AppSprout-dev · jkbennitt · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026
diff --git a/.github/ISSUE_TEMPLATE/bug_report.yml b/.github/ISSUE_TEMPLATE/bug_report.yml
@@ -11,7 +11,7 @@ body:
     id: version
     attributes:
       label: Unsga3 version / commit
-      placeholder: "0.1.2 or git SHA"
+      placeholder: "0.1.4 or git SHA"
     validations:
       required: true
   - type: textarea

diff --git a/.grok/skills/unsga3-oracle/SKILL.md b/.grok/skills/unsga3-oracle/SKILL.md
@@ -16,8 +16,11 @@ Repo root: Unsga3. Confirm `Unsga3.slnx` / `tools/OracleCompare` exist before ru
 | Problem | Partitions | Pop | Gens | C# tournament |
 |---------|------------|-----|------|----------------|
 | zdt1 | 12 | 52 | 100 | default (`RankNicheDistance`) |
+| zdt2 | 12 | 52 | **250** | `--pymoo-mode` (`PymooCompatible`) |
 | dtlz2 | 12 | 92 | 150 | `--pymoo-mode` (`PymooCompatible`) |
 
+ZDT2 **gens=100** is an early-stress snapshot (collapse on Bend, C#, and pymoo), not the quality bar. Quality protocol matches unsga3-bend A/B (gens=250, PymooCompatible). `RankNicheDistance` is an optional unpublished Wilcoxon ZDT2 mating mode — do not silently switch all ZDT defaults to it. ZDT1 and DTLZ2 unchanged.
+
 IGD = **mean** nearest Euclidean distance (pymoo-compatible). Docs: `docs/EQUIVALENCE.md`, `docs/RESEARCH-STANDARDS.md`.
 
 Requires: .NET 10 SDK; Python 3 + `pip install pymoo` for pymoo side / multi-seed.
@@ -31,8 +34,10 @@ Ask which mode if unclear; default **quick** when validating a small fix, **full
 ```powershell
 dotnet build tools/OracleCompare -c Release
 dotnet run --project tools/OracleCompare -c Release --no-build -- --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1
+dotnet run --project tools/OracleCompare -c Release --no-build -- --problem zdt2 --partitions 12 --pop 52 --seed 1 --pymoo-mode
 dotnet run --project tools/OracleCompare -c Release --no-build -- --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1 --pymoo-mode
 python tools/oracle/run_pymoo_oracle.py --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1
+python tools/oracle/run_pymoo_oracle.py --problem zdt2 --partitions 12 --pop 52 --seed 1
 python tools/oracle/run_pymoo_oracle.py --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1
 ```
 

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -15,6 +15,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 
 - Docs and XML comments describe `initialPopulation` and hybrid loops in generic terms (domain-adapter warm-start / grid-seed). No product-repo names.
 
+## [0.1.4] — 2026-09-19
+
+### Changed
+
+- **ZDT2 oracle/smoke protocol:** quality bar is **gens=250** + `PymooCompatible` (docs and harness defaults). gens=100 is an early-stress snapshot (collapse on Bend, C#, and pymoo), not the quality bar. `RankNicheDistance` stays an optional unpublished Wilcoxon ZDT2 mating mode. ZDT1 (100) and DTLZ2 (150) unchanged. **No algorithm or public API change.** Matches [unsga3-bend](https://github.com/AppSprout-dev/unsga3-bend) A/B honesty.
+
 ## [0.1.3] — 2026-08-10
 
 ### Added
@@ -60,7 +66,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 - Benchmarks: ZDT1–4/6, DTLZ1–4/7, Sphere / Ackley / Rosenbrock
 - Metrics: IGD, GD, 2-D HV; self-tests + GitHub Packages publish workflow
 
-[Unreleased]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.2...HEAD
+[Unreleased]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.4...HEAD
+[0.1.4]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.3...v0.1.4
+[0.1.3]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.2...v0.1.3
 [0.1.2]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.1...v0.1.2
 [0.1.1]: https://github.com/AppSprout-dev/Unsga3/compare/v0.1.0...v0.1.1
 [0.1.0]: https://github.com/AppSprout-dev/Unsga3/releases/tag/v0.1.0
diff --git a/CITATION.cff b/CITATION.cff
@@ -21,8 +21,8 @@ keywords:
   - C#
   - .NET
 license: MIT
-version: 0.1.2
-date-released: "2026-08-10"
+version: 0.1.4
+date-released: "2026-09-19"
 references:
   - type: article
     authors:

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -28,6 +28,9 @@ dotnet run --project samples/BasicUsage -c Release
 # single seed
 dotnet run --project tools/OracleCompare -c Release -- --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1 --pymoo-mode
 python tools/oracle/run_pymoo_oracle.py --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1
+# ZDT2 quality protocol: omitted --gens is 250 + --pymoo-mode (matches unsga3-bend).
+# --gens 100 is an early-stress snapshot. RankNicheDistance (omit --pymoo-mode) is optional.
+dotnet run --project tools/OracleCompare -c Release -- --problem zdt2 --partitions 12 --pop 52 --seed 1 --pymoo-mode
 
 # multi-seed Mann–Whitney / Wilcoxon (writes docs/WILCOXON-RESULTS.md)
 python tools/oracle/run_multiseed_wilcoxon.py --problems zdt1 dtlz2 --seeds 15

diff --git a/Directory.Build.props b/Directory.Build.props
@@ -12,7 +12,7 @@
     <PackageProjectUrl>https://github.com/AppSprout-dev/Unsga3</PackageProjectUrl>
     <RepositoryUrl>https://github.com/AppSprout-dev/Unsga3</RepositoryUrl>
     <PackageTags>optimization;nsga;nsga3;unsga3;multi-objective;evolutionary;moea</PackageTags>
-    <Version>0.1.3</Version>
+    <Version>0.1.4</Version>
     <PackageReadmeFile>README.md</PackageReadmeFile>
 
   </PropertyGroup>

diff --git a/README.md b/README.md
@@ -6,7 +6,7 @@
 
 **U-NSGA-III** (Unified NSGA-III) for .NET — single-, multi-, and many-objective evolutionary optimization with Das–Dennis reference directions, SBX crossover, polynomial mutation, and **niching-based tournament selection** ([Seada & Deb, 2016](https://ieeexplore.ieee.org/document/7271063)).
 
-> **v0.1.2** — production-usable core with pymoo-aligned normalization.  
+> **v0.1.4** — production-usable core with pymoo-aligned normalization. ZDT2 quality protocol is gens=250 (docs/defaults; no algorithm/API change).  
 > **15-seed IGD vs pymoo `UNSGA3`:** ZDT1 **median 0.053 vs 0.070** (we win; MWU *p*≈0.05); DTLZ2 **median 0.0045 vs 0.0028** (~1.6×, same order; pymoo still ahead).  
 > Details: [`docs/WILCOXON-RESULTS.md`](docs/WILCOXON-RESULTS.md) · single-seed notes: [`docs/ORACLE-RESULTS.md`](docs/ORACLE-RESULTS.md)
 
@@ -121,7 +121,7 @@ API docs used: [HTTP](https://docs.typesafe.ai/api.md) · [Python SDK](https://d
 |-----|----------|
 | [docs/ORACLE-RESULTS.md](docs/ORACLE-RESULTS.md) | Single-seed C# vs pymoo |
 | [docs/WILCOXON-RESULTS.md](docs/WILCOXON-RESULTS.md) | Multi-seed Mann–Whitney / Wilcoxon |
-| [docs/EQUIVALENCE.md](docs/EQUIVALENCE.md) | Protocol & intentional deltas |
+| [docs/EQUIVALENCE.md](docs/EQUIVALENCE.md) | Protocol & intentional deltas (ZDT2 quality A/B = gens=250, `PymooCompatible`; matches unsga3-bend) |
 | [docs/RESEARCH-STANDARDS.md](docs/RESEARCH-STANDARDS.md) | Literature + indicator standards |
 | [docs/NOTICE.md](docs/NOTICE.md) | Attribution (papers + validation tools) |
 | [docs/ROADMAP.md](docs/ROADMAP.md) | Near / medium term plan |

diff --git a/docs/EQUIVALENCE.md b/docs/EQUIVALENCE.md
@@ -22,6 +22,16 @@ See also **[RESEARCH-STANDARDS.md](RESEARCH-STANDARDS.md)** for the literature +
 5. **Tolerance:** median IGD within ~1–2× of pymoo on ZDT/DTLZ is the practical bar.
    15-seed: ZDT1 median **better** than pymoo (ratio 0.76, MWU n.s.); DTLZ2 median ~**1.6×** (pymoo still ahead).
 
+Published A/B budgets (ZDT1 / DTLZ2 unchanged; ZDT2 matches unsga3-bend protocol honesty):
+
+| Problem | Pop | Gens | Partitions | C# tournament (quality) |
+|---------|-----|------|------------|-------------------------|
+| ZDT1 | 52 | 100 | 12 | `RankNicheDistance` (published Wilcoxon) |
+| ZDT2 | 52 | **250** | 12 | `PymooCompatible` (A/B quality) |
+| DTLZ2 | 92 | 150 | 12 | `PymooCompatible` |
+
+ZDT2 **gens=100** is an early-stress snapshot (collapse on Bend, C#, and pymoo), not the quality bar. `RankNicheDistance` on ZDT2 is an optional unpublished Wilcoxon mating mode — do not silently switch all ZDT defaults to it. C# has no published ZDT2 IGD / Wilcoxon table; do not invent one.
+
 ## Problems (must-pass)
 
 | Class | Problems | M | In library |

diff --git a/docs/ORACLE-RESULTS.md b/docs/ORACLE-RESULTS.md
@@ -20,13 +20,21 @@ Reproduce:
 # Python
 pip install pymoo
 python tools/oracle/run_pymoo_oracle.py --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1
+python tools/oracle/run_pymoo_oracle.py --problem zdt2 --partitions 12 --pop 52 --seed 1
 python tools/oracle/run_pymoo_oracle.py --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1
 
 # C#
 dotnet run --project tools/OracleCompare -c Release -- --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1
+dotnet run --project tools/OracleCompare -c Release -- --problem zdt2 --partitions 12 --pop 52 --seed 1 --pymoo-mode
 dotnet run --project tools/OracleCompare -c Release -- --problem dtlz2 --partitions 12 --pop 92 --gens 150 --seed 1 --pymoo-mode
 ```
 
+Omitted `--gens` on `--problem zdt2` is **250** (quality protocol). `--gens 100` is an early-stress snapshot. ZDT1 stays 100; DTLZ2 stays 150.
+
+### ZDT2 protocol honesty (matches unsga3-bend)
+
+C# never published a hard ZDT2 oracle / Wilcoxon table. The unpublished Wilcoxon harness used **gens=100** and **`RankNicheDistance`**. That budget collapses on Bend, C#, and pymoo (axis pile near `f1≈0`). **gens=250 + `PymooCompatible`** is the quality A/B bar (15-seed: 0/15 collapse on both stacks; see [unsga3-bend PR #17](https://github.com/AppSprout-dev/unsga3-bend/pull/17) / `docs/ZDT2_COLLAPSE.md`). `RankNicheDistance` remains an **optional** Wilcoxon mating mode — do not silently switch all ZDT defaults to it. This repo does not invent a ZDT2 IGD table here.
+
 ## Results (seed=1)
 
 | Problem | Settings | pymoo IGD | C# default IGD | C# `PymooCompatible` IGD | Verdict |
@@ -69,9 +77,12 @@ Deep-dive vs pymoo `HyperplaneNormalization` / `ReferenceDirectionSurvival` (pym
 | Test | Bar |
 |------|-----|
 | ZDT1 seed=1, 100 gen, default tournament | IGD ≤ 1.5 × 0.0629 |
+| ZDT2 seed=2, 250 gen, default `RankNicheDistance` | IGD &lt; 0.75 (loose CI smoke, not oracle parity) |
 | DTLZ2 seed=1, 150 gen, pymoo-mode | IGD ≤ 3 × 0.00350 (currently ~1.15×) |
 | DTLZ2 short smoke (80 gen) | IGD &lt; 0.15 |
 
+ZDT2 quality A/B is gens=250 + `PymooCompatible` (not the loose smoke bar). ZDT1 / DTLZ2 shipping bars are unchanged.
+
 ## Known remaining deltas (intentional / minor)
 
 | Item | Status |

diff --git a/docs/RESEARCH-STANDARDS.md b/docs/RESEARCH-STANDARDS.md
@@ -77,9 +77,12 @@ Typical ZDT: r = (1.1, 1.1). Always document r; never compare HV across differen
 1. Same problem definition (bounds, n, evaluate)  
 2. Same Das–Dennis partitions → identical ref set size  
 3. Same pop, gens, SBX/PM η, p_m = 1/n  
+   - ZDT1: pop=52, **gens=100**  
+   - ZDT2 quality A/B: pop=52, **gens=250**, `PymooCompatible` (matches unsga3-bend). gens=100 is an early-stress snapshot, not the quality bar. `RankNicheDistance` is optional, not the ZDT2 default.  
+   - DTLZ2: pop=92, **gens=150**, `PymooCompatible`  
 4. Fixed seed **or** 15–31 seeds → median + IQR IGD  
 5. Compare IGD (and HV for M=2) to pymoo `UNSGA3`  
-6. Shipping bar: median IGD within ~1–2% of pymoo on ZDT1/DTLZ2 (or non-inferior Wilcoxon)
+6. Shipping bar: median IGD within ~1–2% of pymoo on ZDT1/DTLZ2 (or non-inferior Wilcoxon). ZDT2 has no published C# Wilcoxon table; quality budget is 250 gens.
 
 Export path: dump final `F` as CSV from both sides; compute IGD in this library.
 

diff --git a/docs/WILCOXON-RESULTS.md b/docs/WILCOXON-RESULTS.md
@@ -7,8 +7,11 @@ Generated by `tools/oracle/run_multiseed_wilcoxon.py`. IGD = mean nearest Euclid
 | Problem | Pop | Gens | Partitions | C# tournament | Seeds |
 |---------|-----|------|------------|---------------|-------|
 | zdt1 | 52 | 100 | 12 | RankNicheDistance | 15 |
+| zdt2 (quality A/B; not in this table) | 52 | **250** | 12 | PymooCompatible | — |
 | dtlz2 | 92 | 150 | 12 | PymooCompatible | 15 |
 
+ZDT2 **gens=100** is an early-stress snapshot, not the quality bar. C# never published a hard ZDT2 Wilcoxon table; the unpublished harness used gens=100 + `RankNicheDistance` (optional mating mode — do not silently switch all ZDT defaults to it). Quality protocol matches [unsga3-bend](https://github.com/AppSprout-dev/unsga3-bend) A/B honesty: gens=250 + `PymooCompatible`. ZDT1 and DTLZ2 numbers below are unchanged. Do not invent a ZDT2 IGD table here.
+
 Hypothesis tests (α = 0.05, two-sided):
 
 - **Mann–Whitney U** (Wilcoxon rank-sum): independent samples, H₀: same IGD distribution.

diff --git a/tests/Unsga3.Tests/Benchmarks/IgdSmokeTests.cs b/tests/Unsga3.Tests/Benchmarks/IgdSmokeTests.cs
@@ -9,6 +9,9 @@ namespace Unsga3.Tests.Benchmarks;
 /// <summary>
 /// Fixed-seed IGD smoke bars — not full pymoo equivalence (see docs/EQUIVALENCE.md).
 /// Bounds are intentionally loose; tighten after oracle harness lands.
+/// ZDT2 quality A/B is gens=250 + <see cref="TournamentMode.PymooCompatible"/>
+/// (matches unsga3-bend). The ZDT2 smoke below keeps ctor-default
+/// <see cref="TournamentMode.RankNicheDistance"/> and a loose IGD bar.
 /// </summary>
 public class IgdSmokeTests
 {
@@ -40,10 +43,13 @@ public void Zdt2_runs_with_measurable_igd()
         var problem = new Zdt2Problem();
         var dirs = ReferenceDirections.DasDennis(2, 12);
         var algo = new Unsga3Algorithm(dirs, populationSize: 52, seed: 2);
-        var result = algo.Run(problem, maxGenerations: 150);
+        var result = algo.Run(problem, maxGenerations: 250);
         var obtained = result.NonDominatedSolutions.Select(i => i.Objectives).ToArray();
         double igd = PerformanceIndicators.InvertedGenerationalDistance(obtained, ParetoFronts.Zdt2());
-        // ZDT2 non-convex; mean-IGD smoke bar (not oracle parity).
+        // Loose CI smoke, not oracle parity. Quality A/B is gens=250 + PymooCompatible
+        // (matches unsga3-bend). gens=100 is an early-stress snapshot (collapse on
+        // Bend, C#, and pymoo). This test keeps the ctor default RankNicheDistance
+        // (optional unpublished Wilcoxon ZDT2 mating mode) and IGD < 0.75.
         Assert.True(igd < 0.75, $"ZDT2 IGD={igd}");
     }
 

diff --git a/tools/OracleCompare/Program.cs b/tools/OracleCompare/Program.cs
@@ -10,11 +10,16 @@
 // Fixed-protocol C# side of the pymoo oracle (see tools/oracle/run_pymoo_oracle.py).
 //
 //   dotnet run --project tools/OracleCompare -- --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1 --pymoo-mode
+//   # ZDT2 quality protocol (matches unsga3-bend A/B): gens=250 + --pymoo-mode.
+//   # Omitted --gens on --problem zdt2 is 250. gens=100 is an early-stress snapshot.
+//   # RankNicheDistance (omit --pymoo-mode) is optional, not the ZDT2 A/B default.
+//   dotnet run --project tools/OracleCompare -- --problem zdt2 --partitions 12 --pop 52 --seed 1 --pymoo-mode
 
 string problemName = "zdt1";
 int partitions = 12;
 int? pop = null;
 int gens = 100;
+bool gensExplicit = false;
 int seed = 1;
 bool pymooMode = false;
 string? outDir = null;
@@ -26,13 +31,21 @@
         case "--problem": problemName = args[++i]; break;
         case "--partitions": partitions = int.Parse(args[++i], CultureInfo.InvariantCulture); break;
         case "--pop": pop = int.Parse(args[++i], CultureInfo.InvariantCulture); break;
-        case "--gens": gens = int.Parse(args[++i], CultureInfo.InvariantCulture); break;
+        case "--gens":
+            gens = int.Parse(args[++i], CultureInfo.InvariantCulture);
+            gensExplicit = true;
+            break;
         case "--seed": seed = int.Parse(args[++i], CultureInfo.InvariantCulture); break;
         case "--pymoo-mode": pymooMode = true; break;
         case "--out-dir": outDir = args[++i]; break;
     }
 }
 
+// Quality protocol: ZDT2 A/B default is gens=250 (unsga3-bend honesty).
+// ZDT1 stays 100; DTLZ2 callers still pass --gens 150. Explicit --gens always wins.
+if (!gensExplicit && problemName.Equals("zdt2", StringComparison.OrdinalIgnoreCase))
+    gens = 250;
+
 IProblem problem;
 int m;
 double[][] pf;

diff --git a/tools/oracle/run_multiseed_wilcoxon.py b/tools/oracle/run_multiseed_wilcoxon.py
@@ -5,6 +5,11 @@
 Protocol matches docs/EQUIVALENCE.md and docs/RESEARCH-STANDARDS.md:
   15 independent seeds (default), fixed SBX/PM/Das–Dennis settings.
 
+ZDT2 quality protocol is gens=250 + PymooCompatible (matches unsga3-bend A/B).
+gens=100 is an early-stress snapshot. RankNicheDistance (csharp_pymoo_mode=False)
+is an optional unpublished Wilcoxon mating mode — not the ZDT2 A/B default.
+ZDT1 and DTLZ2 budgets are unchanged.
+
 Usage:
   python tools/oracle/run_multiseed_wilcoxon.py
   python tools/oracle/run_multiseed_wilcoxon.py --problems zdt1 dtlz2 --seeds 15 --skip-pymoo
@@ -38,7 +43,10 @@ class Protocol:
 
 PROTOCOLS: dict[str, Protocol] = {
     "zdt1": Protocol("zdt1", partitions=12, pop=52, gens=100, n_obj=2, csharp_pymoo_mode=False),
-    "zdt2": Protocol("zdt2", partitions=12, pop=52, gens=100, n_obj=2, csharp_pymoo_mode=False),
+    # Quality A/B (unsga3-bend honesty): 250 gens, PymooCompatible.
+    # Optional: csharp_pymoo_mode=False (RankNicheDistance) + gens=100 is the
+    # unpublished Wilcoxon ZDT2 mating snapshot — do not treat it as the default.
+    "zdt2": Protocol("zdt2", partitions=12, pop=52, gens=250, n_obj=2, csharp_pymoo_mode=True),
     "dtlz2": Protocol("dtlz2", partitions=12, pop=92, gens=150, n_obj=3, csharp_pymoo_mode=True),
 }
 

diff --git a/tools/oracle/run_pymoo_oracle.py b/tools/oracle/run_pymoo_oracle.py
@@ -9,6 +9,9 @@
 Usage:
   python run_pymoo_oracle.py
   python run_pymoo_oracle.py --problem zdt1 --partitions 12 --pop 52 --gens 100 --seed 1
+  python run_pymoo_oracle.py --problem zdt2 --partitions 12 --pop 52 --seed 1
+  # omitted --gens on zdt2 is 250 (quality protocol; matches unsga3-bend A/B).
+  # --gens 100 is an early-stress snapshot, not the quality bar.
 """
 from __future__ import annotations
 
@@ -20,15 +23,33 @@
 import numpy as np
 
 
+def default_gens(problem: str) -> int:
+    """Quality-protocol generations when --gens is omitted.
+
+    ZDT2 A/B default is 250 (unsga3-bend honesty). ZDT1 stays 100.
+    DTLZ2 stays 100 here so existing callers that omit --gens are unchanged;
+    the published DTLZ2 oracle still passes --gens 150.
+    """
+    if problem == "zdt2":
+        return 250
+    return 100
+
+
 def main() -> int:
     p = argparse.ArgumentParser()
     p.add_argument("--problem", default="zdt1", choices=["zdt1", "zdt2", "dtlz2"])
     p.add_argument("--partitions", type=int, default=12)
     p.add_argument("--pop", type=int, default=None, help="default = n_ref_dirs")
-    p.add_argument("--gens", type=int, default=100)
+    p.add_argument(
+        "--gens",
+        type=int,
+        default=None,
+        help="generations (default: zdt2=250, else 100; explicit value always wins)",
+    )
     p.add_argument("--seed", type=int, default=1)
     p.add_argument("--out-dir", type=Path, default=Path(__file__).resolve().parent / "out")
     args = p.parse_args()
+    gens = args.gens if args.gens is not None else default_gens(args.problem)
 
     try:
         from pymoo.algorithms.moo.unsga3 import UNSGA3
@@ -56,12 +77,12 @@ def main() -> int:
     algo = UNSGA3(ref_dirs, pop_size=pop)
 
     print(f"pymoo UNSGA3 | problem={args.problem} M={n_obj} refs={len(ref_dirs)} "
-          f"pop={pop} gens={args.gens} seed={args.seed}")
+          f"pop={pop} gens={gens} seed={args.seed}")
 
     res = minimize(
         problem,
         algo,
-        ("n_gen", args.gens),
+        ("n_gen", gens),
         seed=args.seed,
         verbose=False,
         save_history=False,
@@ -71,7 +92,7 @@ def main() -> int:
     igd = float(IGD(pf)(F))
 
     args.out_dir.mkdir(parents=True, exist_ok=True)
-    stem = f"pymoo_{args.problem}_p{args.partitions}_pop{pop}_g{args.gens}_s{args.seed}"
+    stem = f"pymoo_{args.problem}_p{args.partitions}_pop{pop}_g{gens}_s{args.seed}"
     f_path = args.out_dir / f"{stem}_F.csv"
     meta_path = args.out_dir / f"{stem}_meta.json"
     np.savetxt(f_path, F, delimiter=",")
@@ -83,7 +104,7 @@ def main() -> int:
         "partitions": args.partitions,
         "n_ref_dirs": int(len(ref_dirs)),
         "pop_size": pop,
-        "n_gen": args.gens,
+        "n_gen": gens,
         "seed": args.seed,
         "n_solutions": int(F.shape[0]),
         "igd": igd,