Metric Justification

Optional inputs:

• terminology_registry (optional): Canonical terms, symbols, and disallowed variants for consistency control.
• current_draft_context (optional): Existing local draft text or outline to revise, compress, or audit.
• venue_constraints (optional): Page limit, formatting rules, length pressure, or author preference on tone.

Output Contract

• Output format: paragraph
• Required sections: objective_interpretation, main_output, risk_flags, handoff_notes
• Tone and language: technical / English
• Length guidance: Keep outputs compact, handoff-ready, and easy to embed in a prompt library.

Workflow

1. Treat the provided inputs as the only source of truth.
2. Produce only the bounded artifact for this skill, not a whole-paper rewrite.
3. Preserve contribution names, terminology, and claim boundaries from the shared inputs.
4. Make gaps explicit instead of guessing, especially when evidence is incomplete.
5. Before finalizing, run the section checks and self-checklist below.
6. If you need the full original schema or prompt template, read references/spec.yaml.

Guardrails

Core hard rules:

• Do not invent experiments, results, contributions, baselines, theoretical conclusions, or citations.
• Maintain factual faithfulness, claim-evidence alignment, terminology consistency, overclaim suppression, and reviewer-friendly clarity.
• If evidence is missing, weak, or indirect, say so explicitly and reduce claim strength rather than filling gaps.
• Do not report numerical trends, significance, or robustness claims unless they are explicitly supported by supplied evidence.

Section-specific checks:

• Output stays within the scope of Experiments and does not leak unsupported content from other sections.
• Contribution names, task setting, and method references match the global blueprint.
• Claims are phrased at the strength justified by the supplied evidence.
• Each research question or claim is paired with the exact supporting experiment or ablation.
• Uncertainty, caveats, and limited causal interpretation are surfaced clearly.

Revision policy:

• Missing required input: Return a gap-marked scaffold and list exactly which missing inputs block a stronger output.
• Unsupported claim: Downgrade or remove the claim, explain the evidence gap, and never patch the gap with invented detail.
• Multiple valid options: Provide 2-3 bounded options with tradeoffs, but keep all options faithful to the same verified facts.

Coordination

Reasoning focus:

• logical coherence.
• evidence grounding.
• novelty positioning.
• wording restraint.
• comparison discipline.

Self-checklist:

• Every main statement is supported by provided facts or explicitly labeled as a gap.
• No wording overstates novelty, empirical coverage, causality, or theory scope.
• Terminology, symbols, and contribution labels match the shared registry and blueprint.
• The output does not conflict with neighboring sections such as abstract, method, experiments, or conclusion.
• Every reported trend, comparison, or efficiency statement is traceable to a supplied table, figure, or trusted summary.

Upstream skills:

• 00_01_paper_blueprint_skill
• 00_03_contribution_extractor_skill
• 00_04_claim_evidence_alignment_skill
• 00_06_terminology_notation_registry_skill
• 05_11_method_draft_skill
• 07_04_baseline_selection_skill

Downstream handoff:

• 07_06_implementation_details_skill