beyond fully refined referential consistency. Identify what understanding accumulates; do not invent branches, endings or reset states absent from the task. Apply the exceptional benefit specifically to: Terminology, color semantics, diagram symbols, and results-table labels remain consistent when reproduced across pages, so readers do not need to relearn the same encoding; whether semantics carry over between pages is handled by O3, and static visual scanning rhythm is handled by V8. In-task verification cases for that benefit: Complex branches, final states, or resets still maintain identity and causal continuity, and readers do not need later information to patch the current meaning. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 4, "benchmark_source": "vas" }, { "id": "V1", "dimension": "Visuals", "criterion": "Intra-page visual hierarchy and focus.\nWhat is measured: Each page should establish a clear title—claim—evidence hierarchy through font size, position, and whitespace, so that readers first see the page's core information.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: At full-page thumbnail scale, check whether each page's first visual landing point matches that page's claim. Check whether key qualifiers have a distinguishable hierarchy from ordinary body text.\nScoring boundary: The salience relationships among titles, body, evidence, and supporting information; evaluate the entry point, core focus, and strength of emphasis.\nOriginal full refinement inspection basis: Even the densest area or state shows the main information first and the basis and qualifications second, with no substantive competition for attention between the whole and the parts. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: The page has no clear focus, and the title, body text, and figures compete for attention, making it impossible to quickly locate the claim. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Most pages have confused focus or core figures are drowned out by secondary text. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: A few pages have a weak hierarchy, requiring extra effort to find the key point. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: Each page has a recognizable title and content hierarchy. The primary/secondary relationships among main objects, supporting information, and auxiliary notes are distinguishable, with no substantive focus conflict. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: titles, key sentences, and figures form a clear reading order. Quality improvement must be manifested as: a set of positional, scale, or contrast choices makes the main focus clearer and prevents auxiliary information from competing for attention. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: hierarchy serves the narrative, and key qualifiers and figure conclusions can be grasped at first glance. Quality improvement must be manifested as: the strength relationships among the main focus, evidence, and necessary qualifications are coordinated and reasonably adjusted with content complexity. hierarchy is restrained and precise, allowing complex information to be naturally layered while maintaining Metropolis's minimalist character. Even the densest area or state shows the main information first and the basis and qualifications second, with no substantive competition for attention between the whole and the parts. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: Each page should establish a clear title—claim—evidence hierarchy through font size, position, and whitespace, so that readers first see the page's core information. Difficult-case and inspection basis: Even the densest area or state shows the main information first and the basis and qualifications second, with no substantive competition for attention between the whole and the parts. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A non-obvious hierarchy brings the decisive relationship and its necessary qualification into attention together without competing emphasis. Demonstrate a first-look interpretation benefit beyond an impeccably clear primary/secondary hierarchy, not just stronger contrast or larger type. Apply the exceptional benefit specifically to: Each page should establish a clear title—claim—evidence hierarchy through font size, position, and whitespace, so that readers first see the page's core information. In-task verification cases for that benefit: Even the densest area or state shows the main information first and the basis and qualifications second, with no substantive competition for attention between the whole and the parts. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 5, "benchmark_source": "vas" }, { "id": "V2", "dimension": "Visuals", "criterion": "Block proportions, alignment, and whitespace.\nWhat is measured: The proportions of text and image blocks, margins, alignment, and whitespace across the eight pages should be stable and suited to the 16:9 reading scale, without obvious imbalance due to content amount.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check page by page whether content crosses the safe margins or sits too close to the footer. Compare whether mechanism pages and results pages maintain consistent margins despite differences in content density.\nScoring boundary: Area proportions among information blocks, grid, edges, center of gravity, density, and whitespace; evaluate the overall composition and spatial organization.\nOriginal full refinement inspection basis: Even the area with the greatest change in density or state still has a stable spatial order, and overall balance and local spacing both hold. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: Page blocks are severely imbalanced or cross boundaries, and core content cannot be read normally. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Whitespace and block proportions are clearly out of control on multiple pages, and the layout appears crowded or empty. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: A few pages have visible alignment or whitespace issues, but they do not block reading. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: The overall layout is orderly, and text and image blocks are balanced. The required block proportions, edges, and whitespace are usable, and the layout has no substantive crowding, looseness, or imbalance. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: margins and block rhythm are stable across pages, and figures and text are well aligned. Quality improvement must be manifested as: block edges, scale, or whitespace have a clear purpose that improves the spatial relationships of one type of content. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: proportions and whitespace are deliberately designed, and a unified rhythm remains between complex and lightweight pages. Quality improvement must be manifested as: area proportions, alignment, center of gravity, and whitespace jointly adapt to the actual content rather than mechanically applying the same layout. layout control is mature, and whitespace both carries Metropolis's sense of breathing room and precisely guides the reading path. Even the area with the greatest change in density or state still has a stable spatial order, and overall balance and local spacing both hold. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The proportions of text and image blocks, margins, alignment, and whitespace across the eight pages should be stable and suited to the 16:9 reading scale, without obvious imbalance due to content amount. Difficult-case and inspection basis: Even the area with the greatest change in density or state still has a stable spatial order, and overall balance and local spacing both hold. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A content-specific spatial solution reconciles competing demands of overview, detail and available space with exceptional clarity and balance. Identify the real space conflict resolved beyond a flawlessly fitted conventional composition; extra whitespace, density or a fashionable layout is not itself a benefit. Apply the exceptional benefit specifically to: The proportions of text and image blocks, margins, alignment, and whitespace across the eight pages should be stable and suited to the 16:9 reading scale, without obvious imbalance due to content amount. In-task verification cases for that benefit: Even the area with the greatest change in density or state still has a stable spatial order, and overall balance and local spacing both hold. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 8, "benchmark_source": "vas" }, { "id": "V3", "dimension": "Visuals", "criterion": "Fonts, type size, and readability.\nWhat is measured: Body text size, line spacing, line breaks, and figure/table labels should be clearly readable at 16:9 projection and PDF reading scales, and formulas and tables should not become hard to distinguish due to compression.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have a neural network background but have not systematically read BERT.\nObservation basis: Check table values and figure captions at both 100% and projection-thumbnail scales. Check whether long formulas and MLM branch labels overflow or are forced to shrink too small.\nScoring boundary: The complete visibility of text in the actually delivered image, type size, contrast, line spacing, and line breaks; evaluate text presentation and reading at the target scale.\nOriginal full refinement inspection basis: The longest necessary labels, small qualifications, or complex symbols remain complete and easy to read, and typographic quality is consistently stable across areas. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: Core text or table values are unreadable at a normal reading scale. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Font sizes are too small, line spacing is too tight, or line breaks are severe in multiple places, affecting understanding. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "-1": "Below basic requirements: This item has a clear, localized substantive defect. Specific manifestation: Readability is weak for a few figure/table labels or long passages of text. Judge by the actual impact described above; harmless minor differences do not trigger this level.", "0": "Meets basic requirements: Body text and figure/table labels are readable at normal scale, with reasonable line breaks. All necessary text is fully readable at the scale specified by the prompt, and type size, contrast, line spacing, and line breaks have no substantive defects. The content and conditions required by this item are fully satisfied; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: the type-size hierarchy is clear, and formulas, tables, and figure captions all have appropriate font sizes. Quality improvement must be manifested as: the type size, spacing, or line breaks of one type of necessary text reduce reading burden, not merely being readable. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: font choice and line spacing work together, and complex pages still maintain a comfortable reading rhythm. Quality improvement must be manifested as: type size, line spacing, contrast, and semantic line breaks are coordinated, and dense content remains comfortably readable at the scale specified by the prompt. typography is mature, and formulas, figure/table labels, and body text remain stable and clearly legible both in projection and on screen. The longest necessary labels, small qualifiers, or complex symbols remain complete and easy to read, and typographic quality across regions remains consistently stable. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: Body text size, line spacing, line breaks, and figure/table labels should be clearly readable at 16:9 projection and PDF reading scales, and formulas and tables should not become hard to distinguish due to compression. Difficult-case and inspection basis: The longest necessary labels, small qualifications, or complex symbols remain complete and easy to read, and typographic quality is consistently stable across areas. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. Typographic structure makes a difficult meaning or relationship apparent as it is read, reducing parsing beyond fully refined legibility, spacing and hierarchy. Identify the specific semantic contribution at the required viewing scale; a distinctive font or decorative lettering is not sufficient. Apply the exceptional benefit specifically to: Body text size, line spacing, line breaks, and figure/table labels should be clearly readable at 16:9 projection and PDF reading scales, and formulas and tables should not become hard to distinguish due to compression. In-task verification cases for that benefit: The longest necessary labels, small qualifications, or complex symbols remain complete and easy to read, and typographic quality is consistently stable across areas. In this task, focus on the MLM conditional probability illustration and the parts of the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data that are relevant to this item; do not add deliverables, scenarios, or experiments outside the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 7, "benchmark_source": "vas" }, { "id": "V4", "dimension": "Visuals", "criterion": "Accurate decoding of diagrams and tables.\nWhat is measured: The MLM diagram, NSP flow, and SQuAD table should let readers accurately read the selection rate, conditional replacement probability, prediction target, labels, and each result column without ambiguity.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Check whether the MLM diagram marks both the 15% selection of all positions and the 80/10/10 conditional probabilities after selection. Check whether the table indicates the dataset version, Dev/Test, model setting, EM, and F1, and preserves the blank Test cells as blank.\nScoring boundary: The recognizability and semantic correspondence of graphical marks, lines, arrows, symbols, and dynamic states; evaluate accurate decoding and reliable comparison.\nOriginal full refinement inspection basis: The most complex connections, states, or scale relationships can still be reliably decoded, and the legend's correspondence with the actual content is neither ambiguous nor misleading. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: The diagram or table encoding misleads on a core fact, for example, drawing 80/10/10 as a proportion of all tokens. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Key branches lack labels or table column meanings are unclear, so readers will misread. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: One arrow, legend, or table header requires going back to the body text to understand. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: Diagrams and tables can be decoded, and labels are complete. Necessary graphics, symbols, and connections can be accurately decoded, and their meanings match the actual relationships. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: branches, arrows, legends, and table headers are consistent with the body text terminology, and the reading method is clear. Quality improvement must be reflected as: A set of marks, connections, or comparisons makes the object relationships more directly decodable and reduces guesswork about the diagram. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the diagram directly distinguishes the global selection rate from the conditional replacement rate through hierarchy and labeling, and the table clearly expresses version/split/setting through grouping. Quality improvement must be reflected as: Graphical encoding, labels, and actual relationships support one another; complex comparisons do not require extra explanation to correct the diagram's meaning. the decoding design is mature, and readers can accurately restate the MLM mechanism and the boundaries of the results table without verbal supplementation. The most complex connections, states, or scale relationships can still be reliably decoded, and the relationship between the legend and the actual content is unambiguous and not misleading. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The MLM diagram, NSP flow, and SQuAD table should let readers accurately read the selection rate, conditional replacement probability, prediction target, labels, and each result column without ambiguity. Difficult-case and inspection basis: The most complex connections, states, or scale relationships can still be reliably decoded, and the legend's correspondence with the actual content is neither ambiguous nor misleading. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A visual encoding makes a difficult relationship directly perceptible instead of requiring mental reconstruction from fully correct labels and marks. Show what becomes easier to understand and verify fidelity to the source or required state; novelty and additional diagrams are not requirements. Apply the exceptional benefit specifically to: The MLM diagram, NSP flow, and SQuAD table should let readers accurately read the selection rate, conditional replacement probability, prediction target, labels, and each result column without ambiguity. In-task verification cases for that benefit: The most complex connections, states, or scale relationships can still be reliably decoded, and the legend's correspondence with the actual content is neither ambiguous nor misleading. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 5, "benchmark_source": "vas" }, { "id": "V5", "dimension": "Visuals", "criterion": "Integration of non-color graphics with the typographic language.\nWhat is measured: The line weights, label fonts, and typographic tone of diagrams, formulas, and tables should be consistent with Metropolis's concise character, presenting a restrained academic style rather than a decorative collage; the color area and accent color semantics are evaluated independently by V9.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Check whether the MLM diagram is drawn natively in TeX and whether its line weights and fonts are consistent with the theme. Check whether the tables and formulas follow the theme's typographic tone rather than being screenshot collages. Check whether decorations that conflict with Metropolis's concise character are introduced.\nScoring boundary: Audience, purpose, medium character, and the integration of graphical abstraction, typography, images, and component surface language; pure color aesthetics are deferred to V9.\nOriginal full refinement inspection basis: Even the hardest-to-integrate content harmonizes with the overall visual language; when information is complete, there is still a clear, restrained scenario expression. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nScenario reference · visual language (does not add thresholds): Retain the supplied Metropolis title treatment, color scheme, and spacing conventions, and stay restrained and academic overall.\nScenario reference · design constraints (does not add thresholds): Retain the theme appearance and the hard eight-page constraint, while accurately distinguishing the 15% global selection rate from the conditional replacement rate and the various SQuAD settings within the limited layout, and do not treat screenshots or non-editable images as core content.\nScenario reference · design goals (does not add thresholds): The eight pages form a coherent narrative; the MLM conditional probabilities and prediction targets are clear at a glance; the results table accurately indicates version, split, setting, and sources of blanks; boundaries and build are verifiable.\nScoring boundary: V9 independently evaluates the aesthetic harmony and refinement of the color scheme; reading impairment caused by text contrast belongs to V3, mismatch between colors and object meanings belongs to V4, and cross-page color-code identity drift belongs to V7. Attribute across items only when different functional impairments are demonstrated. V5 cannot merely repeat whether V9's color scheme is good or bad, and V9 cannot directly award a professional tier merely for 'fitting the scenario'.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Do not deduct points for not replicating the paper's original layout or for not providing other SQuAD rows not specified in the paper.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Installing Fira Sans is not required; font substitution is allowed but must be recorded in BUILD.md.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Do not treat the Metropolis theme files or paper redistribution rights issues as briefing content requirements.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): When copying or modifying theme files, retain the original copyright and CC BY-SA 4.0 license notice.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: The visual language severely conflicts with the theme, and core diagrams look like foreign screenshots or decorations that overpower the main content. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The graphics style on multiple pages is disconnected from the theme typography and looks like a collage. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: A few graphics or components have inconsistent tone, but the overall result is still acceptable. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: The overall result maintains a restrained academic style, and the diagrams harmonize with the theme. The language of graphics, typography, images, and components suits the audience and purpose, with no substantive style conflict. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: diagram line weights and label fonts are consistent with the typography of tables/formulas, and graphical abstraction does not introduce decorations that conflict with the theme. Quality improvement must be reflected as: Specific choices in graphical abstraction, typography, or image treatment improve audience understanding and the overall impression. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: custom diagrams, tables, and formulas form a unified non-color academic graphical language and retain the theme's character. Quality improvement must be reflected as: The language of typography, graphics, images, or components harmoniously serves this task's purpose, and attractive color schemes are not used as a substitute for analysis. non-color visual integration is mature, graphical abstraction, typographic tone, and component language harmonize, and it does not rely on extra decoration. Even the hardest-to-integrate content harmonizes with the overall visual language; when information is complete, there is still a clear, restrained scenario expression. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The line weights, label fonts, and typographic tone of diagrams, formulas, and tables should be consistent with Metropolis's concise character, presenting a restrained academic style rather than a decorative collage; the color area and accent color semantics are evaluated independently by V9. Difficult-case and inspection basis: Even the hardest-to-integrate content harmonizes with the overall visual language; when information is complete, there is still a clear, restrained scenario expression. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A distinctive but appropriate visual language makes the subject or purpose intrinsic to the graphic form, beyond flawless stylistic integration. Identify a non-color design contribution and its communication benefit; personal taste, modernity and impressive decoration are not evidence. Apply the exceptional benefit specifically to: The line weights, label fonts, and typographic tone of diagrams, formulas, and tables should be consistent with Metropolis's concise character, presenting a restrained academic style rather than a decorative collage; the color area and accent color semantics are evaluated independently by V9. In-task verification cases for that benefit: Even the hardest-to-integrate content harmonizes with the overall visual language; when information is complete, there is still a clear, restrained scenario expression. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 7, "benchmark_source": "vas" }, { "id": "V6", "dimension": "Visuals", "criterion": "Local geometry and craftsmanship.\nWhat is measured: Arrow endpoints, frame lines, table rules, cropping, and optical alignment should be clean and accurate, with no visible jaggies, misalignment, or overlap.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Zoom in to check whether the MLM branch arrows and frame line endpoints are aligned. Check whether table vertical lines, horizontal lines, and text baselines are aligned.\nScoring boundary: Graphical outlines, line weights, endpoints, image proportions and cropping, geometry within components, and optical alignment; evaluate local construction precision.\nOriginal full refinement inspection basis: The finest connections, complex intersections, or irregular contours are still handled adequately, and local craftsmanship does not undermine the overall effect. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: Core diagrams or tables have severe misalignment, overlap, or cropping that impairs decoding. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Multiple local geometric defects are obvious, and the layout looks rough. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: One arrow, frame line, or alignment needs correction but does not affect understanding. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: Local geometry is clean, with no obvious misalignment. The endpoints, outlines, proportions, cropping, and local alignment of necessary graphics and components are correct and usable. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: arrow endpoints, frame lines, and table alignment have been polished. Quality improvement must be reflected as: The treatment of a class of endpoints, line weights, cropping, or optical alignment is more precise than ordinary tidiness. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: local craftsmanship is stable, and complex diagrams remain accurate at the detail scale. Quality improvement must be reflected as: The geometry, boundaries, and image proportions of different components harmonize, and the result is stable both up close and in overall viewing. local craftsmanship remains fully harmonious even in the most complex diagrams and tables, and details withstand magnification. The finest connections, complex intersections, or irregular contours are still handled adequately, and local craftsmanship does not undermine the overall effect. Task-specific object: Arrow endpoints, frame lines, table rules, cropping, and optical alignment should be clean and accurate, with no visible jaggies, misalignment, or overlap. Difficult-case and inspection basis: The finest connections, complex intersections, or irregular contours are still handled adequately, and local craftsmanship does not undermine the overall effect. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. Fine construction resolves a real perceptual ambiguity at the task’s most demanding local junction or component, beyond spotless alignment and uniform polish. Identify the optical or geometric benefit visible at the intended scale; more detail or tiny changes with no perceptible benefit do not qualify. Apply the exceptional benefit specifically to: Arrow endpoints, frame lines, table rules, cropping, and optical alignment should be clean and accurate, with no visible jaggies, misalignment, or overlap. In-task verification cases for that benefit: The finest connections, complex intersections, or irregular contours are still handled adequately, and local craftsmanship does not undermine the overall effect. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 6, "benchmark_source": "vas" }, { "id": "V7", "dimension": "Visuals", "criterion": "Cross-page visible identity consistency.\nWhat is measured: The eight pages maintain a recognizable single visual identity in title position, footer, color semantics, and diagram style, while not sacrificing content hierarchy for consistency.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Spot-check the title position and footer style on the title page, mechanism page, and results page. Check whether the color semantics maintain the same meaning across pages.\nScoring boundary: The stable correspondence of the same object, identity, mark, style, and content across pages, states, and carriers; evaluate visual inheritance. Deduct only for independently impaired functions; harmless anti-aliasing, font rasterization, or slight color differences across carriers are not considered identity drift.\nOriginal full refinement inspection basis: Complex states or the densest content remain consistent across all required carriers; identity, layout, and qualifications remain stably traceable. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Retain the delivered Metropolis title treatment, color scheme, and spacing conventions; font substitutions must be recorded in BUILD.md.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nCheck recurring objects in the actual slides and required exports; do not demand additional export formats.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: The pages look like they come from different templates, and the cross-page identity is unrecognizable. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The title treatment or footer style clearly drifts across multiple pages. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: The visual identity of a few pages is weak or inconsistent. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: The eight pages maintain a recognizable unified theme identity. The same objects, content, and styles correspond across necessary areas, states, and delivery carriers, without substantive drift. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: titles, footers, and diagram styles are consistent across pages. Quality improvement must be reflected as: Repeated objects or corresponding carriers follow clear visual rules, reducing the burden of re-recognition. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: under a unified identity, local variation still distinguishes mechanism pages from evidence pages. Quality improvement must be reflected as: State changes, repeated components, and necessary exports correspond to one another, and content changes are distinguished without identity drift. cross-page identity is mature and stable, and theme constraints and page-function variation are well balanced. Complex states or the densest content remain consistent across all required carriers; identity, layout, and qualifications remain stably traceable. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The eight pages maintain a recognizable single visual identity in title position, footer, color semantics, and diagram style, while not sacrificing content hierarchy for consistency. Difficult-case and inspection basis: Complex states or the densest content remain consistent across all required carriers; identity, layout, and qualifications remain stably traceable. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A deliberately designed recurring visible identity enables recognition across the task’s genuinely different views or presentations with materially less re-identification effort than fully refined consistency alone. Name the recognition benefit and verify it across actual delivered contexts, without requiring nonexistent extra versions. Apply the exceptional benefit specifically to: The eight pages maintain a recognizable single visual identity in title position, footer, color semantics, and diagram style, while not sacrificing content hierarchy for consistency. In-task verification cases for that benefit: Complex states or the densest content remain consistent across all required carriers; identity, layout, and qualifications remain stably traceable. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 3, "benchmark_source": "vas" }, { "id": "V8", "dimension": "Visuals", "criterion": "Static reading rhythm.\nWhat is measured: As a static briefing with no animation, the pages should form a natural visual scan, attention shifts, and reading rhythm through sequence, whitespace, and diagram progression; adding animation or transitions is not required; whether semantics connect between pages belongs to O3, and static screenshots are insufficient to prove dynamic motion experience.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Record first-reading pauses and look-back counts in page order. Check whether the MLM diagram branches guide reading in the intended order. Check whether static pages complete the visual scan without animation.\nScoring boundary: Eye movement between information units and content reveal, dwell, motion, and transitions; evaluate scanning and temporal rhythm.\nOriginal full refinement inspection basis: The densest content or most complex handoffs can still be followed calmly, and the overall and detailed rhythm of a complete reading or playback cycle are consistent. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Adding animation, transitions, or dynamic video is not required.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nFor static slides, assess scanning and backtracking at presentation scale, not motion effects.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: Page order and layout make the reading path jump repeatedly, making one coherent reading difficult. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Information density changes abruptly across multiple pages or the diagram progression order is chaotic, interrupting the reading rhythm. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: A few pages require jumping back and forth to understand. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: Reading in page order is smooth and requires no extra animation support. The necessary scanning, dwell, and handoff processes in this task can be followed, and the information can be read to completion under its actual presentation conditions. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: the within-page reading order and cross-page visual progression form a stable rhythm. Quality improvement must be reflected as: The arrangement of a class of scan paths, dwells, or handoffs makes necessary information easier to finish and follow. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: whitespace and diagram progression intentionally guide pauses and comparisons, and the static reading experience is complete. Quality improvement must be reflected as: Eye movement or timing is coordinated with content difficulty, with no unnecessary back-jumps or waiting between entering, understanding, and leaving. the static rhythm is mature, and readers can complete a continuous visual understanding from mechanism to evidence along the designed path. The densest content or most complex handoffs can still be followed calmly, and the overall and detailed rhythm of a complete reading or playback cycle are consistent. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: As a static briefing with no animation, the pages should form a natural visual scan, attention shifts, and reading rhythm through sequence, whitespace, and diagram progression; adding animation or transitions is not required; whether semantics connect between pages belongs to O3, and static screenshots are insufficient to prove dynamic motion experience. Difficult-case and inspection basis: The densest content or most complex handoffs can still be followed calmly, and the overall and detailed rhythm of a complete reading or playback cycle are consistent. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. A non-obvious reading or temporal organization lets the audience orient, inspect and continue with unusually little backtracking or missed understanding. Identify the timing or scanning tension solved beyond consistently smooth ordinary presentation; add no motion or interaction to a static task. Apply the exceptional benefit specifically to: As a static briefing with no animation, the pages should form a natural visual scan, attention shifts, and reading rhythm through sequence, whitespace, and diagram progression; adding animation or transitions is not required; whether semantics connect between pages belongs to O3, and static screenshots are insufficient to prove dynamic motion experience. In-task verification cases for that benefit: The densest content or most complex handoffs can still be followed calmly, and the overall and detailed rhythm of a complete reading or playback cycle are consistent. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 3, "benchmark_source": "vas" }, { "id": "V9", "dimension": "Visuals", "criterion": "Color area and accent color semantics.\nWhat is measured: The area allocation of the Metropolis color scheme in titles, body text, diagrams, and tables, the use of accent colors, and image background colors should be harmonious; accent colors should have clear semantics, and the color scheme should not disrupt the information hierarchy; color harmony alone does not support a high score.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Check whether accent colors are used only for key branches, qualifiers, or table headers. Check whether diagram and table background colors harmonize with the theme background. Check whether color semantics are consistent with the diagram encoding.\nScoring boundary: Hue, lightness/darkness, saturation × area, accent colors, and the aesthetic relationship between images/backgrounds; demonstrate quality goals separately from V3/V4/V5/V7.\nObservation elements: Hue compatibility and warm-cool relationships. Lightness hierarchy and color focal point. Saturation × area ratio. Quantity, weight, and restraint of accent colors. Integration of images, graphics, panels, and background colors.\nOriginal full refinement inspection basis: Among the hardest-to-match images, dense areas, or changing states, the relationships among lightness, hue, and area remain unified, clear, and restrained. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nScoring boundary: V9 independently evaluates the aesthetic harmony and refinement of the color scheme; reading impairment caused by text contrast belongs to V3, mismatch between colors and object meanings belongs to V4, and cross-page color-code identity drift belongs to V7. Attribute across items only when different functional impairments are demonstrated. V5 cannot merely repeat whether V9's color scheme is good or bad, and V9 cannot directly award a professional tier merely for 'fitting the scenario'.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.\nJudge negative visual grades by substantive harm and extent within this criterion, not a prerequisite of functional unusability. Harmless style differences are not defects; retain the original concrete defect examples.", "answers": { "-3": "Severe failure: This criterion’s visual goal is severely impaired. Specific manifestation: Color areas are unbalanced or accent colors are overused, making it difficult to focus on core information. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The color scheme across multiple pages looks harsh, muddy, or conflicts with the diagram backgrounds. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: On a few pages, accent colors are overused or area allocation is not harmonious enough. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: The color scheme is harmonious overall, and accent colors are used with restraint. The relationship among the subject, background, images, and accent colors is harmonious, with no substantive color conflict; common combinations do not by themselves count as improvement. The content and conditions required by this item are fully met; there are no substantive defects, and performance is ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: accent colors are concentrated on key qualifiers and diagram focal points, and the area allocation is reasonable. Quality improvement must be reflected as: Specify how particular hue, lightness, or saturation and area relationships improve harmony, rather than merely listing color names. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the color scheme forms a unified aesthetic with the diagrams and table backgrounds, and accent colors have clear semantics. Quality improvement must be reflected as: The relationship among the subject, background, images, and accent colors has been harmonized, and the color focal point and area control serve the actual content. hue, lightness, saturation, and area are precisely integrated in an academic context, and color semantics are consistent with the diagram encoding. Among the hardest-to-match images, dense areas, or changing states, the relationships among lightness, hue, and area remain unified, clear, and restrained. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: The area allocation of the Metropolis color scheme in titles, body text, diagrams, and tables, the use of accent colors, and image background colors should be harmonious; accent colors should have clear semantics, and the color scheme should not disrupt the information hierarchy; color harmony alone does not support a high score. Difficult-case and inspection basis: Among the hardest-to-match images, dense areas, or changing states, the relationships among lightness, hue, and area remain unified, clear, and restrained. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. Subtle hue, lightness, saturation and area relationships reconcile the hardest color-balance conflict with a distinctive, highly resolved visual character. Explain the perceptual gain beyond a flawlessly harmonious palette; semantic color correctness, fashion and personal liking do not establish exceptional color design. Apply the exceptional benefit specifically to: The area allocation of the Metropolis color scheme in titles, body text, diagrams, and tables, the use of accent colors, and image background colors should be harmonious; accent colors should have clear semantics, and the color scheme should not disrupt the information hierarchy; color harmony alone does not support a high score. In-task verification cases for that benefit: Among the hardest-to-match images, dense areas, or changing states, the relationships among lightness, hue, and area remain unified, clear, and restrained. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 6, "benchmark_source": "vas" }, { "id": "D1", "dimension": "Delivery", "criterion": "The deliverable can be opened and used directly.\nWhat is measured: artifact/main.tex, artifact/briefing.pdf, and the locally required supporting files should be complete; the PDF should be exactly eight pages, landscape 16:9 (including the title page), and can be opened directly to locate and read; delivering a complete TeX distribution is not required.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Open artifact/briefing.pdf and check the page count, orientation, and title page. Check whether the local files referenced by artifact/main.tex exist in the delivery directory. Confirm that no points are deducted for not including a complete TeX distribution.\nScoring boundary: The opening, locating, and direct use of required files; V7 evaluates the consistency of the actually visible content and does not substitute file existence for quality.\nOriginal full refinement inspection basis: The entire viewing flow required to go from location to completion of this task is smooth, and even the most easily confused files, entry points, or end states have no handoff breaks. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): Retain read-only inputs such as workspace/sources/paper.pdf and workspace/sources/paper.txt, and do not modify the originals.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: The PDF is missing, or the page count or orientation is severely incorrect, so it cannot be used as an eight-page 16:9 briefing. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The PDF exists, but hard delivery conditions such as page count, orientation, or title page are not satisfied. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: The delivery is basically usable, but one supporting file or page specification issue needs manual handling. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: artifact/main.tex, artifact/briefing.pdf, and the locally required supporting files should be complete; the PDF should be exactly eight pages, landscape 16:9 (including the title page), and can be opened directly to locate and read; delivering a complete TeX distribution is not required. All necessary requirements of this item in the prompt are met, with no substantive defects; do not award points for extra deliverables or increased explanatory length. The delivery conditions required by this item are fully met, with no substantive defects; structure and explanation may be ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: file organization is clear, and the PDF page specifications and opening experience are stable. Quality improvement must be reflected as: The entry point, naming, or order saves the user one search or trial-and-error step, and the improvement can be verified through the actual opening process. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the delivery directory is self-consistent, and all local resources referenced by main.tex can be located. Quality improvement must be reflected as: The entry point, file roles, and necessary viewing order work together, so no assembly or version guessing is needed when used in the manner specified by the task. the delivery is mature and can be used immediately in the reading-group scenario, and the directory structure facilitates reproduction of the deliverable. The entire viewing flow required to go from location to completion of this task is smooth, and even the most easily confused files, entry points, or end states have no handoff breaks. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: artifact/main.tex, artifact/briefing.pdf, and the locally required supporting files should be complete; the PDF should be exactly eight pages, landscape 16:9 (including the title page), and can be opened directly to locate and read; delivering a complete TeX distribution is not required. Difficult-case and inspection basis: The entire viewing flow required to go from location to completion of this task is smooth, and even the most easily confused files, entry points, or end states have no handoff breaks. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. Within the required deliverables and entry points, an economical handoff removes a real selection, opening or first-use ambiguity that persists in a conventionally flawless delivery. Identify the materially reduced effort; extra files, packaging or software are not required. If no additional benefit is observable, +2 is the ceiling. Apply the exceptional benefit specifically to: artifact/main.tex, artifact/briefing.pdf, and the locally required supporting files should be complete; the PDF should be exactly eight pages, landscape 16:9 (including the title page), and can be opened directly to locate and read; delivering a complete TeX distribution is not required. In-task verification cases for that benefit: The entire viewing flow required to go from location to completion of this task is smooth, and even the most easily confused files, entry points, or end states have no handoff breaks. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 4, "benchmark_source": "vas" }, { "id": "D2", "dimension": "Delivery", "criterion": "Editable structure and offline resources.\nWhat is measured: main.tex, the MLM diagram, and the results table should remain TeX-editable, and the required theme styles and resources should be embedded locally, without relying on the network or paper screenshots.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Check whether the MLM diagram is drawn by TeX code such as TikZ. Check whether the table is generated by a TeX table environment rather than an image. Check whether the theme styles and resources exist locally and do not rely on the network.\nScoring boundary: The editable structure, resources, and material organization of the objects required by the prompt; engineering scripts or extra tests are not required.\nOriginal full refinement inspection basis: Check the relationship between the required editable objects and resources involved in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; original raster assets permitted by the prompt are allowed, and vectorization is not mandatory.\nApplicable boundary for this task (applies to all related entries; does not add scoring items): The core MLM diagram and results table must remain TeX-editable and must not be replaced by paper screenshots.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: The core diagram or results table is replaced by a screenshot and cannot be edited in TeX. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: Multiple core contents are not editable, or key resources are missing, making offline building impossible. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: One core content item or resource requires extra manual completion. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: main.tex, the MLM diagram, and the results table remain TeX-editable, and the required theme styles and resources are embedded locally, without relying on the network or paper screenshots. The delivery conditions required by this item are fully met, with no substantive defects; structure and explanation may be ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: the grouping, naming, or source definitions of necessary objects make locating one class of labels, graphics, or values more direct, and specific improvements can be pointed out from the actual structure. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the labels, graphics, data, and necessary resource relationships of the complex objects in this item are organized coherently, and locating and maintaining them does not require rebuilding unrelated content. even the most complex necessary objects have clear source definitions and correspondences, and structure and resource organization remain stably maintained throughout the delivery; out-of-scope parameterization or actually modifying the submitted files is not required. Task-specific object: main.tex, the MLM diagram, and the results table should remain TeX-editable, and the required theme styles and resources should be embedded locally, without relying on the network or paper screenshots. Difficult-case and inspection basis: Check the relationship between the required editable objects and resources involved in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; original raster assets permitted by the prompt are allowed, and vectorization is not mandatory. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. The native editable organization reveals a meaningful relationship among task objects so directly that a targeted maintenance step requires materially less structural reconstruction than in a flawlessly editable conventional artifact. Show this through native evidence, without requiring an extra refactoring, unrequested source code or speculative editing claims. Apply the exceptional benefit specifically to: main.tex, the MLM diagram, and the results table should remain TeX-editable, and the required theme styles and resources should be embedded locally, without relying on the network or paper screenshots. In-task verification cases for that benefit: Check the relationship between the required editable objects and resources involved in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; original raster assets permitted by the prompt are allowed, and vectorization is not mandatory. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 3, "benchmark_source": "vas" }, { "id": "D3", "dimension": "Delivery", "criterion": "Build instructions and sources are verifiable.\nWhat is measured: artifact/BUILD.md should give usable XeLaTeX commands and font requirements, and the briefing should give the paper version and brief source citations to the relevant sections, figures, or tables.\nObservation scope: A landscape 16:9 PDF generated by XeLaTeX/Beamer, with TeX source files and build instructions; members of an ML reading group who have neural network fundamentals but have not systematically read BERT.\nObservation basis: Check whether the XeLaTeX command recorded in BUILD.md matches the actual main file. Check whether the source citations in the briefing point to the paper version and specific sections, figures, or tables. Check whether font substitutions are recorded in BUILD.md.\nScoring boundary: The verifiability of required sources, calculations/changes/captures, or usage instructions; choose according to the actual requirements of this task, and do not require all document types.\nOriginal full refinement inspection basis: Even the most easily confused sources, conditions, or changes can be checked along the instructions, and the complete verification path is concise with no gaps or overclaims. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task.\nRevised grade boundary: the original high-grade descriptions are +2 quality obligations, not sufficient evidence for +3. +2 covers the entire applicable scope; +3 additionally requires the exceptional benefit specified in answers[\"3\"]. Levels inherit quality obligations, not the shortcomings explaining lower grades.", "answers": { "-3": "Severe failure: The core understanding or usage goal of this item cannot be achieved. Specific manifestation: Build instructions or source citations are missing, making it impossible to verify how it is built and what it is based on. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-2": "Clearly unacceptable: The main area or key process of this item is persistently impaired. Specific manifestation: The build commands are unusable or font requirements are missing, and source citations are missing in multiple places. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "-1": "Below basic requirements: This item has a clear local substantive defect. Specific manifestation: One build instruction or source citation is incomplete and requires manual checking. Judge by the actual impact above; harmless minor differences do not trigger this tier.", "0": "Basically acceptable: artifact/BUILD.md should give usable XeLaTeX commands and font requirements, and the briefing should give the paper version and brief source citations to the relevant sections, figures, or tables. All necessary requirements of this item in the prompt are met, with no substantive defects; do not award points for extra deliverables or increased explanatory length. The delivery conditions required by this item are fully met, with no substantive defects; structure and explanation may be ordinary.", "1": "Accomplished but not yet fully refined: meets tier 0 and shows a substantive observed benefit beyond ordinary correctness, completeness or readability through the following: build commands and font requirements are clear, and source citations point to specific sections, figures, or tables. Quality improvement must be reflected as: At least one class of key evidence has a precise location and corresponds to the actual object, reducing the need to search through vague sources. In reason explain the concrete benefit and the observable limitation keeping it below +2. Do not invent a defect to force this grade; missing required observation must be acquired first.", "2": "Fully refined within scope: meets tier 0 and achieves complete quality in every applicable part, not only an attractive local example. Retain and jointly verify these quality obligations: the build instructions are reproducible, and source citations correspond one-to-one with figure and table content. Quality improvement must be reflected as: Claims, materials, and the explanations required by the prompt correspond to one another, so that one can both trace the evidence and know where that evidence is used. the build and source explanations are precise and verifiable, and readers can reproduce and trace each key fact according to the documentation. Even the most easily confused sources, conditions, or changes can be checked along the instructions, and the complete verification path is concise with no gaps or overclaims. Verify the complex parts relevant to this item: the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data. Task-specific object: artifact/BUILD.md should give usable XeLaTeX commands and font requirements, and the briefing should give the paper version and brief source citations to the relevant sections, figures, or tables. Difficult-case and inspection basis: Even the most easily confused sources, conditions, or changes can be checked along the instructions, and the complete verification path is concise with no gaps or overclaims. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Within this criterion and the stated use conditions, there is no identifiable substantive weakness or feasible task-relevant improvement of material benefit. Establish full applicable coverage and actively inspect the strongest plausible shortcoming; finding no error, a single successful example or an author assertion does not prove full refinement.", "3": "Exceptional beyond conventional full refinement: first fully meet +2. An economical provenance/handoff organization makes a difficult fact, calculation or asset relationship directly auditable, reducing reconstruction beyond a complete accurate conventional source list. Identify the actual trace and saved ambiguity, without requiring new evidence or new documentation beyond the task. Apply the exceptional benefit specifically to: artifact/BUILD.md should give usable XeLaTeX commands and font requirements, and the briefing should give the paper version and brief source citations to the relevant sections, figures, or tables. In-task verification cases for that benefit: Even the most easily confused sources, conditions, or changes can be checked along the instructions, and the complete verification path is concise with no gaps or overclaims. In this task, focus on verifying the parts relevant to this item in the MLM conditional probability illustration and the SQuAD comparison rows that differ in Dev/Test, Single/Ensemble, and additional data; do not add deliverables, scenarios, or experiments beyond the task. Correct completion of these cases is only a foundation, not evidence of exceptional benefit. In reason identify the non-obvious choice, real difficulty and substantial benefit beyond an equally correct, complete and polished conventional solution, with actual evidence. No comparison artifact or unique design is required. Complexity, extra features, ornament, praise or repetition of +2 strengths is insufficient. Without demonstrated extra benefit the ceiling is +2; no score quota applies." }, "weight": 3, "benchmark_source": "vas" } ]