feat(rhbi): Relativistic Hellinger-Born-Infeld transport family (rhbi-v3: 27/27 Tier 1, 76/115 view reqs, 21 full-suite PASS)

configs/forge/campaigns/dcgan-performance-cloud-2edfe08a4bf5.json

@@ -0,0 +1,5 @@
+{
+  "budget_seconds": 10800,
+  "candidate_budget_seconds": 1200,
+  "id": "dcgan-performance-cloud-2edfe08a4bf5"
+}

configs/forge/campaigns/dcgan-performance-cloud-38b89eb060b8.json

@@ -0,0 +1,5 @@
+{
+  "budget_seconds": 10800,
+  "candidate_budget_seconds": 1200,
+  "id": "dcgan-performance-cloud-38b89eb060b8"
+}

configs/forge/family-documentation/rhbi.json

@@ -0,0 +1,83 @@
+{
+  "overview": "Relativistic Hellinger-Born-Infeld Transport (RHBI) is a clock-free, constant-learning-rate ParticleGAN formulation with no EMA critic copy and no task-shape or step-count gates. It unifies nonlocal scale-equivariant Born-Infeld Optimal Transport, Feydy diagonal-included self-coupling entropic debiasing, local relativistic critic log-odds ascent, and fibre-centered conditional relativistic slack routing inside a single (B+1)x(B+1) slack-screened Sinkhorn variational operator, paired with a smooth C-infinity Born-Infeld endpoint input-gradient cap and degree-of-freedom-aware SBNP particle prior regularization.",
+  "symbols": "$G$ is the generator, $D$ the raw-score critic, $P$ the task prior with $N$ locations in $\\mathbb R^{d_z}$, $y=\\{y_j\\}_{j=1}^{B_r}$ a real batch, $z\\sim P$ a latent batch and $x=\\{x_i\\}_{i=1}^{B_f}$ with $x_i=G(z_i)$ the generated batch. $\\ell_t$ is the exponential moving average ($\\beta_\\ell=0.1$) of the detached 25th-quantile real nearest-neighbor Euclidean distance. $c_{\\mathrm{phys}}(x_i,y_j)=\\sqrt{\\ell_t^2+\\lVert x_i-y_j\\rVert_2^2}-\\ell_t$ is the physical Born-Infeld ground cost and $\\hat c_{ij}=c_{\\mathrm{phys}}(x_i,y_j)/\\ell_t$ its dimensionless linear screening form. $\\tilde d(u)=C_{\\max}\\tanh((D(u)-\\bar D)/C_{\\max})$ is the batch-centered, bounded critic log-odds with $\\bar D=\\frac{1}{2}(\\mathbb E[D(x)]+\\mathbb E[D(y)])$. $P^{\\star(XY)}$ and diagonal-included $P^{\\star(XX)}$ are deterministic Sinkhorn couplings on the $(B+1)\\times(B+1)$ active-plus-slack polytope with linear screening horizon $\\rho=3$, relativistic gain $\\tau=2$ and entropic temperature $\\varepsilon=0.1$. $g_r=\\nabla_y D(y)$ and $g_f=\\nabla_x D(x)$ are critic input gradients capped at $\\kappa=1$ with Born-Infeld softness $\\beta=20$ and sub-unit slope floor $\\lambda_{\\mathrm{floor}}=0.005$.",
+  "pseudocode": [
+    "Fix one global configuration with constant learning rates and DualNorm updates across generator, encoder, critic and row-normalized particle prior.",
+    "For each task-owned training update:",
+    "  Sample real batch y and latent batch z from the task prior P; form detached fake batch x_det = stop_gradient(G(z)).",
+    "  Evaluate calibrated non-saturating logistic critic loss on D(y) and D(x_det) plus the smooth Born-Infeld endpoint input-gradient cap (with soft slope floor 0.005) on y and x_det.",
+    "  Backpropagate and step D via DualNorm; clear transient critic input-capture ring.",
+    "  With D fixed, resample latent z and compute differentiable fake batch x = G(z) together with critic forward captures.",
+    "  If continuous conditioning accompanies critic tuples D(c, x), evaluate fibre-centered paired relativistic log-odds softplus(D(c_i, y_i) - D(c_i, x_i)).",
+    "  Otherwise update detached local horizon ell_t on device from the 25th-quantile nearest-neighbor distance of y and solve K=8 deterministic log-domain Sinkhorn scaling steps for active-plus-slack coupling P*(XY) and Feydy diagonal self-coupling P*(XX).",
+    "  Evaluate the Danskin envelope RHBI generator loss combining active Born-Infeld pull, slack-routed critic ascent and Feydy self-coupling debiasing (or slack-screened relativistic log-odds on joint BiGAN hosts), plus SBNP prior regularization and any task-owned auxiliary losses.",
+    "  Project conflicting protected directions via direction-blend constraint geometry when declared, and step G, P and optional encoder E."
+  ],
+  "training_details": {
+    "loss": "Critic minimizes non-saturating logistic loss; unconditional and discrete-class generators minimize the RHBI slack-screened unbalanced Sinkhorn envelope objective combining physical Born-Infeld transport, Feydy diagonal self-coupling debiasing and slack-weighted centered critic log-odds ascent; continuous-conditioned fibres minimize fibre-centered paired relativistic log-odds; joint BiGAN generator/encoder hosts minimize slack-screened relativistic log-odds across active pairs and slack marginals.",
+    "optimizer": "Full DualNorm across generator, encoder, critic and row-normalized particle prior (zero momentum, polar smoothing 0.0061, exact CPU-LAPACK smoothed polar factor on dense 2D weights, quintic per-offset convolution rule on wide CUDA kernels, direction-blend constraint geometry) without an EMA critic copy.",
+    "learning_rates": "Strictly constant across all tasks: global LR $0.014$, critic multiplier $1.5$ and prior multiplier $2.9$, with network and prior LR floors fixed at $1.0$ (zero schedule annealing and zero horizon dependence).",
+    "gradient_clipping": "None on parameter tensors; active transport coordinate gradients are analytically bounded inside $[-1,+1]$ by the Born-Infeld ground cost and critic log-odds are smoothly bounded by $C_{\\max}\\tanh(\\cdot/C_{\\max})$.",
+    "critic_penalty": "Smooth C-infinity Born-Infeld endpoint input-gradient cap evaluated on real and fake inputs ($\\lambda=1$, $\\kappa=1$, sharpness $\\beta=20$, sub-unit slope floor $\\lambda_{\\mathrm{floor}}=0.005$) with zero EMA critic anchor and automatic non-concatenated evaluation on stateful batch-bound critic view wrappers.",
+    "damping_and_guards": "None: critic spike guard ratio=$0$, latent damping=$0$, and zero shape or step-count gates; slack reservoir marginal $S=2$ keeps Sinkhorn iterations strictly inside the Birkhoff polytope while degree-of-freedom-aware SBNP regularization ($\\lambda_{\\mathrm{prior}}=0.02$) stabilizes particle geometry across dimensions.",
+    "noise": "Zero additive critic-input or generator-output training noise ($0.0$). Task-declared prior intrinsic sigma and serving contracts are preserved unchanged.",
+    "averaging": "Generator EMA decay=$0$ and serve_average=$0$; all evaluations score live unaveraged weights."
+  },
+  "configuration_notes": [
+    "RHBI uses one global trainer configuration across every Tier 1, Tier 2, Tier 3 and Tier 4 task without per-task tuning, step-count gates or input-shape regimes.",
+    "When discrete class labels accompany critic inputs, cross-class active pairs are masked to minus infinity in log-Sinkhorn space so unmatched class draws route cleanly into the slack reservoir; when continuous conditioning accompanies critic tuples D(c, x), rhbi_conditional_fibre='slack' routes singleton fibres to fibre-centered paired relativistic log-odds.",
+    "Both passive observer records and formulation critic records checkpoint ell_ema and ell_initialized with zero per-step host-device synchronizations, ensuring exact bit-for-bit continuation across restarts and rejecting incompatible checkpoints.",
+    "Setting rhbi_weight=0.0 executes an exact zero-strength bypass that leaves horizon buffers and RNG streams untouched."
+  ],
+  "references": [
+    {
+      "title": "Sinkhorn Distances: Lightspeed Computation of Optimal Transport",
+      "url": "https://arxiv.org/abs/1306.0895",
+      "component": "Entropic Optimal Transport Sinkhorn-Knopp scaling extended to the (B+1)x(B+1) slack-screened unbalanced polytope."
+    },
+    {
+      "title": "Unbalanced Optimal Transport: Dynamic and Kantorovich Formulations",
+      "url": "https://arxiv.org/abs/1508.05216",
+      "component": "Kantorovich-Hellinger unbalanced mass creation/destruction theory motivating slack-reservoir screening of Poisson batch singletons and empty-valley outliers."
+    },
+    {
+      "title": "Interpolating between Optimal Transport and MMD using Sinkhorn Divergences",
+      "url": "https://arxiv.org/abs/1810.08278",
+      "component": "Feydy diagonal-included symmetric Sinkhorn self-coupling debiasing adapted to the active-plus-slack unbalanced polytope."
+    }
+  ],
+  "sources": [
+    "particlegan/rhbi.py",
+    "particlegan/_rhbi/primitive.py",
+    "particlegan/_rhbi/loss.py",
+    "particlegan/_rhbi/critic.py",
+    "particlegan/_rhbi/penalty.py",
+    "particlegan/optim/dualnorm.py",
+    "particlegan/recipes.py",
+    "particlegan/training.py",
+    "configs/forge/ideas/rhbi-v1.json",
+    "configs/forge/ideas/rhbi-v2.json",
+    "configs/forge/ideas/rhbi-v3.json",
+    "configs/forge/studies/rhbi-v1-study.json",
+    "configs/forge/studies/rhbi-v2-study.json",
+    "configs/forge/studies/rhbi-v3-study.json",
+    "reports/forge/rhbi/README.md"
+  ],
+  "equations": [
+    {
+      "label": "Scale-Equivariant Born-Infeld Cost and Local Horizon",
+      "latex": "c_{\\mathrm{phys}}(x_i,y_j)=\\sqrt{\\ell_t^2+\\lVert x_i-y_j\\rVert_2^2}-\\ell_t,\\qquad \\hat c_{ij}=\\frac{c_{\\mathrm{phys}}(x_i,y_j)}{\\ell_t},\\qquad \\ell_t=(1-\\beta_\\ell)\\ell_{t-1}+\\beta_\\ell\\,\\operatorname{stopgrad}\\!\\left(Q_{0.25}(d^{\\mathrm{NN}}_Y)+10^{-6}\\right)",
+      "explanation": "Evaluated via float64 GEMM and algebraic rationalization so gradients remain smooth at zero distance and finite across the full float32 range, matching exact quadratic W2 inside modes while bounding inter-mode pull slopes inside $[-1,+1]$."
+    },
+    {
+      "label": "Slack-Screened Unbalanced Sinkhorn Envelope Objective",
+      "latex": "\\mathcal L_{\\mathrm{RHBI}}(X,Y;\\psi)=\\lambda_{\\mathrm{rhbi}}\\left[\\sum_{i=1}^{B_f}\\sum_{j=1}^{B_r}\\operatorname{stopgrad}\\!\\left(P^{\\star(XY)}_{ij}\\right)c_{\\mathrm{phys}}(x_i,y_j)-\\sum_{i=1}^{B_f}\\operatorname{stopgrad}\\!\\left(P^{\\star(XY)}_{i,0}\\right)\\frac{\\tau}{2}\\tilde d(x_i)-\\frac{1}{2}\\sum_{i,k=1}^{B_f}\\operatorname{stopgrad}\\!\\left(\\sqrt{a_i a_k}P^{\\star(XX)}_{ik}\\right)c_{\\mathrm{phys}}(x_i,x_k)\\right]",
+      "explanation": "Linear horizon screening ($\\rho=3$) routes valley outliers ($\\hat c_{ij}>\\rho$) to unit-slope centered critic log-odds ascent ($\\tau/2=1$), while Feydy diagonal self-coupling cancels intra-mode entropic contraction without inflating isolated singletons."
+    },
+    {
+      "label": "Smooth Born-Infeld Endpoint Critic Penalty",
+      "latex": "R_{\\mathrm{RHBI}}(D)=\\frac{\\lambda}{2}\\sum_{u\\in\\{y,x\\}}\\mathbb E_{u}\\!\\left[\\phi_\\beta(\\lVert\\nabla_u D(u)\\rVert_2-\\kappa)^2+\\lambda_{\\mathrm{floor}}\\lVert\\nabla_u D(u)\\rVert_2^2\\right],\\qquad \\phi_\\beta(s)=\\frac{1}{\\beta}\\log\\!\\left(1+e^{\\beta s}\\right)",
+      "explanation": "Softly caps critic input slopes at $\\kappa=1$ ($\\beta=20$) and damps residual sub-unit slope chatter ($\\lambda_{\\mathrm{floor}}=0.005$) on real and fake endpoints without transport-chord interpolation, shape gates or an EMA critic anchor."
+    }
+  ]
+}

configs/forge/ideas/rhbi-v1.json

@@ -0,0 +1,181 @@
+{
+  "api_changes": [
+    "particlegan/__init__.py",
+    "particlegan/_rhbi/__init__.py",
+    "particlegan/_rhbi/critic.py",
+    "particlegan/_rhbi/loss.py",
+    "particlegan/_rhbi/penalty.py",
+    "particlegan/_rhbi/primitive.py",
+    "particlegan/optim/dualnorm.py",
+    "particlegan/recipe_schedules.py",
+    "particlegan/recipes.py",
+    "particlegan/rhbi.py",
+    "particlegan/training.py"
+  ],
+  "api_version": "forge-api-v1",
+  "changed_factors": [
+    "Introduce Relativistic Hellinger-Born-Infeld Transport (critic_formulation=rhbi, rhbi_weight=1.0, rhbi_rho=2.0, rhbi_tau=2.0, rhbi_eps=0.1, rhbi_iters=8, rhbi_c_max=20.0, rhbi_beta_ell=0.1, rhbi_nn_quantile=0.25, rhbi_critic_weight=0.0): replace unscreened adversarial generator loss and piecewise critic caps with a single debiased slack-screened unbalanced Sinkhorn variational generator objective paired with a smooth C-infinity Born-Infeld endpoint input-gradient cap (reg_coeff=1.0, reg_kappa=1.0, ebi_beta=20.0), global moment-tether prior regularization (prior_reg=0.01, prior_lr_mult=2.5), and DualNorm updates across generator, encoder, critic and row-normalized particle prior (lr=0.012, d_lr_mult=1.5, optimizer_smoothing=0.001, optimizer_convolution=per_offset, optimizer_convolution_rule=quintic, optimizer_dense_profile=polar_cpu_v1, constraint_geometry_mode=direction_blend) under strictly constant learning rates (lr_floor=1.0, network_lr_floor=1.0) and zero shape or step gates."
+  ],
+  "claim_contract": {
+    "sampling_law": "task_declared",
+    "schedule": "scheduled",
+    "scoring_weights": "live"
+  },
+  "execution_path": "public_trainer",
+  "guide": "reports/forge/rhbi/README.md",
+  "host_adaptation": {
+    "recipe_fields": [
+      "alpha_bar",
+      "conditioning",
+      "distance_reduction",
+      "encoder_mode",
+      "model",
+      "num_classes",
+      "observation_sigma",
+      "prior_reg",
+      "reconstruction_weight",
+      "routing_temperature",
+      "ucd_target",
+      "ucd_weight"
+    ],
+    "schema_version": 1
+  },
+  "id": "rhbi-v1",
+  "mechanism_class": "structural",
+  "mechanism_rationale": "Relativistic Hellinger-Born-Infeld Transport (RHBI) unifies nonlocal scale-equivariant Born-Infeld Optimal Transport, active-fraction-matched hollow self-coupling entropic debiasing, and slack-routed relativistic critic log-odds ascent inside a single (B+1)x(B+1) unbalanced Sinkhorn operator with zero RNG draws, zero EMA critic copies, zero shape/step gates, and constant learning rates across all tasks. Unmatched batch singletons route to the slack reservoir instead of hallucinating cross-mode jumps, while Born-Infeld ground costs smoothly bound active transport coordinate gradients inside [-1, +1] to match the unit Lipschitz critic slope.",
+  "parent": "recipe-prior-refactor-bcap-pure-v1",
+  "recipe_overrides": {
+    "adam_variant": "pytorch",
+    "alpha_bar": [
+      1.0,
+      0.9,
+      0.5,
+      0.05,
+      0.0001
+    ],
+    "amsgrad": false,
+    "beta2_anneal_end": 0.2,
+    "beta2_end": null,
+    "betas": [
+      0.0,
+      0.999
+    ],
+    "birth_death_backend": "knn",
+    "birth_death_cells": 64,
+    "birth_death_chunk": 256,
+    "birth_death_feature_scale": "none",
+    "birth_death_isolation": false,
+    "birth_death_metric_rank": 8,
+    "birth_death_parent_policy": "real_anchor",
+    "birth_death_space": "data",
+    "conditioning": "scalar",
+    "constraint_geometry_mode": "direction_blend",
+    "continuous_policy": null,
+    "critic_formulation": "rhbi",
+    "critic_payoff_damping": true,
+    "critic_r1_real": true,
+    "critic_step_mode": "none",
+    "d_betas": null,
+    "d_eps": null,
+    "d_guard_min_steps": 200,
+    "d_guard_ratio": 0.0,
+    "d_lr_mult": 1.5,
+    "direct_particle_betas": [
+      0.0,
+      0.9
+    ],
+    "direct_particle_gain": false,
+    "distance_reduction": "sum",
+    "ebi_beta": 20.0,
+    "ema_decay": 0.0,
+    "encoder_mode": "none",
+    "eps": 1e-08,
+    "generator_gradient_response": "none",
+    "input_noise_anneal_end": 0.1,
+    "input_noise_std": 0.0,
+    "kinetic_transport_local_weight": 0.0,
+    "kinetic_transport_projections": 32,
+    "kinetic_transport_weight": 0.0,
+    "latent_damping_max_rate": 0.0,
+    "loss": "non_saturating",
+    "loss_labels": [
+      0.0,
+      1.0,
+      1.0
+    ],
+    "lr": 0.012,
+    "lr_anneal_start": 0.6,
+    "lr_control": "mobility",
+    "lr_decay_rate": 0.96,
+    "lr_decay_staircase": false,
+    "lr_decay_steps": 50000,
+    "lr_floor": 1.0,
+    "lr_schedule": "cosine",
+    "model": "gan",
+    "name": "rhbi",
+    "network_lr_floor": 1.0,
+    "network_lr_horizon_cap": null,
+    "num_classes": null,
+    "observation_sigma": 0.03,
+    "optimizer_adam_lr": null,
+    "optimizer_convolution": "per_offset",
+    "optimizer_convolution_rule": "quintic",
+    "optimizer_dense_profile": "polar_cpu_v1",
+    "optimizer_family": "dualnorm",
+    "optimizer_momentum": 0.0,
+    "optimizer_smoothing": 0.001,
+    "optimizer_svd_backend": "native",
+    "output_noise_mode": "fixed",
+    "output_noise_std": 0.0,
+    "output_noise_warmup": 0.0,
+    "particle_birth_death": false,
+    "prior_betas": null,
+    "prior_eps": null,
+    "prior_l2": 0.0,
+    "prior_lr_mult": 2.5,
+    "prior_reg": 0.01,
+    "prior_reg_eps": 0.0001,
+    "prior_reg_target_std": 1.0,
+    "prior_regularizer": "moment_tether",
+    "prior_update": "learned",
+    "reconstruction_weight": 1.0,
+    "reg_anchor_min_decay": 0.9,
+    "reg_anchor_weight": 0.0,
+    "reg_arm": null,
+    "reg_coeff": 1.0,
+    "reg_coeff_anneal_end": 0.2,
+    "reg_coeff_end": null,
+    "reg_every": 1,
+    "reg_kappa": 1.0,
+    "reopen_anchor": "hold",
+    "reopen_guard": null,
+    "reopen_signal": "data",
+    "rhbi_beta_ell": 0.1,
+    "rhbi_c_max": 20.0,
+    "rhbi_critic_weight": 0.0,
+    "rhbi_eps": 0.1,
+    "rhbi_iters": 8,
+    "rhbi_nn_quantile": 0.25,
+    "rhbi_rho": 2.0,
+    "rhbi_tau": 2.0,
+    "rhbi_weight": 1.0,
+    "routing_temperature": 0.25,
+    "row_evidence_exclude": true,
+    "row_evidence_gate": false,
+    "row_evidence_hold": true,
+    "row_evidence_hot": true,
+    "row_evidence_null": "theory",
+    "row_policy": "independent",
+    "serve_average": 0.0,
+    "table_release_rule": "any",
+    "ucd_target": "class",
+    "ucd_weight": 0.02
+  },
+  "recipe_preset": null,
+  "requires_capabilities": [
+    "learned_locations",
+    "named_rng"
+  ],
+  "schema_version": 3,
+  "trainer_family": "rhbi"
+}