superlab 0.1.70 → 0.1.72

package/package-assets/claude/commands/lab:write.md

@@ -9,4 +9,5 @@ Use the installed `lab` skill at `.claude/skills/lab/SKILL.md`.
 Execute the requested `/lab-write` command against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
+Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations.
 This command runs the `write` stage of the lab workflow. Use `.claude/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

package/package-assets/claude/commands/lab/357/274/232idea.md

@@ -7,4 +7,4 @@ argument-hint: idea or research problem
 Use the installed `lab` skill at `.claude/skills/lab/SKILL.md`.
 
 Execute the requested `/lab-idea` command against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `idea` stage of the lab workflow. Use `.claude/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.
+This command runs the `idea` stage of the lab workflow. Use `.claude/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, contribution-vs-insight separation, insight evidence chain, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what insight the idea teaches beyond the artifact, what evidence chain supports that insight, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the core insight, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.

package/package-assets/claude/commands/lab/357/274/232report.md

@@ -7,4 +7,5 @@ argument-hint: report context
 Use the installed `lab` skill at `.claude/skills/lab/SKILL.md`.
 
 Execute the requested `/lab-report` command against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
+The report must put the core insight near the top: what was learned beyond the produced artifact, what evidence supports it, what action or design implication follows, and what boundary still applies. Use main tables and ablations as diagnostic evidence for that insight rather than only containers for metric values.
 This command runs the `report` stage of the lab workflow. It must produce a user-facing final report plus the managed `main-tables.md` artifact, explicitly carry the approved primary and secondary metrics forward, explain the selected metrics in plain language, say which metrics are only health or support metrics, and explain what each main table proves or does not prove.

package/package-assets/claude/commands/lab/357/274/232write.md

@@ -9,4 +9,5 @@ Use the installed `lab` skill at `.claude/skills/lab/SKILL.md`.
 Execute the requested `/lab-write` command against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
+Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations.
 This command runs the `write` stage of the lab workflow. Use `.claude/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

package/package-assets/codex/prompts/lab/idea.md

@@ -6,4 +6,4 @@ argument-hint: idea or research problem
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:idea` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.
+This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, contribution-vs-insight separation, insight evidence chain, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what insight the idea teaches beyond the artifact, what evidence chain supports that insight, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the core insight, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.

package/package-assets/codex/prompts/lab/report.md

@@ -6,4 +6,5 @@ argument-hint: report context
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:report` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
+The report must put the core insight near the top: what was learned beyond the produced artifact, what evidence supports it, what action or design implication follows, and what boundary still applies. Use main tables and ablations as diagnostic evidence for that insight rather than only containers for metric values.
 This command runs the `/lab:report` stage. It must produce a user-facing final report plus the managed `main-tables.md` artifact, explicitly carry the approved primary and secondary metrics forward, explain the selected metrics in plain language, say which metrics are only health or support metrics, and explain what each main table proves or does not prove.

package/package-assets/codex/prompts/lab/write.md

@@ -8,4 +8,4 @@ Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 Execute the requested `/lab:write` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
-This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.
+This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, the insight integration audit, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

package/package-assets/codex/prompts/lab-idea.md

@@ -6,4 +6,4 @@ argument-hint: idea or research problem
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:idea` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.
+This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, contribution-vs-insight separation, insight evidence chain, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what insight the idea teaches beyond the artifact, what evidence chain supports that insight, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the core insight, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.

package/package-assets/codex/prompts/lab-report.md

@@ -6,4 +6,5 @@ argument-hint: report context
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:report` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
+The report must put the core insight near the top: what was learned beyond the produced artifact, what evidence supports it, what action or design implication follows, and what boundary still applies. Use main tables and ablations as diagnostic evidence for that insight rather than only containers for metric values.
 This command runs the `/lab:report` stage. It must produce a user-facing final report plus the managed `main-tables.md` artifact, explicitly carry the approved primary and secondary metrics forward, explain the selected metrics in plain language, say which metrics are only health or support metrics, and explain what each main table proves or does not prove.

package/package-assets/codex/prompts/lab-write.md

@@ -8,4 +8,4 @@ Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 Execute the requested `/lab:write` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
-This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.
+This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, the insight integration audit, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

package/package-assets/codex/prompts/lab:idea.md

@@ -6,4 +6,4 @@ argument-hint: idea or research problem
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:idea` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.
+This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, contribution-vs-insight separation, insight evidence chain, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what insight the idea teaches beyond the artifact, what evidence chain supports that insight, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the core insight, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.

package/package-assets/codex/prompts/lab:report.md

@@ -6,4 +6,5 @@ argument-hint: report context
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:report` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
+The report must put the core insight near the top: what was learned beyond the produced artifact, what evidence supports it, what action or design implication follows, and what boundary still applies. Use main tables and ablations as diagnostic evidence for that insight rather than only containers for metric values.
 This command runs the `/lab:report` stage. It must produce a user-facing final report plus the managed `main-tables.md` artifact, explicitly carry the approved primary and secondary metrics forward, explain the selected metrics in plain language, say which metrics are only health or support metrics, and explain what each main table proves or does not prove.

package/package-assets/codex/prompts/lab:write.md

@@ -8,4 +8,4 @@ Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 Execute the requested `/lab:write` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
-This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.
+This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, the insight integration audit, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

package/package-assets/codex/prompts/lab/357/274/232idea.md

@@ -6,4 +6,4 @@ argument-hint: idea or research problem
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:idea` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.
+This command runs the `/lab:idea` stage. Use `.codex/skills/lab/stages/idea.md` as the single source of truth for the two brainstorm passes, two literature sweeps, closest-prior comparison, source-backed proposal memo, evaluation sketch, contribution-vs-insight separation, insight evidence chain, tentative contributions, user guidance, minimum viable experiment, convergence status, and approval gate. An explicit `/lab:idea` request defaults to a full-stage rerun, even if the user asks to reorganize or rewrite an existing idea memo. Only switch to rewrite-only mode when the user explicitly asks to only reorganize or rewrite the existing idea artifact without new searching. Rewrite-only mode may improve structure or readability, but it must not change the recommendation, paper-fit judgment, convergence status, or canonical context; it must say that the stage remains unconverged because the literature sweeps were not rerun. Start with brainstorm pass 1 over 3-4 candidate directions. For each candidate direction, explain what it is, why it matters, roughly how it would work, what problem it solves, and its main risk. Run literature sweep 1 with real closest-prior references for each direction, narrow the field with brainstorm pass 2 to 1-2 surviving directions, explain why each survivor remains, why each rejected direction was dropped, and why the narrowed recommendation is stronger now, then run literature sweep 2 to build the final source bundle before producing a collaborator-readable recommendation. Materialize or update `.lab/writing/idea.md` and `.lab/writing/idea-source-log.md` before any final recommendation, paper-fit judgment, or mission writeback. Do not end the stage with a chat-only brainstorm; if the work is still unconverged, say so explicitly, list what is still missing, and stop there. The final idea memo must explain the real-world scenario, the problem solved, why current methods fall short, roughly how the idea would work, how it would be evaluated, what the tentative contributions are, what insight the idea teaches beyond the artifact, what evidence chain supports that insight, what is already source-backed, what is still hypothesis-only, and what the user should decide next. It must also include a user-visible literature summary naming the closest prior found, the recent strong papers found, and what existing work still does not solve. In the final user-facing summary, say what current methods do, why they still fall short, how the proposed direction differs, the core insight, the rough approach, the main risk, and where to read the full idea artifact and source log. Keep `.lab/writing/idea-source-log.md` synchronized with the actual search queries, bucketed sources, and final source count used in both sweeps. The literature bundle should default to about 20 sources unless the field is genuinely narrow and that smaller bundle is explicitly justified. Only after `.lab/.managed/scripts/validate_idea_artifact.py` passes may the stage present a final recommendation as converged.

package/package-assets/codex/prompts/lab/357/274/232report.md

@@ -6,4 +6,5 @@ argument-hint: report context
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:report` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
+The report must put the core insight near the top: what was learned beyond the produced artifact, what evidence supports it, what action or design implication follows, and what boundary still applies. Use main tables and ablations as diagnostic evidence for that insight rather than only containers for metric values.
 This command runs the `/lab:report` stage. It must produce a user-facing final report plus the managed `main-tables.md` artifact, explicitly carry the approved primary and secondary metrics forward, explain the selected metrics in plain language, say which metrics are only health or support metrics, and explain what each main table proves or does not prove.

package/package-assets/codex/prompts/lab/357/274/232write.md

@@ -8,4 +8,4 @@ Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 Execute the requested `/lab:write` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
 When the user provides reference PDFs, paper URLs, local reference-paper paths, or asks to write by reference, stay within the write stage and switch to reference-guided deep-write. Extract structure, map section/subsection slots, paragraph roles, table/figure roles, and bridge logic to the current paper, record the consumption plan, and only then draft prose. The current section must visibly realize the mapped slots; do not treat a consumption plan as enough. Reuse structure only; do not copy wording, claims, metrics, captions, or conclusions. Keep service-style or AI-assistant meta language and workflow-only placeholder language out of paper-facing prose.
 When Method, Experiments, captions, tables, or analysis assets introduce or revise reported metrics, create or update `.lab/writing/metric-glossary.md` before prose polish. Each metric must define its paper-facing name, approved short name, table/header label, plain-language definition, calculation, unit or denominator, direction, scope or conditions, allowed aliases, forbidden aliases, and first-use location. Use the same metric names across prose, captions, table notes, table headers, and result summaries. Run `validate_metric_glossary.py` and remove forbidden aliases from reader-facing LaTeX before finalizing the round.
-This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.
+This command runs the `/lab:write` stage. Use `.codex/skills/lab/stages/write.md` as the single source of truth for template choice, paper-plan requirements, section references, validator gates, asset coverage, and final manuscript rules. Carry the same core insight anchor through the paper: Introduction creates the cognitive contrast, Method turns the insight into design motivation, Experiments diagnose it with evidence, and Conclusion states the broader principle and boundary. Do not create a standalone `Our Insights` section just to satisfy this; weave the insight into motivation, mechanism, evidence, and limitations. Read the matching paper-writing reference, the current section block in `section-style-policies.md`, and any bundled example-bank files for the requested section, revise only one section, and keep draft rounds warning-only while final-draft or export rounds must satisfy the write-stage acceptance gates. Draft ordinary manuscript rounds in `workflow_language`, and ordinary `.tex` section drafts must stay in `workflow_language` instead of treating `paper_language` as the default draft language. When `workflow_language` and `paper_language` differ, treat the workflow-language paper layer as the default ordinary working layer. Resolve the active paper topology from `.lab/config/workflow.json` before drafting: the active canonical root is `<deliverables_root>/paper/`, and when workflow-language is active its root is `<deliverables_root>/paper/workflow-language/`. Ordinary write rounds should still edit one target paper layer at a time rather than silently refreshing both language layers. If the user names a concrete file or layer, treat that as the only target for the round unless they also explicitly request synchronization. Classify the named target path before editing it. Only active-layer targets count as managed manuscript rounds; legacy side layers such as `review_zh`, `translation_zh`, `sections_zh`, or stale `deliverables/.../workflow-language/*.md` paths are out-of-band/legacy edits and must not silently replace the active paper topology. If a workflow-language paper layer is active and the round still targets the canonical manuscript, record why canonical-only writing was acceptable in the write iteration artifact. If `paper_language_finalization_decision=convert-to-paper-language`, explicit canonical-manuscript work may target the canonical `paper_language` manuscript, but that does not make canonical the default ordinary working layer while workflow-language remains active. Treat the workflow-language paper layer as a real persisted artifact rather than a review layer, and preserve it as a full LaTeX mirror with `workflow-language/main.tex`, `workflow-language/references.bib`, `workflow-language/sections/*.tex`, `workflow-language/tables/*.tex`, `workflow-language/figures/*.tex`, and `workflow-language/analysis/analysis-asset.tex`. Do not write new workflow-language output to deprecated review-layer paths such as `docs/lab/paper/review_zh/`. Maintain `.lab/writing/terminology-glossary.md` as the write-stage glossary for full forms, approved short forms, reader-facing explanations, and aliases. Apply the same academic readability standard in every language: when the round introduces or revises key terms, abbreviations, metrics, mechanism names, or system labels, use the full form first, define any short form at first mention, explain what the term is and why it matters here, keep one natural-language paper-facing name per concept, use natural-language full names in prose, do not use labels containing `_` or `-` in reader-facing prose, apply the same first-mention rule to table headers, table captions, table notes, and figure captions or labels, do not assume a fixed drafting order such as Method before Experiments, add a local naming bridge when a section uses canonical short names before their defining section has been drafted, and reuse the canonical label instead of replacing it with a narrative alias. Follow the current section's encouraged, discouraged, and banned expression lists from `section-style-policies.md`; section-specific banned expressions take priority over prose-polish goals. Before any additional tighten, compress, or polish pass on the same section, run a section-level acceptance gate first. That gate must explicitly confirm naming consistency, adjacent-section consistency, claim, metric, and ranking consistency with the current evidence, local clarity, local concision, and section-style compliance. If the round changes the paper's canonical experiment or evaluation protocol, treat that change as a canonical replacement unless the user explicitly scoped it as supplementary or appendix-only, run a paper-wide impact audit before more polishing, update the highest-impact stale sections and assets first, and do not default to translation/workflow-layer sync work unless the user explicitly asked for it or the language-finalization workflow requires it. Only edit both the canonical manuscript and the workflow-language paper layer in the same round when the user explicitly asks for cross-language synchronization or when a final-draft/export language-finalization step requires both layers to be refreshed together. Do not treat a routine tighten/compress/polish request as an instruction to sync the workflow-language companion. For export or remote-publication rounds, if `paper_language_finalization_decision=convert-to-paper-language`, include the workflow-language paper layer in the exported or pushed bundle by default. Allow canonical-only export or remote publication only when the user explicitly asked for it or when the remote target forbids extra files. If any gate item is unresolved, or if a banned expression or move from the current section policy remains, spend the round fixing that blocker instead of polishing sentences further, and do not default the next-step recommendation to another polish pass. Main tables must be locally self-contained: the title, header, note, and adjacent prose should tell the reader what each row and column means, the metric direction, and any relevant unit, denominator, or event condition. Short headers remain allowed, but abbreviations in paper-facing tables must be expanded locally in the same table. If Method or Experiments prose promises a metric family, the main table set must either expose those metrics directly or explicitly mark the missing ones as appendix-only and explain why. If a metric is measured but omitted because it is zero everywhere, redundant, or appendix-only, state that disposition explicitly in the table note instead of silently dropping it. Do not treat `\resizebox{\linewidth}{!}{...}` as the default way to fit a main table. Fit 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, keep width changes readable, and explain the width-control rationale locally in the same table note. Do not use `\scriptsize` or `\tiny` as the default main-table fit strategy. Keep internal identifiers out of reader-facing prose unless they are mapped once for the reader and then moved back out of prose, and record the terminology-clarity self-check, the section-level acceptance gate, section-style policy compliance, the protocol/scope impact audit, the export or remote bundle audit, the round target layer, any canonical-only justification while workflow-language was active, any cross-language sync justification, the active canonical/workflow-language roots, the resolved target path role, any out-of-band justification, the insight integration audit, and the table-semantics audit in the write iteration artifact. If the manuscript would start from the managed scaffold and no template decision is recorded yet, ask once whether to keep the default scaffold or attach a template directory first. If finalization reaches a round where `workflow_language` and `paper_language` differ, finish and preserve the workflow-language paper layer first, then ask once whether to keep the draft language or convert the canonical manuscript to `paper_language`, persist that answer, record both the language decision and the workflow-language paper-layer path in the latest write iteration, and only then edit the final manuscript in the chosen language.

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

@@ -66,6 +66,32 @@ METRIC_GUIDE_DETAIL_MARKERS = {
     "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："),
 }
+READER_SUMMARY_INSIGHT_MARKERS = {
+    "core insight": ("Core insight:", "Key insight:", "核心洞见：", "关键洞见："),
+    "supporting evidence": (
+        "Evidence that supports the insight:",
+        "What evidence makes this insight plausible:",
+        "支撑该洞见的证据：",
+        "为什么这个洞见可信：",
+    ),
+    "decision or action implication": (
+        "Decision or action implication:",
+        "Action implication:",
+        "Design implication:",
+        "决策或行动含义：",
+        "行动含义：",
+        "设计含义：",
+    ),
+}
+METHOD_INSIGHT_MARKERS = {
+    "mechanism": ("Mechanism tested or explained:", "Underlying mechanism:", "解释或验证的机制：", "底层机制："),
+    "why design follows": (
+        "Why the design follows from the insight:",
+        "Why the method follows from the insight:",
+        "为什么该设计来自这个洞见：",
+        "为什么方法由这个洞见推出：",
+    ),
+}
 
 
 def parse_args():
@@ -143,6 +169,34 @@ def validate_metric_guide_detail(text: str, label: str) -> list[str]:
     ]
 
 
+def validate_insight_detail(text: str, label: str) -> list[str]:
+    issues = []
+    reader_summary = extract_section_body(text, REPORT_REQUIRED_SECTIONS["Reader Summary"])
+    if reader_summary:
+        missing = [
+            detail_name
+            for detail_name, markers in READER_SUMMARY_INSIGHT_MARKERS.items()
+            if not has_marker_with_value(reader_summary, markers)
+        ]
+        if missing:
+            issues.append(
+                f"{label} section 'Reader Summary' must carry insight-level interpretation: {', '.join(missing)}"
+            )
+
+    method_overview = extract_section_body(text, REPORT_REQUIRED_SECTIONS["Method Overview"])
+    if method_overview:
+        missing = [
+            detail_name
+            for detail_name, markers in METHOD_INSIGHT_MARKERS.items()
+            if not has_marker_with_value(method_overview, markers)
+        ]
+        if missing:
+            issues.append(
+                f"{label} section 'Method Overview' must explain mechanism-level insight: {', '.join(missing)}"
+            )
+    return issues
+
+
 def validate(path_str: str, required_sections: dict[str, list[str]], label: str) -> list[str]:
     path = Path(path_str)
     if not path.exists():
@@ -152,7 +206,7 @@ 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) + validate_metric_guide_detail(text, label)
+        return validate_source_sections(text, label) + validate_metric_guide_detail(text, label) + validate_insight_detail(text, label)
     return []
 
 

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

@@ -21,6 +21,8 @@ REQUIRED_SECTIONS = {
         r"^##\s+Literature Summary for Recommendation\s*$",
         r"^##\s+用于最终推荐的文献摘要\s*$",
     ],
+    "Contribution vs Insight": [r"^##\s+Contribution vs Insight\s*$", r"^##\s+贡献与洞见\s*$"],
+    "Insight Evidence Chain": [r"^##\s+Insight Evidence Chain\s*$", r"^##\s+洞见证据链\s*$"],
     "Rough Approach": [r"^##\s+Rough Approach\s*$", r"^##\s+我们准备怎么做\s*$"],
     "Problem Solved": [r"^##\s+Problem Solved\s*$", r"^##\s+解决了什么问题\s*$"],
     "Evaluation Sketch": [r"^##\s+Evaluation Sketch\s*$", r"^##\s+评测草图\s*$"],
@@ -63,6 +65,48 @@ LITERATURE_SUMMARY_FIELDS = {
     ),
 }
 
+CONTRIBUTION_INSIGHT_FIELDS = {
+    "contribution": ("Contribution", "What we build", "What this contributes", "贡献", "我们做了什么"),
+    "insight": ("Insight", "What we learn", "洞见", "我们学到了什么"),
+    "core_insight_anchor": (
+        "Core insight anchor sentence",
+        "Core insight anchor",
+        "Insight anchor",
+        "核心洞见锚点句",
+        "核心洞见锚点",
+        "洞见锚点",
+    ),
+    "beyond_artifact": (
+        "Why the insight matters beyond the artifact",
+        "Why it matters beyond the artifact",
+        "为什么这个洞见不只属于这个工件",
+        "为什么不只是工程产物",
+    ),
+    "action_or_community_value": (
+        "Action or community value",
+        "Community value",
+        "Decision value",
+        "行动或社区价值",
+        "行动价值",
+        "社区价值",
+    ),
+}
+
+INSIGHT_CHAIN_FIELDS = {
+    "observation": ("Observation", "现象", "反直觉现象"),
+    "existing_explanations_fail": (
+        "Why existing explanations fail",
+        "Existing explanations fail",
+        "为什么现有解释不够",
+        "旧解释为什么失败",
+    ),
+    "core_insight": ("Core insight", "核心洞见"),
+    "mechanism": ("Mechanism", "机制", "机制展开"),
+    "validation": ("Validation tests", "Validation plan", "验证实验", "验证计划"),
+    "generalization": ("Generalization or action implication", "Generalization", "Action implication", "推广性或行动含义", "推广性", "行动含义"),
+    "prediction": ("Prediction", "预测能力", "可验证预测"),
+}
+
 SOURCE_LOG_SECTIONS = {
     "Search Intent": [r"^##\s+Search Intent\s*$", r"^##\s+检索意图\s*$"],
     "Sweep 1 Log": [r"^##\s+Sweep 1 Log\s*$", r"^##\s+第一轮检索记录\s*$"],
@@ -322,6 +366,37 @@ def validate_content(text: str) -> list[str]:
         if count_references(literature_summary) < 2:
             issues.append("idea artifact literature summary for recommendation is missing real references")
 
+    contribution_insight = extract_section_body(text, REQUIRED_SECTIONS["Contribution vs Insight"])
+    if not contribution_insight:
+        issues.append("idea artifact is missing contribution vs insight")
+    else:
+        for field_name, labels in CONTRIBUTION_INSIGHT_FIELDS.items():
+            if not has_field_value(contribution_insight, labels):
+                issues.append(f"idea artifact contribution vs insight is missing {field_name}")
+        contribution_value = ""
+        insight_value = ""
+        for labels, target_name in (
+            (CONTRIBUTION_INSIGHT_FIELDS["contribution"], "contribution"),
+            (CONTRIBUTION_INSIGHT_FIELDS["insight"], "insight"),
+        ):
+            for label in labels:
+                pattern = re.compile(rf"^\s*(?:-|\d+\.)\s*{re.escape(label)}[:：][ \t]*([^\n]+?)\s*$", flags=re.MULTILINE | re.IGNORECASE)
+                match = pattern.search(contribution_insight)
+                if match and target_name == "contribution":
+                    contribution_value = match.group(1).strip()
+                if match and target_name == "insight":
+                    insight_value = match.group(1).strip()
+        if contribution_value and insight_value and contribution_value == insight_value:
+            issues.append("idea artifact must distinguish contribution from insight instead of repeating the same statement")
+
+    insight_chain = extract_section_body(text, REQUIRED_SECTIONS["Insight Evidence Chain"])
+    if not insight_chain:
+        issues.append("idea artifact is missing an insight evidence chain")
+    else:
+        for field_name, labels in INSIGHT_CHAIN_FIELDS.items():
+            if not has_field_value(insight_chain, labels):
+                issues.append(f"idea artifact insight evidence chain is missing {field_name}")
+
     rough_approach = extract_section_body(text, REQUIRED_SECTIONS["Rough Approach"])
     if not contains_any(rough_approach, ("plain-language", "how this would work", "粗略做法", "怎么做", "why this design", "为什么")):
         issues.append("idea artifact is missing a rough plain-language approach")