paper-workflow-orchestrator

MathModel Standard Orchestrator

Use this skill as the only entry router for a complete competition-paper task. Standard targets strong models that can maintain a long evidence chain and execute tools reliably while keeping cost controlled. It does not use Pro multi-agent tournaments or approval checkpoints.

Start Or Resume

From the contest project root, run:

python .claude/skills/paper-workflow-orchestrator/scripts/preflight_check.py
python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --status

Read paper_output/qa/workflow_guard_report.json and follow recommended_skill plus next_action. The current files and hashes override conversational memory.

Do not run downstream skills before their guard requirement passes. After a child skill finishes, return here and evaluate status again.

S0-S8

S0 Input Admission

preflight_check.py inventories problem_files/, hashes every input, checks runtime dependencies, prepares paper_output/, and rejects mixed MathModel editions. Required outputs:

• paper_output/preflight_report.json
• paper_output/input_manifest.json
• paper_output/OUTPUT_LAYOUT.md

S1 Problem Analysis

Use $problem-doc-model-selector to create paper_output/step1/problem_analysis.json. Every question, attachment, field, objective, constraint, ambiguity, and required output must be traceable.

S2 Model And Rubric Route

Use $modeling-paper-rubric-and-model-selector. Produce:

• paper_output/plan/model_route.json
• paper_output/plan/rubric_alignment.json
• paper_output/plan/scoring_strategy.md

Use $authoritative-data-harvester only when public external data is necessary. Keep source identity and retrieval notes.

S3 Data And Visualization Plan

Use $data-cleaning-and-visualization. Read only files classified in the input manifest and produce a fresh load report, data plan, visualization plan, figure index, and cleaned data. Contest-specific code belongs under paper_output/code/, never inside installed skills.

S4 Reproducible Model Code

Use $model-code-and-result-generator to write question-specific code under paper_output/code/modeling/, including run_modeling.py and per-question modules. Code must emit machine-readable result contracts.

S5 Real Execution

Run the modeling code. Preserve script, input, output, exit-code, size, and SHA-256 records in paper_output/results/run_manifest.json. Required evidence includes model results, finite metrics, conclusions, tables, and usable figures. Draft placeholders do not count.

S6 Evidence Gate

Use $quality-assurance-auditor and run official evidence validation:

python .claude/skills/quality-assurance-auditor/scripts/evidence_gate.py --mode official

Do not enter formal writing until paper_output/qa/evidence_gate_report.json is PASS and all recorded inputs are still fresh.

S7 Adaptive Formal Writing

Use $paper-formal-writer as the sole formal author:

python .claude/skills/paper-formal-writer/scripts/build_paper_outline.py
python .claude/skills/paper-formal-writer/scripts/prepare_authoring.py --mode auto

Normal competition papers use complete-section drafting, possibly over several turns. Preserve the formal writer's declared competition scope; a short report needs an explicit user-requested scope and reason. Audit every draft with validate_authoring.py --section; global repeated failure falls back to section mode. A section’s second repeated category creates a micro-repair route; only then may $paper-micro-unit-generator repair the queued location. The third repeated category blocks S7 and suggests Lite without switching automatically.

After every active unit passes:

python .claude/skills/paper-formal-writer/scripts/assemble_sections.py
python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --assembled

The Agent must then globally revise the full assembly into paper_output/final_paper_source.md; a copy-only promotion is rejected. Finish with:

python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --final
python .claude/skills/paper-formal-writer/scripts/format_formal_docx.py

S8 Format And Render Gate

Run:

python .claude/skills/paper-formal-writer/scripts/check_paper_format.py --render required

Delivery requires a fresh PASS in paper_output/format_check_report.json. Fix the reported source, formula, citation, figure/table, DOCX, pagination, or PDF issue and rerun; never edit the report to force PASS.

Formal Outputs

paper_output/plan/writing_plan.json
paper_output/context/authoring_state.json
paper_output/qa/draft_audit.json
paper_output/qa/repair_queue.json
paper_output/drafts/sections/*.md
paper_output/drafts/assembled_draft.md
paper_output/final_paper_source.md
paper_output/final_paper.docx
paper_output/format_check_report.json

Legacy micro-unit and quickstart outputs remain under paper_output/drafts/legacy/ and paper_output/quickstart/. They can never satisfy S7.

Non-Negotiable Invariants

• Never invent model results, data sources, successful runs, citations, or validation.
• Every critical numeric claim must trace to current machine-readable evidence.
• Any upstream hash change invalidates dependent S6-S8 reports.
• Each included figure/table must exist, be indexed, cited, and interpreted.
• Formal formulas must become editable Word OMML; no screenshot or plain-text substitution.
• Do not expose skill names, guard commands, or workflow prose in the paper body.
• Do not write contest-specific code into installed skill directories.
• Install one MathModel edition per contest project.

Recovery

For an interrupted or long task:

python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --status
python .claude/skills/context-memory-keeper/scripts/update_workflow_memory.py

Read the guard report, current stage contracts, and paper_output/context/workflow_memory.json; continue from the first failing stage instead of replaying completed work.