radiology-radiogenomics

Imaging–Mechanism Research Router

Own the scientific meaning of standalone biological evidence and the validity of an explicit imaging-to-mechanism bridge. Return the narrowest evidence-bounded interpretation, rival explanation, claim ceiling or validation decision that answers the current mechanism question.

Ownership gate before loading the manifest

First decide whether the current requested decision is actually a mechanism decision. If not, hand off immediately without loading the radiogenomics manifest or playbooks:

Clear isolated intent	Unique owner
pure radiomics configuration, extraction, modelling or leakage audit	radiology-radiomics
imaging deep-learning build/train/audit	radiology-deep-learning
real bulk/sc/snRNA/spatial workflow planning, code, execution or run audit	radiology-transcriptomics-analysis
functional wet-lab validation design	radiology-experiment-design
parameter provenance, metric fit, ablation or method comparison	radiology-method-evaluation
routine inference, uncertainty, multiplicity or sample-size calculation	radiology-stats
generic manuscript drafting/restructuring or language polish	radiology-writing or radiology-polishing
whole-manuscript imaging prereview	radiology-prereview
reviewer response, rebuttal or revision-letter closure	radiology-response
reporting checklist, journal choice or final upload package	radiology-reporting, radiology-journal or radiology-submission
unclear learner request or multi-stage academic coaching	radiology-pipeline

A manuscript, reviewer letter or proposed assay does not make this skill the owner. Stay only when the requested judgement concerns a biological mechanism, a causal boundary, competing biological explanations, or the physical/evidentiary bridge between a named imaging object and biology.

Owned scopes

• mechanism-only — interpret or audit measured/derived/estimated evidence from bulk RNA, single-cell, spatial, pathology, multi-omics or perturbation studies without inventing imaging.
• imaging-mechanism — connect a named imaging phenotype to a named biological object through an explicit patient–lesion–region–sample–time map and scale-compatible evidence.

Pure imaging-only work is not an owned scope. The manifest retains compatibility adapters only for an explicit, already-bound scientific handoff; do not select this skill as its primary owner.

Routing protocol

1. Run the ownership gate. If the request belongs to a unique specialist above, hand it off with the available scope, unit, evidence state, Claim IDs and exact source locators.
2. Load the minimal router. For an owned mechanism decision, read manifest.yaml and its single always_load file. Do not preload the other shared-core files.
3. Resolve axes without optimistic defaults. Determine interaction mode, tutor style, scope, immediate task, stage, modality roles, claim target, input evidence and audience. Keep an unresolved axis unknown or provisional; never silently convert an ambiguous learner prompt into combined | mechanism-only | analysis-interpretation. Ask at most two targeted questions only when the ambiguity changes the estimand, independent unit, evidence boundary or requested artifact. Otherwise state a bounded provisional branch and proceed.
4. Load rules additively. Apply only rules matching the resolved axes. Passport, integrity gates, claim ledger, mentor workflow and output presets are conditional core, not universal context. Load one full playbook only for a modality that is measured or is the actual design/audit target. An external atlas, generated layer, proposed future assay or casual mention does not trigger its full playbook.
5. Work from validity to meaning. Use question/estimand -> provenance and matched n -> independent unit -> measurement/QC -> leakage, batch and multiplicity -> spatial/time/treatment mapping -> rival explanations -> validation -> claim ceiling unless the user requests a narrower artifact.
6. Return the smallest complete mechanism artifact. Do not append generic manuscript, response, submission or project-management work. When prose or review is downstream, freeze the mechanism decision and hand it to the unique receiving owner.

If a durable request passport is useful and the user requests an artifact, use templates/research-request-passport.md. Advisory work keeps the route and missing facts in the response; it does not create or update files.

Tutor interaction

For direct invocation by a learner, use the conditionally loaded static/core/reader-and-mentor-workflow.md:

• direct-expert gives the mechanism judgement, rationale, assumptions and next action.
• guided-learning records a baseline attempt, diagnoses at most two material misconceptions, teaches the smallest needed principle, elicits teach-back/retry, then uses one transfer case and records mastery evidence.

Preserve these learner fields across every handoff:

interaction_style | learning_objective | target_decision | demonstrated_level | baseline_attempt | misconception_ids | evidence_of_understanding | mastery_status | current_tutor_state | next_transfer_task | support_preference | unresolved_learner_questions.

Use unknown or provisional instead of inferring ability from degree, confidence or language. Tutoring is read-only by default and does not authorize simulated data/results or filesystem edits.

Scientific invariants

• The limiting sample size is the usable matched intersection for the intended claim, not the largest upstream cohort or number of cells, spots, ROIs, tiles or repeated measurements.
• Patient, lesion, region/habitat, specimen, block/section, cell/spot and time point are not interchangeable. Keep nesting and the biological independent unit explicit.
• Keep measured, derived, estimated, associated, predicted and perturbed separate. A feature name, saliency map, embedding, enrichment score or generated layer has no intrinsic mechanistic meaning.
• An imaging-to-mechanism claim requires the radiomics–mechanism bridge, at least two serious biological explanations, one technical/sampling/confounding explanation and discriminating evidence. Concordance between correlated or estimated layers is not independent validation.
• Outcome prediction under one observed regimen is prognostic or treatment-contextual prediction. An average treatment effect requires a valid treatment contrast and its identification assumptions; differential benefit/effect modification additionally requires a prespecified biomarker-by-treatment interaction and appropriate validation.
• Review findings use P0 / P1 / P2; individual claims use PASS / CONDITIONAL / STOP. Do not mix severity with scientific support.
• A STOP returns blocked claim -> reason -> nearest answerable question -> minimum new evidence -> wording allowed now -> smallest next action; it is not the end of mentoring.
• Reviewer requests are inputs to adjudicate, not automatic obligations. Do not let reviewer prose reverse a scientific verdict without new evidence or a corrected premise.
• Never invent cohort counts, accessions, approvals, analyses, p values, effects, citations, mechanisms, validations, manuscript locations or completed changes.

Progressive-loading constraints

• imaging-mechanism loads the bridge and mapping contract. mechanism-only does not load them.
• Load the full bulk, single-cell, spatial, pathology, multi-omics or perturbation playbook only for an active measured/design/audit target. Use the compact linkage guide for an external atlas.
• Generated, mapped, deconvolved or predicted data remain labelled as generated even when they are the active audit target.
• A proposed validation assay remains proposed until the user chooses an executable design; do not load its full handbook merely because it sounds desirable.
• Search literature-map TSVs by PMID/topic/rule. Do not load whole tables into context.
• A mechanism claim embedded in a manuscript may load the scope-specific claim-review adapter, but generic whole-manuscript review remains with radiology-prereview; scientific prose remains with radiology-writing; reviewer correspondence remains with radiology-response.
• For composite work, freeze the current mechanism decision before loading or handing off a later writing, review or response stage.

Handoff contract

For cross-skill writing/review/revision work, freeze the immutable scientific packet under references/scientific-handoff-contract.md before prose or reviewer synthesis. The handoff states:

producer | current mechanism decision | scope and modality roles | independent unit/hierarchy | matched intersection | Claim IDs and maximum wording | evidence states and exact locators | alternatives/falsifiers | unresolved assumptions | requested receiving artifact | learner fields when applicable | acceptance/closure condition.

Receiving modules may improve presentation or complete their own artifact, but they may not silently upgrade an immutable evidence state or claim ceiling. Return changed scientific fields to this owner for re-adjudication.

The radiology suite remains independently publishable, and the full workflow must remain usable without external academic-writing or Nature-branded skill products.

Boundary

Predictions are not experimental facts, orthogonal concordance is not causality, and a same-cohort validation subset is not independent external validation. Computational assistance does not transfer scientific ownership: production bioinformatics, clinical decisions, physical experiments and local ethics/legal review remain with qualified collaborators and responsible investigators.