Jinko Calibration Cmaes

Jinkō CMA-ES Calibration SDK Workflows

The calibration manager API is CMA-ES only — no type/method discriminator exists. "Subsampling" is an unrelated VPop-generator feature, not a calibration type. This skill is pure SDK mechanics: no defaults, no diagnostics, no when-to-calibrate guidance.

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.

Minimum Calibration

• parameters: priors to calibrate (required, ≥1).
• At least one fitness-function source (required): dataTableDesigns (data table must report metadata.public.validForFitnessFunction: True, see jinko-data-table) and/or an advanced output set with objectives (see jinko-output-set). This skill creates neither input.
• CalibrationOptions: seed + thresholdWeightedScore are schema-required and have contract defaults 0 and 1; populationSize + numberOfIterations have no contract defaults and are functionally required. The bundled creation script requires population size and iteration count, and uses the contract defaults for omitted seed and threshold; pass all four explicitly when reproducibility policy requires it.

Two encoding rules are mandatory before creation:

• With log_transform=True, mean and std are in log10(x) coordinates, but min_bound and max_bound remain in the original physical coordinates of x. If planned bounds are written in log10 coordinates, exponentiate them first: physical_bound = 10**log10_bound. A calibration sanity warning such as MAX_BOUND_LOWER_THAN_MEAN_LOG indicates this mapping is inconsistent.
• For every attached fitness data table, set options.log_transform_wide_bounds to every distinct obsId in that table unless the user explicitly requests linear bound scaling for a named observable. This is the SDK field behind the UI's Scale bounds option.

Create

model = client.get_model("cm-...")
data_table = client.get_data_table("dt-...")
calibration = model.create_calibration(
    parameters=[
        {
            "id": "k_elim",
            "mean": -1.0,
            "std": 0.5,
            "log_transform": True,
            "min_bound": 0.001,
            "max_bound": 10.0,
        }
    ],
    data_tables=[
        {
            "data_table": data_table,
            "include": True,
            "options": {
                "weight": 1.0,
                "log_transform_wide_bounds": sorted({
                    row["obsId"] for row in data_table.export()
                }),
            },
        }
    ],
    calib_seed=42,
    calib_threshold_weighted_score=0.0,
    calib_number_of_iterations=100,
    calib_population_size=12,
)

Equivalent client-level call: client.create_calibration(model=model, ...). calibrationOptionsOverride, solvingOptionsOverride, coreVersion have no typed kwarg — use client.create_calibration_from_json(json_content=payload) / client.calibrations.create_raw(payload). See references/creating-a-calibration.md for full field tables.

Solving times can be set post-creation with calibration.set_solving_times(t_max=timedelta(days=28), t_step="P1D"); each duration may be a timedelta or ISO 8601 string.

Run & Poll

calibration.run()
final_status = calibration.wait_until_completed(timeout=3600)

See references/running-and-polling.md for .get_sanity(), .status(), and StoppingReason values.

Results

calibration.performance()  # raw dict
calibration.results_summary()  # raw dict
calibration.objective_weights()  # raw dict, {objective_id: weight}
calibration.results.sorted_patients(
    sort_by="optimizationWeightedScore desc"
)  # raw, low-level

All results accessors return unparsed dicts today. See references/results-and-inspection.md.

Project Folder Hygiene

Same as jinko-trial/jinko-data-table: propose a YYYY-MM-DD-<experiment> folder, reuse an exact-name match via client.get_folder_by_name(name, exact_match_only=True), create only on confirmation or --create-folder --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_cmaes_calibration: dry-run by default, creates a calibration with --apply.
• jinko.cli.run_calibration: runs and polls an existing calibration with --apply.
• jinko.cli.inspect_calibration: prints/writes raw performance/results_summary/objective_weights/sorted_patients JSON.

python -m jinko.cli.create_cmaes_calibration --model-sid cm-... --data-table-sid dt-... --parameter "k_elim:-1.0:0.5:0.001:10.0:log" --seed 42 --threshold-weighted-score 0.0 --iterations 100 --population-size 12
python -m jinko.cli.create_cmaes_calibration --model-sid cm-... --data-table-sid dt-... --parameter "k_elim:-1.0:0.5:0.001:10.0:log" --seed 42 --threshold-weighted-score 0.0 --iterations 100 --population-size 12 --folder 2026-07-07-calib --create-folder --apply
python -m jinko.cli.run_calibration --calibration-sid ca-... --apply --timeout 3600
python -m jinko.cli.inspect_calibration --calibration-sid ca-... --performance --results-summary --objective-weights --output-dir calib-results

Reference Routing

• references/creating-a-calibration.md: full field tables, three creation patterns.
• references/running-and-polling.md: run/stop/status/sanity, JobStatus, StoppingReason.
• references/results-and-inspection.md: performance/results_summary/objective_weights/results.* field tables and caveats.