jkbennitt · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -10,17 +10,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 ### Added
 
 - Optional **TypeSafe / Jev System One** pass (`tools/typesafe-pareto`) for Score + Choice judgments over a small Pareto candidate sample. Additive semantic layer — does not replace NSGA-III / U-NSGA-III objectives. Live calls only when `TYPESAFE_API_KEY` is set; CI uses mocks. Metrics append to `metrics/typesafe-runs.jsonl`.
+- `ReferenceDirections.TwoLayerDasDennis`: outer Das–Dennis layer plus an inside layer scaled toward the centroid (Deb & Jain, NSGA-III Part I; default scale 0.5). Duplicates are removed. Single-layer `DasDennis` / `WithDasDennis` stays the default. M≤3 oracles do not need the second layer.
+- Constrained problems `OsyProblem`, `TnkProblem`, and `C1Dtlz1Problem`, with formulation and survival self-tests. No IGD table.
 - Opt-in `MatingPoolMode.TwoShuffledPasses`: two shuffled passes of consecutive tournaments (Seada & Deb Algorithm 1 and pymoo 0.6.2 `TournamentSelection`, pressure 2). When N parents are drawn from N members, each member is a contestant twice. The default pool stays N independent tournaments with replacement, which is the published ZDT1 Wilcoxon protocol.
 - Opt-in `TournamentMode.Algorithm2`: a same-niche perpendicular-distance tie keeps the second parent. `PymooCompatible` still coin-flips. `RankNicheDistance` is unchanged.
 
-### Fixed
-
-- Odd population sizes no longer skip `parents[N-1]` in SBX. The leftover parent is paired with parent 0. Even-N pair order is unchanged (ZDT1 pop 52, DTLZ2 pop 92).
-
 ### Changed
 
+- Environmental selection follows pymoo `Survival.filter_infeasible` (Jain & Deb, NSGA-III Part II). Feasible members are niched. If fewer than N are feasible, the rest are filled by ascending constraint violation and are not assigned a niche. When any member is feasible, ideal, worst, and ASF extremes come from feasible objectives. An all-infeasible generation does not move the hyperplane. This replaces the locked whole-pool ideal fixture.
 - Docs and XML comments describe `initialPopulation` and hybrid loops in generic terms (domain-adapter warm-start / grid-seed). No product-repo names.
 
+### Fixed
+
+- `CompareConstraintDominated` no longer falls through to Pareto when two infeasible individuals share a constraint violation. Deb's constraint-domination leaves them mutually non-dominated. Tournament comparators are unchanged.
+- Odd population sizes no longer skip `parents[N-1]` in SBX. The leftover parent is paired with parent 0. Even-N pair order is unchanged (ZDT1 pop 52, DTLZ2 pop 92).
+
 ## [0.1.5] — 2026-09-22
 
 Forensic-audit wave (G1–G10): document known gaps against pymoo and Seada & Deb, and lock them with tests. The published 15-seed table is not rewritten. **Default tournament stays `RankNicheDistance`.**

diff --git a/README.md b/README.md
@@ -73,13 +73,14 @@ var algo = new Unsga3Algorithm(dirs, populationSize: 92, seed: 1,
 | `Unsga3Algorithm` | Main entry — `Run(problem, gens)` |
 | `Individual` | Variables / Objectives / Constraints |
 | `OptimizationResult` | Final population + non-dominated set |
-| `ReferenceDirections.DasDennis` | Structured reference points |
+| `ReferenceDirections.DasDennis` | Single-layer reference points (default) |
+| `ReferenceDirections.TwoLayerDasDennis` | Outer layer plus an inside layer for larger M |
 | `SimulatedBinaryCrossover` / `PolynomialMutation` | Variation operators |
 | `PerformanceIndicators` | IGD, GD, 2-D hypervolume |
 | `TournamentMode` | Default rank→niche, `PymooCompatible` (distance-tie coin flip), or opt-in `Algorithm2` (keep the second parent) |
 | `MatingPoolMode` | Default independent tournaments; opt-in `TwoShuffledPasses` (paper / pymoo 0.6.2) |
 
-Built-in problems: ZDT1–4/6, DTLZ1–4/7, Sphere, Ackley, Rosenbrock.
+Built-in problems: ZDT1–4/6, DTLZ1–4/7, Sphere, Ackley, Rosenbrock, and constrained OSY, TNK, and C1-DTLZ1 (self-tests only; no IGD table yet).
 
 ## Build & test
 
@@ -154,6 +155,7 @@ Algorithm papers (please cite these when publishing results):
 
 - Seada, H. & Deb, K. (2016). *A Unified Evolutionary Optimization Procedure for Single, Multiple, and Many Objectives.* IEEE Trans. Evol. Comput.
 - Deb, K. & Jain, H. (2014). *An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach (NSGA-III), Part I.* IEEE Trans. Evol. Comput.
+- Jain, H. & Deb, K. (2014). *NSGA-III, Part II: Handling Constraints and Extending to an Adaptive Approach.* IEEE Trans. Evol. Comput.
 - Das, I. & Dennis, J. E. (1998). *Normal-Boundary Intersection.* SIAM J. Optim.
 
 ## License

diff --git a/docs/EQUIVALENCE.md b/docs/EQUIVALENCE.md
@@ -11,6 +11,7 @@ See also **[RESEARCH-STANDARDS.md](RESEARCH-STANDARDS.md)** for the literature +
 - COIN Report 2014022; Seada & Deb, IEEE TEVC 2016  
 - [pymoo `UNSGA3`](https://pymoo.org/algorithms/moo/unsga3.html)  
 - pymoo `nsga3.py` — `HyperplaneNormalization`, `associate_to_niches`, `niching`  
+- pymoo `Survival.filter_infeasible` — feasible niching, then ascending-CV fill (Jain & Deb, NSGA-III Part II)  
 - Indicators: [pymoo performance indicators](https://pymoo.org/misc/indicators.html) (GD, IGD, IGD+, HV)
 
 ## Protocol
@@ -44,12 +45,13 @@ ZDT2 **gens=100** is an early-stress snapshot (collapse on Bend, C#, and pymoo),
 | Bi | ZDT1–4, ZDT6 | 2 | yes |
 | Many | DTLZ1–4, DTLZ7 | 3+ | yes |
 | Hard | WFG1, WFG2, WFG9 | 3–5 | planned |
-| Constrained | OSY, TNK, C1-DTLZ1 | 2–3 | planned |
+| Constrained | OSY, TNK, C1-DTLZ1 | 2–3 | yes (self-tests; no IGD table) |
 
 ## Unit checks
 
 - Reference association & niche counts  
-- Non-dominated ranks / constraint domination  
+- Non-dominated ranks / constraint domination (equal CV is mutual non-domination)  
+- Feasible niching, CV fill, and OSY / TNK / C1-DTLZ1 (`ConstrainedSurvivalTests`, `ConstrainedProblemTests`). No IGD table.  
 - **Normalization intercepts / ASF axis extremes** (`NormalizationTests`)  
 - Tournament pressure  
 - `populationSize: null` ⇒ `|refs|`  
@@ -71,14 +73,17 @@ ZDT2 **gens=100** is an early-stress snapshot (collapse on Bend, C#, and pymoo),
 | Mating pool (`TwoShuffledPasses`, opt-in) | two shuffled passes of consecutive pairs. When N parents are drawn from N members, each index is a contestant twice | pymoo 0.6.2 `TournamentSelection` (pressure 2) **is** this pool: `n_perms = ceil(2N / N)` permutations, then consecutive pairs |
 | Niche ids at mating | `PrepareForSelection` re-normalizes survivors and re-associates | ids written during survival are kept |
 | `WithDasDennis(1, 1)` | throws. One objective has a single direction, and N must be ≥ 2 | pass `populationSize` ≥ 2 for the single-objective degeneration |
-| Reference layers | single-layer Das–Dennis only. Two-layer directions are absent | many-objective NSGA-III adds an inside layer for larger M |
+| Reference layers | `DasDennis` / `WithDasDennis` stay single-layer (the M≤3 oracles). `TwoLayerDasDennis` adds Part I's inside layer, default scale 0.5, duplicates removed | multi-layer factory; NSGA-III consumes whatever `ref_dirs` it is given |
 | Duplicate elimination | default **on**. Key is `G12` (12 significant digits), not 12 decimal places. Attempts are capped when mutation cannot produce a new key; remaining slots may be duplicates | `eliminate_duplicates=True` |
 | Survival RNG | optional RNG niche pick | random among equal niches |
 | IGD | **mean** nearest distance | same (verified pymoo 0.6.2) |
 | Scored set | full non-dominated front | `res.F` niche optimum |
 | Hyperplane norm | persistent ideal, ND extremes, correct ASF | `HyperplaneNormalization` |
 | Collapsed nadir | if the span is still ≤ 1e-6, nadir = ideal + 1 | stop at worst-of-population |
-| Infeasible points in the hyperplane | ideal and worst from the whole pool, including infeasible points. ASF extremes use the ND index set when supplied | pymoo niching can restrict the normalized set to feasible members |
+| Constrained survival | Niche feasible members only. If fewer than N are feasible, fill by ascending CV (original index breaks ties) with no niche assignment. All-infeasible generations truncate by CV | `Survival.filter_infeasible` |
+| Feasible hyperplane | Ideal, worst, and ASF use feasible objectives when any feasible member exists. An all-infeasible generation does not move the hyperplane | `HyperplaneNormalization.update` sees the feasible subset passed by `filter_infeasible` |
+| Equal CV in the sort | Two infeasible individuals with the same violation are mutually non-dominated (Deb). Objectives are not compared | constraint-domination; CV ordering is separate from Pareto |
+| Tournament equal CV | unchanged. `RankNicheDistance` still falls through to rank, niche, and distance. `PymooCompatible` still coin-flips | pymoo coin-flips |
 
 ## Mating pool
 

diff --git a/docs/NOTICE.md b/docs/NOTICE.md
@@ -12,7 +12,8 @@ It is **not** affiliated with, endorsed by, or derived from the source code of p
 | Work | Role in Unsga3 |
 |------|----------------|
 | Seada, H. & Deb, K. (2016). *A Unified Evolutionary Optimization Procedure for Single, Multiple, and Many Objectives.* IEEE TEVC. | U-NSGA-III algorithm |
-| Deb, K. & Jain, H. (2014). *NSGA-III.* IEEE TEVC. | Reference-point survival, hyperplane normalization |
+| Deb, K. & Jain, H. (2014). *An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints.* IEEE TEVC. | Reference directions (including the inside layer), hyperplane normalization, niching |
+| Jain, H. & Deb, K. (2014). *An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part II: Handling Constraints and Extending to an Adaptive Approach.* IEEE TEVC. | Feasible-only niching and constraint-violation fill |
 | Das, I. & Dennis, J. E. (1998). *Normal-Boundary Intersection.* SIAM J. Optim. | Das–Dennis reference directions |
 | Deb, K. et al. ZDT / DTLZ suites | Standard test problems |
 | Deb & Agrawal — SBX; Deb et al. — polynomial mutation | Variation operators |

diff --git a/docs/ORACLE-RESULTS.md b/docs/ORACLE-RESULTS.md
@@ -131,7 +131,7 @@ ZDT2 quality A/B is gens=250 + `PymooCompatible` (not the loose smoke bar). ZDT1
 | IGD mean-distance | **aligned** |
 | ASF / axis intercepts | **aligned** |
 | Collapsed nadir (span ≤ 1e-6) | **delta**: nadir = ideal + 1 after the worst-of-pop fallback. pymoo 0.6.2 stops at worst-of-pop. Locked by `Collapsed_span_sets_nadir_to_ideal_plus_one` (`{2, 2+1e-8}` → nadir 3). |
-| Infeasible points | **current rule, locked**: ideal and worst include them. Fixture is feasible (1, 1) vs infeasible (0, 0) → ideal (0, 0). No constrained benchmark yet. |
+| Infeasible points | **feasible hyperplane**. Ideal, worst, and ASF use feasible objectives when any feasible member exists. An all-infeasible generation does not move the hyperplane. Survival niches the feasible set and fills a shortfall by ascending CV. The old whole-pool fixture (feasible (1, 1) vs infeasible (0, 0) → ideal (0, 0)) is replaced. OSY, TNK, and C1-DTLZ1 have self-tests. No constrained IGD table. |
 | Mating re-association | **delta**: after survival, `PrepareForSelection` normalizes the survivors again and overwrites niche ids. pymoo keeps the survival ids. Fixture: `PrepareForSelection_overwrites_survival_niche_ids`. |
 | Persistent ideal + ND extremes | **aligned** |
 | `TournamentMode.PymooCompatible` | **implemented** |

diff --git a/docs/RESEARCH-STANDARDS.md b/docs/RESEARCH-STANDARDS.md
@@ -19,7 +19,7 @@ Sources consulted (2026-08):
 | Bi-objective | ZDT1–4, ZDT6 | Convex / non-convex / disconnected / multimodal / biased density |
 | Many-objective | DTLZ1–4, DTLZ7 | Linear simplex, sphere, multimodal sphere, bias, disconnected |
 | Hard (later) | WFG1–9 | Non-separable, scaled, deceptive |
-| Constrained (later) | OSY, TNK, C1-DTLZ1 | Constraint-domination path |
+| Constrained | OSY, TNK, C1-DTLZ1 | In the library, with survival self-tests. No IGD table until measured runs exist |
 
 **Seada selection paper** stresses ZDT4 (many local fronts) and DTLZ1 for M=3,5.
 
@@ -35,7 +35,7 @@ Default dimensions (Deb / pymoo convention):
 |------|----------------|
 | Crossover | SBX, η_c = **30**, p_c = **1.0** (pymoo). Paper section 4 uses p_c = **0.9** |
 | Mutation | Polynomial, η_m = **20**, p_m = **1/n** |
-| Reference set | **Das–Dennis** (uniform) on the unit simplex, **single layer**. Two-layer directions for larger M are absent |
+| Reference set | **Das–Dennis** on the unit simplex. **Single layer is the default** (`DasDennis`, `WithDasDennis`, M≤3 oracles). `TwoLayerDasDennis` adds Part I's inside layer (default scale 1/2) for larger M |
 | Population size | Often = #reference directions (or slightly larger). N ≥ 2. `WithDasDennis(1, 1)` throws because |H| = 1; pass an explicit population size for single-objective runs |
 | Selection | U-NSGA-III **tournament** (not NSGA-III random mating) |
 

diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
@@ -18,7 +18,8 @@ Living plan for Unsga3. Issues track concrete work; this page is the narrative.
 - [ ] `net8.0` (+ `net10.0`) multi-target for broader NuGet consumers  
 - [ ] nuget.org publish (in addition to GitHub Packages)  
 - [ ] GD+ indicator (IGD+ has a hand-case test; GD and IGD were already implemented)
-- [ ] Constrained demos (OSY / TNK) with self-tests  
+- [x] Constrained demos (OSY, TNK, C1-DTLZ1) with formulation and survival self-tests  
+- [ ] Constrained IGD / Wilcoxon vs pymoo (no table until measured runs exist)  
 - [ ] API docs site (DocFX or similar)  
 
 ## Medium term (0.3+)

diff --git a/src/Unsga3/Algorithm/Unsga3Algorithm.cs b/src/Unsga3/Algorithm/Unsga3Algorithm.cs
@@ -85,7 +85,9 @@
     }
 
     /// <summary>
-    /// Convenience: build Das–Dennis directions then construct the algorithm.
+    /// Convenience: build single-layer Das–Dennis directions then construct the algorithm.
+    /// Two-layer directions are <see cref="ReferenceDirections.TwoLayerDasDennis"/>, passed
+    /// to the constructor; M≤3 oracles stay on this single-layer factory.
     /// One objective produces a single direction, so the default population size is 1.
     /// The constructor requires N ≥ 2, and <c>WithDasDennis(1, 1)</c> throws.
     /// Single-objective runs must pass <paramref name="populationSize"/> ≥ 2
@@ -118,8 +120,8 @@
    /// are pad-filled randomly.
    /// </param>
    public OptimizationResult Run(
        IProblem problem,
        int maxGenerations,
        IReadOnlyList<Individual>? initialPopulation) =>
        Run(problem, TerminationCriterion.MaxGenerations(maxGenerations), initialPopulation);


diff --git a/src/Unsga3/Core/NonDominatedSort.cs b/src/Unsga3/Core/NonDominatedSort.cs
@@ -85,7 +85,10 @@ public static int CompareConstraintDominated(Individual a, Individual b)
         {
             if (a.ConstraintViolation < b.ConstraintViolation) return -1;
             if (a.ConstraintViolation > b.ConstraintViolation) return 1;
-            // equal violation → fall through to objective comparison
+            // Deb constraint-domination: equal violation ⇒ neither dominates.
+            // Objectives are not consulted. Tournament comparators are separate
+            // and still break an equal-CV tie their own way.
+            return 0;
         }
 
         return ComparePareto(a.Objectives, b.Objectives);