Which of your evaluated designs should you actually pick?

Paste the designs from an optimization run or a design sweep. Your browser finds the feasible Pareto front the way pymoo does - dominance, crowding, hypervolume, preference picks - and plots it, free, nothing uploaded. A paid run then writes the trade-off decision or a pymoo script to refine the search.

Each example has a saved model run, so you can see the whole page for free.

Problem

Evaluated designs (CSV, TSV or a pymoo dump; one row per design; the front is searched over up to 100,000 rows)

Drop a .csv / .tsv / .txt or

Tag each header: [min], [max], [<=250], [>=10], [g] (already g <= 0), [var]. Columns named F1, G1, X1 are read the pymoo way; a headerless table is read as objectives to minimise.

Roles, weights and reference point

Roles line (overrides tags)

Preference weights

Hypervolume reference point

Weights drive the two picks: pymoo's pseudo-weights (the design whose position on the front best matches them) and its achievement scalarisation (ASF). Give a reference point in original units, one value per objective, to get a hypervolume you can compare between runs.

Notes (optional: what the objectives mean, what matters most, hard limits, how the designs were made)

Load set .json

Pareto front (free, in your browser, no AI)

Across Up

Front designs in the accent colour, dominated feasible designs grey, infeasible designs as red rings; the two picks are ringed.

Pareto front

Flags

Free: no account, no run.

Anything to ask (optional)

Paste your designs to price the run.

Raw model reply

The reply did not parse as the structured result, so it is shown as it arrived.

Your recent runs

What this does, and what it does not

The front is computed the way pymoo computes it: maximised objectives are negated so everything is minimised, constraints are put in g <= 0 form and summed into a violation, and fast non-dominated sorting ranks the feasible designs. Crowding distance, the ideal and nadir points, the hypervolume of the front normalised by them, pymoo's pseudo-weights and its achievement scalarisation (ASF) pick follow the pymoo source. The implementation was checked against pymoo 0.6.2 on 600 random problems - two to five objectives, ties, duplicates, mixed scales - with identical fronts, crowding distances and picks, and hypervolumes equal to 13 significant figures.

The paid run reads only what the browser computed and your notes. It is told never to compute a new number, and the page checks every number it writes. A refinement script is a starting point: you supply the function that evaluates a design. Derived from the agent skill @k-dense-ai/pymoo (k-dense-ai/scientific-agent-skills, K-Dense Inc.). The example data are illustrative, not measured.