# Pareto Desk

> Paste the evaluated designs from an optimization run or a design sweep and get the feasible Pareto front computed in the browser with pymoo semantics, then a trade-off decision you can defend or a pymoo script that refines the search. https://pareto-desk.skillsafe.ai/

## What it does

- Free, in the browser, nothing uploaded: reads CSV, TSV, semicolon- or whitespace-separated tables and raw pymoo dumps (the front is searched over up to 100,000 rows); assigns column roles from header tags - name[min], name[max], name[<=b], name[>=b], name[g] (already g <= 0), name[var], name[label] - or pymoo names F#, G#, X#, or a roles line; a headerless numeric table is read as objectives to minimise.
- Computes with pymoo semantics: maximised objectives negated; constraints in g <= 0 form, violation = sum of positive parts; fast non-dominated sorting of the feasible designs; crowding distance per front; ideal and nadir of the first front; exact hypervolume of the front normalised by them (reference point 1.1 by default, or yours in original units, which also gives a run-to-run comparable raw hypervolume); pymoo PseudoWeights per design and the pick nearest your weights; the ASF pick (ASF on normalised values with 1/weights).
- Checked against pymoo 0.6.2 on 600 random problems (2-5 objectives, ties, duplicates, mixed scales): identical fronts, front order, crowding distances, pseudo-weights and picks; hypervolume equal to moocore's within 1e-13 relative.
- Flags: no or few feasible designs, mostly infeasible sample, one dominant design, thin front, weak dominance, many objectives (NSGA-III advice), objectives that do not conflict (Spearman > 0.9), degenerate objectives, duplicates on the front, gaps in a 2-objective front, constraints active on the front, disagreeing picks, scale disparity, skipped rows, no decision variables.
- Scatter plot of any two objectives; front, ranked-rows and Markdown exports whose headers carry the role tags, so an export pasted back reproduces the analysis; the pick copied as one CSV row with its decision variables.
- The loop: after a refinement run, paste the new designs and the page says whether the previous pick is now dominated, and can combine the earlier designs with the new ones into one front.

## Paid lanes (model gpt-terra, signed-in users)

- task "tradeoff": verdict (recommend, close_call, rerun), a pick, a shortlist of 1-5 front designs with what each gives and gives up, up to 6 measured moves between designs, constraint notes and caveats.
- task "refine": an algorithm (NSGA2, NSGA3, RNSGA2, SMSEMOA, MOEAD), a search space with bounds from the front or table ranges, run settings, a complete pymoo script with an evaluate_design(x) stub that writes pareto_refined.csv back in this page's tag format, assumptions and checks. Paste the new results back to decide again; the page compares with the last run on the same problem.
- Every reply is reconciled in the browser: flags answered once, pick a sent front design, verdict no looser than the browser's read, every number in the prose present in the browser's facts or your notes, and for scripts n_obj / n_ieq_constr / n_var, the allowed imports, the column tags and the bounds.

## Limits

- A front says which designs are not beaten in the objectives given - not that a design is safe or right to build.
- The model never recomputes dominance, crowding or hypervolume and never evaluates a design; a refinement script needs your own evaluation function.
- Up to 100,000 rows read; above 5,000 usable rows the front, hypervolume and picks still cover every row, while later fronts, the plot and the ranked export use the front plus the last rows (the page, the export names and the run all say so). A table cut at 100,000 rows is marked PARTIAL; a run is sent up to 40 front designs (picks, extremes, then the most isolated).

## Pages

- App: https://pareto-desk.skillsafe.ai/
- API tutorial: https://pareto-desk.skillsafe.ai/api.html

## Source

Derived from the agent skill @k-dense-ai/pymoo (https://skillsafe.ai/skill/@k-dense-ai/pymoo), part of k-dense-ai/scientific-agent-skills by K-Dense Inc. (https://github.com/k-dense-ai/scientific-agent-skills). pymoo: J. Blank and K. Deb, "pymoo: Multi-Objective Optimization in Python", IEEE Access 8 (2020) 89497-89509. The example data are illustrative, not measured.