Research Defense Radar

Research Defense Radar / 研究防御雷达

Build a decision-oriented map of literature that can pre-empt, narrow, support, falsify, or improve a research project. Treat the project—not a keyword list—as the unit of analysis.

Reply in the user's language. Preserve authoritative paper titles in their original language. If you provide a translated title, label it as an assistant translation rather than an official title. For a Chinese operating summary, read SKILL.zh-CN.md; this file remains authoritative if the two differ.

Select one operating mode

Choose the smallest mode that answers the request:

1. Known-paper comparison — compare one or more supplied papers with the project. Do not run a full baseline unless the user asks or the comparison exposes a material gap.
2. Baseline competitor map — establish the first research fingerprint, search broadly, and produce a provisional novelty assessment.
3. Project-revision refresh — diff a revised project against its prior fingerprint, re-score known papers, and search only newly relevant threat surfaces before broadening.
4. Incremental monitor — use durable prior state and report only new or materially changed literature.
5. Method/data discovery — prioritize transferable measurement, data, identification, estimation, validation, and implementation practices; substantive-topic overlap is optional.

If a vague input leaves multiple materially different interpretations, ask only the questions that would change the search. Otherwise proceed with a partial fingerprint and label assumptions.

When the user supplies a full working-paper draft, distinguish access to the manuscript from public disclosure. If no stable public record establishes when the user's project or a comparison WP first became publicly accessible, ask for that earliest public date. Do not infer it from PDF metadata, a file-modification timestamp, a date printed inside a private draft, or private circulation.

1. Resolve inputs and state

Accept research questions, proposals, LaTeX, PDF, Word, slides, code, specifications, data dictionaries, referee responses, and chat notes. Read enough of the actual material to recover framing, data, methods, and claimed contributions.

Separate each fingerprint statement by provenance:

• user-stated — explicitly supplied by the user;
• source-extracted — directly supported by the project files;
• agent-inferred — a provisional interpretation that needs confirmation.

For local/project-backed work, copy assets/RESEARCH_PROFILE_TEMPLATE.md and assets/RADAR_STATE_TEMPLATE.json into a user-controlled project directory such as:

research-radar/<project-slug>/
├── fingerprint.md
├── radar-state.json
├── observations.json
├── coverage.json
├── search-log.jsonl
└── alerts/

Never write project data into this installed skill directory. Skill upgrades may replace it, and private research material does not belong in a distributable package.

For web/cloud runs without a durable folder, use one of these patterns:

• schedule inside the ongoing research chat so prior context remains available;
• attach the fingerprint and state to an accessible project/task;
• store them in an explicitly authorized connected service; or
• embed a compact fingerprint plus known-paper registry in the scheduled prompt when the state is small.

If prior state is unavailable, run a fresh baseline and say so. Never call an item "new since last run" without readable prior state.

2. Build or update the research fingerprint

Use assets/RESEARCH_PROFILE_TEMPLATE.md. Capture:

• core research question and main claim;
• mechanisms and falsifiable predictions;
• population, setting, unit, geography, and period;
• data sources, granularity, linkage, and access constraints;
• key constructs, measures, and validation risks;
• identification and identifying assumptions;
• estimation, theory, simulation, ML/LLM, or data-engineering methods;
• primary and secondary contributions;
• known nearest papers and the user's current novelty claim;
• earliest public working-paper date, its stable record or user-stated provenance, and whether the project is not yet public;
• threat surfaces: combinations that would materially reduce novelty;
• search vocabulary and privacy-safe public query terms.

The default rubric is strongest for empirical economics, finance, management, and adjacent quantitative social science. For theory, experiments, computer science, life sciences, qualitative research, or humanities, read references/DOMAIN_ADAPTATION.md and mark inapplicable fields NA instead of forcing an empirical template.

3. Search current literature

Before any novelty judgment, read references/SEARCH_PLAYBOOK.md and execute the relevant query families. Search the contribution space separately by question, mechanism, data/setting, measurement, identification, method, outcome, known-paper citation neighborhood, authors, and dangerous combinations.

Use multiple independent source families appropriate to the field. Treat working papers, preprints, job-market papers, conference drafts, accepted/forthcoming work, and online-first publications as potentially priority-relevant.

For each search run, record:

• search date and literature cutoff date;
• query family and privacy-safe query;
• source or index checked;
• coverage status: completed, partial, inaccessible, or not applicable;
• material access limitations and likely blind spots.

Do not paste confidential proposal passages, proprietary dataset names, or unpublished hypotheses into public search. Generalize them to the minimum conceptual query needed. If generalization would destroy the search value, show the proposed public query to the user before sending it.

4. Verify and deduplicate

Do not infer a paper's contribution from its title. Assign an evidence level:

• M — Metadata only: identity/status only; insufficient for substantive overlap claims.
• A — Abstract verified: supports a provisional description and overlap assessment.
• F — Full text or detailed manuscript verified: supports method, identification, mechanism, and contribution comparisons.

Every priority item needs a stable URL or DOI, evidence level, verification date, earliest public date and its provenance when available, current version/status, and confidence level. If the WP content is available but its public date is not, ask the user rather than treating the manuscript date as public chronology.

Track a paper lineage rather than counting NBER, SSRN, arXiv, conference, author-page, and journal versions as separate papers. Prefer DOI, then repository ID, then canonical URL, then normalized title plus author overlap. Note changed claims, samples, data, or methods across versions.

If local state is available, read references/STATE_SCHEMA.md, copy the observation and coverage templates, and use scripts/update_radar_state.py to merge verified observations deterministically. Read its --help output before use and prefer --dry-run before an in-place update.

5. Apply the public-working-paper priority cutoff

If the user's own WP is already public and its earliest public date is verified or user-stated, use that date as the priority cutoff:

• literature public before the project WP can pre-empt or narrow its public-time novelty and requires differentiation analysis;
• literature public after the project WP is post-disclosure convergence: it may matter for citation, monitoring, or method/data learning, but it does not negate the project's novelty at disclosure and does not, by default, require a new distinction;
• same-day or unknown timing remains uncertain until better chronology evidence is available.

Keep the public-time novelty verdict separate from the current literature landscape. If the project is not yet public, or its public date is unknown, do not apply this protection.

6. Classify relevance

Choose one primary class and optional secondary tags:

• A — Direct competitor: occupies substantially the same contribution space.
• B — Potential threat: pre-empts or narrows a novelty, mechanism, data, measurement, or design claim.
• C — Method/data lift: offers a transferable research step.
• D — Supporting/positioning: strengthens motivation, theory, interpretation, or external validity.
• E — Contradictory/falsification evidence: challenges a maintained assumption, expected sign, or mechanism.

Class A is not a mechanical keyword threshold. A paper may be a direct competitor with different data or methods if it makes substantially the same primary contribution. Conversely, shared keywords do not make a paper a threat.