superlab 0.1.69 → 0.1.70

package/package-assets/shared/lab/.managed/scripts/validate_collaborator_report.py

@@ -51,6 +51,21 @@ SOURCE_SECTION_PATH_MARKERS = (
 SOURCE_SECTION_CITATION_MARKERS = ("Citation:", "引用：")
 SOURCE_SECTION_ROLE_MARKERS = ("What it established:", "What it does:", "What it measures:", "做了什么：", "衡量什么：")
 SOURCE_SECTION_LIMITATION_MARKERS = ("Limitation", "局限")
+METRIC_GUIDE_DETAIL_MARKERS = {
+    "evaluation target": ("Evaluation target:", "What is evaluated:", "评估对象：", "评估什么："),
+    "test-set prediction": ("Test-set prediction used:", "Prediction used:", "测试集预测：", "预测量："),
+    "ranking or grouping": ("Ranking or grouping step:", "Ranking step:", "Grouping step:", "排序或分组：", "排序步骤：", "分组步骤："),
+    "calculation sketch": (
+        "Aggregation / calculation sketch:",
+        "Calculation sketch:",
+        "Approximate calculation:",
+        "大致计算：",
+        "近似公式：",
+        "聚合方式：",
+    ),
+    "direction and scale": ("Direction and scale:", "Metric direction:", "方向与尺度：", "方向：", "越高/越低："),
+    "comparability boundary": ("Comparability boundary:", "What not to compare:", "可比性边界：", "不能比较："),
+}
 
 
 def parse_args():
@@ -99,6 +114,35 @@ def validate_source_sections(text: str, label: str) -> list[str]:
     return issues
 
 
+def has_marker_with_value(body: str, markers: tuple[str, ...]) -> bool:
+    for line in body.splitlines():
+        stripped = line.strip()
+        for marker in markers:
+            if marker not in stripped:
+                continue
+            value = stripped.split(marker, 1)[1].strip()
+            if value and value not in {"-", "—", "TODO", "TBD", "待补", "待定"}:
+                return True
+    return False
+
+
+def validate_metric_guide_detail(text: str, label: str) -> list[str]:
+    body = extract_section_body(text, REPORT_REQUIRED_SECTIONS["Metric Guide"])
+    if not body:
+        return []
+    missing = [
+        detail_name
+        for detail_name, markers in METRIC_GUIDE_DETAIL_MARKERS.items()
+        if not has_marker_with_value(body, markers)
+    ]
+    if not missing:
+        return []
+    return [
+        f"{label} section 'Metric Guide' must explain metric computation details: "
+        f"{', '.join(missing)}"
+    ]
+
+
 def validate(path_str: str, required_sections: dict[str, list[str]], label: str) -> list[str]:
     path = Path(path_str)
     if not path.exists():
@@ -108,7 +152,7 @@ def validate(path_str: str, required_sections: dict[str, list[str]], label: str)
     if missing:
         return [f"{label} is missing required sections: {', '.join(missing)}"]
     if label == "report.md":
-        return validate_source_sections(text, label)
+        return validate_source_sections(text, label) + validate_metric_guide_detail(text, label)
     return []
 
 

package/package-assets/shared/lab/.managed/scripts/validate_manuscript_delivery.py

@@ -38,6 +38,7 @@ REQUIRED_TABLE_NOTE_MARKERS = (
     "% Important caveat:",
 )
 WIDTH_CONTROL_NOTE_MARKER = "% Width control:"
+WIDE_PLAIN_TABULAR_COLUMN_LIMIT = 7
 TABLE_ABBREVIATION_EXCEPTIONS = {"TODO", "TBD"}
 PLACEHOLDER_TABLE_NOTE_PREFIXES = (
     "explain ",
@@ -97,6 +98,109 @@ def contains_any(text: str, needles: tuple[str, ...]) -> bool:
     return any(needle.lower() in lowered for needle in needles)
 
 
+def read_braced_group(text: str, start: int) -> tuple[str, int] | None:
+    if start >= len(text) or text[start] != "{":
+        return None
+    depth = 0
+    content_start = start + 1
+    for index in range(start, len(text)):
+        char = text[index]
+        if char == "{":
+            depth += 1
+        elif char == "}":
+            depth -= 1
+            if depth == 0:
+                return text[content_start:index], index + 1
+    return None
+
+
+def skip_whitespace(text: str, index: int) -> int:
+    while index < len(text) and text[index].isspace():
+        index += 1
+    return index
+
+
+def extract_plain_tabular_specs(text: str) -> list[str]:
+    specs: list[str] = []
+    needle = r"\begin{tabular}"
+    search_from = 0
+    while True:
+        index = text.find(needle, search_from)
+        if index == -1:
+            return specs
+        spec_start = skip_whitespace(text, index + len(needle))
+        group = read_braced_group(text, spec_start)
+        if group is not None:
+            specs.append(group[0])
+            search_from = group[1]
+        else:
+            search_from = index + len(needle)
+
+
+def count_column_spec(spec: str) -> tuple[int, bool]:
+    count = 0
+    has_width_aware_column = False
+    index = 0
+    while index < len(spec):
+        char = spec[index]
+        if char in "lcr":
+            count += 1
+            index += 1
+            continue
+        if char == "X":
+            count += 1
+            has_width_aware_column = True
+            index += 1
+            continue
+        if char in "pmb":
+            count += 1
+            has_width_aware_column = True
+            index = skip_whitespace(spec, index + 1)
+            if index < len(spec) and spec[index] == "{":
+                group = read_braced_group(spec, index)
+                index = group[1] if group is not None else index + 1
+            continue
+        if char == "*":
+            index = skip_whitespace(spec, index + 1)
+            repeat_group = read_braced_group(spec, index)
+            if repeat_group is None:
+                continue
+            repeat_text, index = repeat_group
+            index = skip_whitespace(spec, index)
+            repeated_spec_group = read_braced_group(spec, index)
+            if repeated_spec_group is None:
+                continue
+            repeated_spec, index = repeated_spec_group
+            try:
+                repeat_count = int(repeat_text.strip())
+            except ValueError:
+                repeat_count = 1
+            nested_count, nested_width_aware = count_column_spec(repeated_spec)
+            count += repeat_count * nested_count
+            has_width_aware_column = has_width_aware_column or nested_width_aware
+            continue
+        if char in "@!<>":
+            index = skip_whitespace(spec, index + 1)
+            if index < len(spec) and spec[index] == "{":
+                group = read_braced_group(spec, index)
+                index = group[1] if group is not None else index + 1
+            continue
+        index += 1
+    return count, has_width_aware_column
+
+
+def has_width_control_command(text: str) -> bool:
+    return any(
+        token in text
+        for token in (
+            r"\begin{tabularx}",
+            r"\begin{tabular*}",
+            r"\resizebox{",
+            r"\setlength{\tabcolsep}",
+        )
+    )
+
+
 def find_workflow_config(start_path: Path) -> Path | None:
     search_roots = [start_path, *start_path.parents]
     for root in search_roots:
@@ -315,6 +419,18 @@ def check_table_file(path: Path, issues: list[str], label: str):
             continue
         if value < 3.0:
             issues.append(f"{label} sets \\tabcolsep below the safe range for paper-facing main tables")
+    for spec in extract_plain_tabular_specs(text):
+        column_count, has_width_aware_column = count_column_spec(spec)
+        if (
+            column_count >= WIDE_PLAIN_TABULAR_COLUMN_LIMIT
+            and not has_width_aware_column
+            and not has_width_control_command(text)
+        ):
+            issues.append(
+                f"{label} uses a wide plain tabular layout ({column_count} columns) without a width-aware strategy; "
+                "use tabularx or p columns, split the table, move secondary metrics to appendix, "
+                "or document last-resort width control"
+            )
 
 
 def check_figure_file(path: Path, issues: list[str], label: str):

package/package-assets/shared/lab/.managed/scripts/validate_section_draft.py

@@ -241,6 +241,22 @@ INTERNAL_EXPERIMENT_PROVENANCE_PHRASES = (
     "调参运行",
     "调参轮次",
 )
+INTERNAL_EXPERIMENT_PLANNING_PATTERNS = (
+    r"current\s+[\d.]+\s+only\s+shows?.*need(?:s|ed)?\s+(?:a\s+)?(?:new\s+)?holdout",
+    r"(?:new|additional)\s+holdout\s+(?:and|or)\s+(?:more\s+)?natural(?:ized)?\s+(?:payload|attack|statement)",
+    r"(?:small[- ]batch|pilot[- ]batch).*(?:gate|gating)",
+    r"(?:freeze|freezing).*(?:payload|attack statement|trigger)",
+    r"(?:api|API).*(?:budget|cost|scale)",
+    r"新增\s*(?:holdout|外部|样本|实验).*验证",
+    r"还需要\s*新增.*验证",
+    r"后文.*边界",
+    r"当前\s*[\d.]+\s*只能说明.*不能外推.*(?:还需要|需要)",
+    r"小批量.*(?:门控|gate)",
+    r"(?:冻结|固定).*(?:payload|载荷|攻击语句|触发语句)",
+    r"(?:不能|不得).*边跑边调",
+    r"API\s*(?:规模|预算|成本)",
+    r"(?:按设计|设计上).*(?:失败|不通过).*(?:过拟合|调参)",
+)
 INTERNAL_CONFIG_LABEL_PATTERN = re.compile(
     r"\b[a-z]{1,4}\d+(?:[-_][a-z]?\d+(?:\.\d+)?){1,4}\b",
     flags=re.IGNORECASE,
@@ -265,6 +281,10 @@ def check_common_section_gate_risks(text: str, issues: list[str]):
         issues.append(
             "reader-facing prose appears to contain internal experiment provenance or tuning/config labels; move run provenance to workflow notes or map it to paper-facing diagnostic terminology"
         )
+    if any(re.search(pattern, prose_text, flags=re.IGNORECASE) for pattern in INTERNAL_EXPERIMENT_PLANNING_PATTERNS):
+        issues.append(
+            "reader-facing prose appears to contain internal experiment planning or holdout-expansion rationale; keep plans, gates, payload-freezing notes, and future validation logistics in workflow artifacts instead of the manuscript"
+        )
     if contains_any(
         prose_text,
         (

package/package-assets/shared/lab/.managed/templates/final-report.md

@@ -51,6 +51,12 @@
 - Primary metric plain-language explanation:
 - Secondary metric plain-language explanation:
 - Health or support metrics and why they are not the main claim:
+- Evaluation target:
+- Test-set prediction used:
+- Ranking or grouping step:
+- Aggregation / calculation sketch:
+- Direction and scale:
+- Comparability boundary:
 
 ## Background Sources
 

package/package-assets/shared/lab/.managed/templates/main-tables.md

@@ -17,6 +17,12 @@
 - Primary metric plain-language explanation:
 - Secondary metric plain-language explanation:
 - Health or support metrics and how to read them:
+- Evaluation target:
+- Test-set prediction used:
+- Ranking or grouping step:
+- Aggregation / calculation sketch:
+- Direction and scale:
+- Comparability boundary:
 
 ## Final Performance Summary
 

package/package-assets/shared/lab/.managed/templates/paper-table.tex

@@ -2,18 +2,18 @@
 \caption{One-sentence message of the table and the evaluation protocol.}
 \label{tab:placeholder}
 \centering
-\begin{tabular}{lcc}
+\begin{tabularx}{\linewidth}{>{\raggedright\arraybackslash}Xcc}
 \toprule
 Method & Metric 1 $\uparrow$ & Metric 2 $\uparrow$ \\
 \midrule
 Ours & 0.0000 & 0.0000 \\
 Baseline & 0.0000 & 0.0000 \\
 \bottomrule
-\end{tabular}
+\end{tabularx}
 % Rows: explain what each row represents.
 % Columns: explain what each column represents and its direction.
 % Metric definitions: expand local abbreviations, units, denominators, or event conditions.
 % Comparison scope: explain which setting, split, attack family, or benchmark scope this table covers.
 % Important caveat: state any omitted metrics, zero-valued metrics, or appendix-only reporting decision.
-% Width control: first shorten headers, move secondary metrics out of the main table, and reduce or split columns; only then adjust \setlength{\tabcolsep}{...} conservatively or use \resizebox{\linewidth}{!}{...} as a documented last resort.
+% Width control: default to bounded columns with tabularx or p{...}; first shorten headers, move secondary metrics out of the main table, and reduce or split columns; only then adjust \setlength{\tabcolsep}{...} conservatively or use \resizebox{\linewidth}{!}{...} as a documented last resort.
 \end{table}

package/package-assets/shared/lab/.managed/templates/paper.tex

@@ -4,6 +4,8 @@
 \usepackage{hyperref}
 \usepackage{graphicx}
 \usepackage{booktabs}
+\usepackage{array}
+\usepackage{tabularx}
 
 \title{Paper Title}
 \author{Author Name}

package/package-assets/shared/lab/.managed/templates/write-iteration.md

@@ -86,6 +86,9 @@
 - Were all abbreviations expanded at local first mention:
 - Did each main table include a local table note:
 - Can a reader interpret rows and columns without chasing Method:
+- Table width audit:
+- Did any main table use a wide plain `tabular` layout:
+- If width control was needed, was the table first shortened, split, moved partly to appendix, or converted to `tabularx` / bounded columns before using `\tabcolsep` or `\resizebox`:
 - If this section used canonical short names before their defining section, was a local naming bridge added:
 - Did model and ablation labels stay canonical instead of drifting into narrative aliases:
 
@@ -141,6 +144,7 @@
 - Did the round avoid copying reference wording, claims, metrics, captions, or conclusions:
 - Did final prose avoid service-style or AI-assistant meta language:
 - Did final prose avoid workflow-only placeholder language:
+- Did final prose avoid internal experiment planning, future-holdout logistics, gates, payload-freezing notes, API-budget notes, and automation triage language:
 - Validator command and result:
 
 ## Decision

package/package-assets/shared/skills/lab/SKILL.md

@@ -210,6 +210,7 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Read `.lab/context/mission.md`, `.lab/context/state.md`, `.lab/context/workflow-state.md`, `.lab/context/decisions.md`, `.lab/context/evidence-index.md`, and `.lab/context/data-decisions.md` before drafting.
 - Read `.lab/context/eval-protocol.md` before choosing tables, thresholds, or final result framing.
 - Keep metric definitions, comparison semantics, and implementation references anchored to the approved evaluation protocol instead of re-deriving them during reporting.
+- In `report.md`, explain each primary metric with a computation guide: what is evaluated, which test-set predictions or scores are used, whether examples are sorted, grouped, bucketed, or paired, how the value is aggregated or approximately calculated, what direction and scale mean, and what cannot be compared across datasets, splits, or implementations.
 - Aggregate them with `.lab/.managed/scripts/summarize_iterations.py`.
 - Write the final document with `.lab/.managed/templates/final-report.md`, the managed table summary with `.lab/.managed/templates/main-tables.md`, and the internal handoff with `.lab/.managed/templates/artifact-status.md`.
 - Keep failed attempts and limitations visible.
@@ -272,10 +273,12 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Use the same metric names across Method, Experiments, captions, table headers, table notes, and result summaries; remove forbidden aliases from reader-facing LaTeX instead of letting legacy metric names drift.
 - Run `.lab/.managed/scripts/validate_metric_glossary.py` in metric-bearing draft, final-draft, or export rounds and record the result in the latest write iteration artifact.
 - Do not treat `\resizebox{\linewidth}{!}{...}` as the default main-table fit strategy.
-- Fit paper-facing main tables by redesign first: shorten headers, move secondary metrics out of the main table, reduce or split columns, then adjust `\tabcolsep` conservatively; only use `\resizebox` as a last resort and document why.
+- Wide plain `tabular` layouts with many columns are not manuscript-ready by default; prefer `tabularx` or bounded `p{...}` columns for text-heavy or multi-metric tables.
+- Fit paper-facing main tables by redesign first: shorten headers, move secondary metrics out of the main table, reduce or split columns, prefer `tabularx` or bounded columns, then adjust `\tabcolsep` conservatively; only use `\resizebox` as a last resort and document why.
 - Keep `\tabcolsep` adjustments conservative and avoid shrinking below a roughly readable floor for paper-facing main tables.
 - Do not rely on `\scriptsize` or `\tiny` as the default way to make a main table fit.
 - Keep internal identifiers, tuning-run labels, probe names, config strings, rerun ids, and package labels out of prose unless they are mapped once for the reader and then moved back out of prose.
+- Keep internal experiment planning out of manuscript prose: future holdout expansion, small-batch gates, payload freezing, API budgets, automation decisions, and overfitting triage logic belong in lab artifacts, not paper-facing sections.
 - Do not rely on unexplained jargon density as a substitute for academic tone.
 - Bind each claim to evidence from `report`, iteration reports, or normalized summaries.
 - Use the write-stage contract in `.codex/skills/lab/stages/write.md` or `.claude/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section-specific references, validator calls, asset coverage, and final manuscript gates.

package/package-assets/shared/skills/lab/references/paper-writing/section-style-policies.md

@@ -122,6 +122,7 @@ These are paper-facing defaults. They are not project-specific branding rules.
 - Self-evaluations such as "结果也很清楚", "the defense results are very clear", or "the table is self-explanatory".
 - Layout-process commentary in scientific prose, such as "由于表列较多，这里采用页宽自适应排版" or "we use page-width adaptive layout here".
 - Claims that a table "proves" something when the evidence only supports a bounded empirical result.
+- Internal experiment-planning prose, such as "还需要新增 holdout", "小批量门控", "冻结 payload", "不能边跑边调", "API 规模估计", or "if all scores are 1.0000, treat it as overfitting".
 - Service-style or AI-assistant meta language such as "用户说", "按你的要求", "我来解释", "let me explain", or "as requested by the user".
 - Workflow-only placeholder language such as "图的意图", "资产意图", "占位符", "workflow-language", or "sync this wording".
 

package/package-assets/shared/skills/lab/stages/report.md

@@ -10,6 +10,7 @@
 - method overview
 - selected metrics summary
 - plain-language metric guide
+- metric computation guide that explains what is evaluated, which test-set predictions are used, whether examples are sorted or grouped, how values are aggregated or approximately calculated, metric direction and scale, and comparability boundaries
 - background sources
 - method and baseline sources
 - metric sources
@@ -52,6 +53,8 @@
 - Do not restate metric definitions, baseline behavior, or comparison implementations from memory; use the approved evaluation protocol and its recorded sources.
 - Carry the approved `Primary metrics`, `Secondary metrics`, and `Required terminal evidence` into both the report and the managed main-tables artifact.
 - Explain the selected primary and secondary metrics in plain language for the user: what each metric measures, whether higher or lower is better, and whether it is a main result metric or only a health/support metric.
+- For every primary metric, also explain enough of the computation for a collaborator to reproduce the idea without reading code: what is evaluated, which test-set predictions or scores are used, whether the examples are sorted, bucketed, grouped, or paired, how the resulting values are aggregated or approximately calculated, what direction and scale mean, and which comparisons are invalid across datasets, splits, or metric implementations.
+- If a metric depends on ranking, the report must name the ranking score and the order. If it depends on a contrast, the report must name the compared conditions or groups. If it depends on an average, rate, area, threshold crossing, or recovery amount, the report must give a simple calculation sketch.
 - If coverage, completeness, confidence, or similar health metrics appear, explicitly say that they describe experimental reliability rather than the main scientific effect.
 - Pull the core background references, method or baseline references, and metric references out of the approved evaluation protocol instead of hiding them in `.lab/context/*`.
 - Treat `report.md` as an external-review-ready memo. Source sections must not rely on local file paths or internal provenance notes; they must give a few human-readable anchor references instead.

package/package-assets/shared/skills/lab/stages/write.md

@@ -165,6 +165,8 @@ Do not enter prose polish until the current section has passed the reference-con
 - Do not use labels containing `_` or `-` in reader-facing prose.
 - Keep internal identifiers, config keys, and experiment package labels out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose.
 - Keep run provenance such as tuning-run labels, probe names, internal config strings, rerun ids, and package labels out of reader-facing prose. If the evidence is useful, rewrite it as a bounded paper-facing diagnostic or move the raw provenance to workflow notes or appendix metadata.
+- Keep internal experiment planning out of reader-facing prose. Do not write paper sentences that explain future holdout expansion, small-batch gates, payload freezing, API budget, "if all scores are 1.0000 then treat as overfitting", or why a next automation round is needed.
+- When an experiment boundary matters, report only the scientific scope already supported by the evidence. Put the operational plan for collecting new attacks, new papers, new markers, or additional holdout cases into `.lab/changes/`, `.lab/iterations/`, or report artifacts, not into manuscript sections.
 - Do not use unexplained terminology density as a substitute for academic tone.
 - Keep service-style or AI-assistant meta language out of manuscript prose. Phrases such as "用户说", "按你的要求", "我来解释", "下面我", "this version", or "as requested by the user" belong in workflow notes, not in paper-facing sections, captions, table notes, or analysis assets.
 - Keep workflow-only placeholder language out of manuscript prose. Phrases such as "图的意图", "资产意图", "占位符", "workflow-language", "translation layer", or "sync this wording" belong in authoring artifacts, not in reader-facing LaTeX.
@@ -178,10 +180,12 @@ Do not enter prose polish until the current section has passed the reference-con
 - If a metric's denominator, event condition, score scale, or comparison scope differs by setting, define a separate entry or explicitly scope the metric in `.lab/writing/metric-glossary.md`.
 - Deprecated or forbidden metric aliases must be removed from reader-facing LaTeX instead of explained away locally.
 - Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table.
-- Main-table width control should follow this order: shorten headers while preserving local explanations, move secondary metrics to appendix-only, reduce or split columns, adjust `\tabcolsep` conservatively, and only then consider `\resizebox` as a last resort.
+- Wide plain `tabular` layouts with many columns are not manuscript-ready by default; final/export validation should force a width-aware table design instead of silently accepting likely overfull tables.
+- Main-table width control should follow this order: shorten headers while preserving local explanations, move secondary metrics to appendix-only, reduce or split columns, prefer `tabularx` or bounded `p{...}` columns, adjust `\tabcolsep` conservatively, and only then consider `\resizebox` as a last resort.
 - When `\tabcolsep` is adjusted for a paper-facing main table, keep it in a safe range and avoid shrinking below roughly `3pt`; prefer `4pt` or `5pt` when a small reduction is enough.
 - Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. If a table only fits after aggressive font shrinking, redesign the table instead of forcing it into the page.
 - If a paper-facing main table uses `\resizebox` or non-default width control, explain the width-control rationale in the same table note.
+- Prefer `tabularx` for paper-facing main tables whose first column or text-heavy columns need bounded line wrapping; use plain `tabular` only for compact tables with a small column count.
 - Every main table should have a short table-introduction sentence before it and a short interpretation sentence after it so the reader knows what question the table answers and how to read the result.
 - Build the paper asset plan before prose when the section carries introduction, experimental, method, related-work, or conclusion claims:
   - record the asset coverage targets and gaps for the current paper
@@ -221,6 +225,7 @@ Do not enter prose polish until the current section has passed the reference-con
 - Table assets must also include a local table note that explains row meaning, column meaning, metric definitions, comparison scope, and any important caveat.
 - The local table note must contain real reader-facing explanations, not the default template phrases such as "explain what each row represents" or "expand local abbreviations".
 - Table assets must not rely on aggressive width hacks by default; if width control is still needed after table redesign, document it locally and keep it readable.
+- Table assets with seven or more columns should be split, moved partly to appendix, or written with width-aware columns such as `tabularx` or `p{...}` instead of a plain `tabular` layout.
 - Figure placeholders may record what the final figure should show and why the reader needs it in authoring comments, the paper plan, or the write-iteration artifact, but the caption itself must remain paper-facing and must not contain "Figure intent", "图的意图", "asset intent", "占位符", or similar workflow language.
 - Core asset coverage for a paper-facing final draft should include a problem-setting or teaser figure, a method overview figure, a results overview figure, a main-results table, an ablation table, and one additional analysis asset.
 - Keep `.lab/writing/plan.md` synchronized with the current table plan, figure plan, citation plan, and section-to-asset map whenever manuscript assets change.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlab",
-  "version": "0.1.69",
+  "version": "0.1.70",
   "description": "Strict /lab research workflow installer for Codex and Claude",
   "keywords": [
     "codex",