Jinko Calibration Subsampling

Jinkō Subsampling SDK Workflows

Subsampling creates a derived, smaller Vpop by selecting patients from the Vpop simulated in a completed Trial so that the selected population best matches specified population-level targets. It neither calibrates the model nor creates new patients. Use jinko-trial to create, sanity-check, and run the source Trial, and jinko-vpop to inspect the generated Vpop. Scientific choices belong to a workflow or domain expert, not this skill.

PREREQUISITE: This skill needs an initialized jinko-sdk connection and an SDK satisfying its metadata.requires_sdk range. Run the jinko-sdk-setup skill (../jinko-sdk-setup/SKILL.md) and proceed only once its check passes. If that skill is not found, install it from novainsilico/jinko-skills.

Canonical Flow

1. Retrieve the Trial, require trial.status()["status"] == "completed", then inspect trial.descriptors.scalars and trial.descriptors.categoricals; descriptor IDs and arms must be taken from this Trial, not guessed from display labels.
2. Build a SubsamplingDesign with filters and population targets through trial.create_subsampling_design(...).
3. Read design.diagnostics; do not generate while it has errors. Use design.diagnostics.errors().explain() to relate an error to its target or filter and its source-Trial descriptor.
4. Call design.generate_vpop(...) with all simulated-annealing options. The returned Vpop is immutable.
5. Inspect generated artifacts through design.generated_vpops.list_with_details(). Reuse a compatible design with design.set_trial(other_trial) before it is used, or edit its typed components such as design.marginals.

Scalar Discovery and Candidate Estimates

Read references/scalar-discovery-and-estimates.md before choosing a Trial scalar or using platform-fitted law estimates. It distinguishes descriptor discovery, per-patient values, and candidate target forms without making the scientific choice for the user.

For a complete Python flow and the meaning of generation options, read references/generation-and-diagnostics.md.

Typed Targets and Edits

• Numeric filters: descriptor builders such as scalar.gte(18), or design.numeric_filters.create_gte(...) after creation.
• Scalar targets: scalar.normal(...), .uniform(...), .weibull(...), and design.marginals.create_* / persisted-handle setters.
• Other supported SDK target surfaces: design.categorical_filters, .categoricals, .correlations, .survivals, .summary_statistics, and .observables. Read references/creating-and-editing.md before using one.
• The older UI guide says categorical constraints are unsupported, whereas the current SDK exposes typed categorical builders and services. Treat support as backend/version-dependent: create the design and require clean diagnostics before generation.

Use design.edit(...) only for advanced full-slice replacement. Prefer typed subservices so immutable IDs and existing content are preserved. A design can be pointed at another Trial only when descriptor/arm pairs remain compatible; use design.set_trial(...) and validate diagnostics again.

Project Folder Hygiene

Propose a YYYY-MM-DD-<experiment> folder and reuse an exact-name match via client.get_folder_by_name(name, exact_match_only=True). Create folders and remote project items only after confirmation or when a script receives --apply.

SDK Scripts

These are on PATH as console scripts once the SDK is installed, and also runnable via python -m as shown below.

• jinko.cli.create_subsampling_design: dry-run creation of numeric filters, normal scalar marginals, and observables; --apply creates the design.
• jinko.cli.generate_subsampled_vpop: dry-run generation plan; --apply checks diagnostics and creates the Vpop.
• jinko.cli.inspect_subsampling_design: prints design content, diagnostics, source Trial, and generated-Vpop options/fitness without mutating anything.

Read references/scripts.md for invocation examples.

Reference Routing

• references/creating-and-editing.md: descriptors, builders, target types, typed edits, and compatible Trial reuse.
• references/scalar-discovery-and-estimates.md: output-scalar discovery, per-patient scalar values, and platform candidate law estimates.
• references/generation-and-diagnostics.md: validation, annealing options, and generated-Vpop semantics.
• references/inspection.md: artifact listing, stored options, and fitness payload caveats.
• references/scripts.md: bundled-script invocation examples.