radiology-grant

Radiology Grant Review and Development

Use this skill for funding proposals, not for manuscript peer review or presentation production. Treat a grant as a call-bound argument for a consequential, answerable question and a feasible research programme. A polished narrative cannot cure an ineligible application, an unsupported premise, an underpowered design, or an uncontrolled imaging measurement chain.

Non-negotiable boundaries

• Lock the live call before compliance advice. Record the funder, mechanism, cycle, programme, official sources and retrieval date in a grant call passport. If those cannot be verified, emit CALL_VERSION_UNRESOLVED, restrict the review to scientific content, and do not claim submission readiness.
• For NSFC, lock six routing fields. Record application year, programme/type and subtype, science division/office, year-bound application code candidate, research attribute and any special call/clinical track. Read nsfc-2024-2026-change-map.md; current official-source conflicts become CONFLICTING / ADMIN_CONDITIONAL, never an inferred eligibility answer.
• Separate two gates. Administrative acceptance (eligibility, limits, forms, attachments, budget regime, ethics/security) is not scientific merit. Passing either gate does not imply passing the other.
• Use the target funder's current criteria. Do not translate every call into a generic NSFC, NIH or numeric rubric. Do not invent scores, weights, deadlines, page limits or eligibility.
• Evidence before rhetoric. Preliminary findings must point to real source artifacts. Keep verified facts, user attestations, calculations, hypotheses, planned work and missing evidence visibly distinct. Never turn planned work into preliminary data.
• Protect claim ceilings. Prediction is not measurement, mechanism, causality or clinical utility. A model-performance aim cannot silently license any of those stronger claims.
• Review is read-only by default. Critique and propose changes; modify the submitted file only when the user explicitly requests revision.
• NSFC generative-AI boundary. For a live NSFC application, read call-and-compliance-gate.md. The 2026 annual application and completion notice permits bounded research-tracking/reference assistance only with human verification, applicant responsibility, truthful declaration and marking; it prohibits directly generated applications and unverified content. Label AI-assisted text as advisory for applicant verification and rewriting; never represent it as applicant-authored or submission-ready.

Route by task

User need	Mode and resources	Main output
Live NSFC rule, year comparison or submission-readiness question	read nsfc-2024-2026-change-map.md, call-and-compliance-gate.md and source-registry.md	six-field current-year lock, exact-source rule map and conflict states
NSFC title, abstract, 立项依据, 研究内容, 研究基础 or innovation development	read nsfc-application-content-playbook.md and grant-architecture.md	scientific-function map and author decision points, not legacy headings
NSFC imaging code, H18/H27/H28/H29 or clinical/special-call choice	read nsfc-radiology-code-and-call-router.md	year-bound candidate-code/call record and unresolved confirmation gate
Fast fatal-flaw or one-section check	triage/targeted; read review-modes-and-criteria.md	bounded findings, missing inputs and next gate
Whole application review	full audit; also read call-and-compliance-gate.md, radiology-feasibility-audit.md, and evidence-contract-and-stopping-rules.md	criterion-by-criterion report
Simulated panel or pre-submission challenge	mock panel; read review-modes-and-criteria.md	independent views, disagreement record, panel synthesis
Response to prior critiques / resubmission	revision/resubmission; read review-modes-and-criteria.md	critique-to-change matrix and residual risks
Final funder-system fields, files and institutional handoff	funder-package-finalization; read call-and-compliance-gate.md and use the submission contract in grant-call-passport.md	closed field/file/attachment inventory, unresolved blockers and human institutional action gate
Develop question, aims or architecture	read grant-architecture.md and reframe-and-innovation.md	proposal spine and author decision points
Feasibility, milestones, budget or contingency	read radiology-feasibility-audit.md and feasibility-and-pitfalls.md	aim/milestone/risk register
Choose a framework for imaging AI, diagnostic accuracy, prediction, radiomics, generative AI or quantitative imaging	read imaging-methodology-router.md	task-specific design-function map; no checklist-as-validity claim
NIH, ERC, Wellcome or another international route	read international-grants.md, then verify the named call live	funder-specific mapping; never generic substitution

Do not load all references for a narrow request.

Funder-package finalization boundary

Use funder-package-finalization only for one exact funder + mechanism + cycle + programme/subcall + applicant/host configuration. Recheck the live call, amendments, current electronic form/system and institutional route. Freeze:

call-passport ID/version | proposal/source versions | field/section/attachment inventory | exact file locators and hashes where files exist | page/character/format checks | budget/ethics/security/AI-use handoffs | applicant/co-applicant attestations | institutional steps and deadlines | unresolved blockers | responsible human role.

This mode audits a package; it does not log in, paste fields, upload files, attest for an applicant, confirm institutional eligibility, authorize a budget or press a submission control. Scientific or budget content that changes during finalization returns to the grant/scientific owner and creates a new source version before the inventory is rechecked.

Use only these machine states:

• FUNDER_PACKAGE_INCOMPLETE — required fields, files, current rules or source versions are missing;
• FUNDER_PACKAGE_BLOCKED — a known compliance, authority, confidentiality or unresolved-live-rule condition prevents institutional handoff; or
• HUMAN_INSTITUTIONAL_SUBMISSION_REQUIRED — the bounded machine audit is complete and every remaining portal, attestation, approval and submission action belongs to authorized humans.

HUMAN_INSTITUTIONAL_SUBMISSION_REQUIRED is the maximum machine state. Never report SUBMITTED, ACCEPTED, ELIGIBLE_CONFIRMED, official receipt, funding probability or funder/institutional approval without the corresponding authoritative external evidence.

Review workflow

1. Intake and version lock. Identify the review mode and material version. Build the call passport. Record absent proposal sections instead of inferring them.
2. Administrative gate. Check eligibility, application limits, required forms/attachments, research category/code, budget route, ethics/security and AI-use rules only against current official sources. For NSFC, freeze all six routing fields before any compliance verdict. Preserve official-text conflicts instead of resolving them by convenience. Return ADMIN_PASS, ADMIN_CONDITIONAL, ADMIN_FAIL, or ADMIN_NOT_ASSESSABLE with evidence.
3. Freeze the proposal spine. Extract clinical/scientific need → gap → central question → hypothesis → aims → expected inference/impact. Flag aims that do not serve the same question.
4. Apply the funder criteria. Create one criterion finding per material issue: criterion, verdict, evidence anchor, consequence, actionable repair and residual uncertainty. Preserve the funder's language and scoring model.
5. Run the radiology gate. Audit the clinical-phenomenon → image-phenotype → biological/clinical proposition bridge; intended use and unit; acquisition/reconstruction; multimodal intersection denominator; reference standard; annotation; data/model-lineage leakage; aim-level sample size; external-independence vector; statistics; transportability; mechanism evidence; clinical pathway; data/compute/ethics; and measurement-to-claim consistency.
6. Test execution. Map every aim to personnel, data, dependencies, milestone, decision rule, fallback, time and budget. Use the aim/milestone/risk register.
7. Synthesize without laundering disagreement. Prioritize fatal compliance defects and inference-breaking weaknesses before competitive or editorial improvements. In panel mode, retain minority concerns.
8. Close with a decision boundary. State what was assessed, what remains unverified, which claims are supportable, and whether the next state is READY_FOR_AUTHOR_REVISION, READY_FOR_INSTITUTIONAL_CHECK, REVIEWABLE_BUT_NOT_READY, or STOP_AND_REFRAME.

Output contract

For a full review use grant-review-report.md and include:

1. Review scope — mode, proposal version, sections actually read, target call and limitations.
2. Call passport status — current official evidence and unresolved live rules.
3. Administrative gate — separate from merit, with no implied funding judgment.
4. Proposal spine — question, hypothesis, aims, claim ceiling and logic breaks.
5. Criterion findings — PASS / CONDITIONAL / FAIL / NOT_ASSESSABLE, each evidence-anchored.
6. Radiology feasibility gate — measurement, data, reference standard, analysis and translation.
7. NSFC code/call and methodology route, when applicable — year-bound candidate code, clinical or special-call boundary, selected design frameworks and explicit non-applicability.
8. Aim–milestone–risk–budget alignment — go/no-go logic and scientifically meaningful fallback.
9. Prioritized revision plan — P0 compliance, P1 inference-threatening, P2 competitive, P3 editorial.
10. Evidence and uncertainty ledger — verified, attested, planned, inferred, missing and stale items.
11. Readiness boundary — never a guarantee of eligibility, score, review outcome or funding.

Handoffs

• Study design and validation → radiology-design; acquisition/reconstruction/QC → radiology-acquisition-qc; ground truth → radiology-annotation; data → radiology-data.
• Sample size and inference → radiology-stats / radiology-method-evaluation; clinical reader-study design → radiology-translation (stats supplies MRMC analysis and sizing); ethics and governance → radiology-ethics; clinical pathway → radiology-clinical-domain.
• Live call, deadline, eligibility or frontier verification → radiology-search / radiology-frontier; project-level provenance → radiology-pipeline.
• A frozen, author-approved proposal spine for a defence or pitch deck → radiology-paper2ppt. This skill does not design slides.
• Award acceptance, post-award conditions, amendments, annual/interim/final technical reporting and sponsor deliverable tracking → radiology-research-ops; changed scientific content returns to its original scientific owner before the funder-facing artifact is refreshed.
• Final portal use, applicant attestations, institutional approvals and submission remain with the applicant, host research office and funder. A completed package audit stops at HUMAN_INSTITUTIONAL_SUBMISSION_REQUIRED.
• This skill does not replace the funder, the host institution's research office, ethics review, financial approval or human expert judgment, and never guarantees funding.