[martyn]

Summary

Implements EMPF (Exponential Moving Prototype Flow) (particlegan/empf.py, particlegan/_empf/), a standalone ParticleGAN trainer family (critic_formulation="empf", prior_regularizer="empf", preset "empf") inspired by e-means (Bessa et al., arXiv:2611.09501), together with its full Forge CUDA qualification (empf-v1), 9-family DCGAN scaling campaign, and current leaderboard publication.

DEFAULT_RECIPE = "bispf_optimized" and all historical receipts, snapshots, and incumbent recipe digests remain untouched.

Formulation Overview (critic_formulation="empf", prior_regularizer="empf")

1. Online (S, N) Sufficient-Statistics Prototype Bank (particlegan/_empf/prototypes.py):
   • Replaces noisy raw real-minibatch Sinkhorn endpoints with persistent prototype sums and effective counts (S_k, N_k) updated via a clock-free (K, B) hit-rate EMA (alpha = clip(alpha_base * (B / K)^0.45, 0.03, 0.35)), dropping dataset-size and epoch-horizon clocks.
   • Deterministic proportional-split dead-prototype revival: When N_k < revival_tau * mean(N), splits an active donor selected via golden/plastic (phi_1, phi_2) quasi-Monte Carlo CDF sampling while conserving exact total mass ((N_donor + N_dead) / 2) with zero RNG consumption.
2. Mass-Weighted Prototype-to-Fake Entropic Sinkhorn Chords (particlegan/_empf/transport.py, penalty.py):
   • Solves scale-free log-domain Sinkhorn with non-uniform prototype row marginals w_k = N_k / sum_j N_j against uniform generated column marginals 1 / B_f in $O(B \cdot K)$ rather than $O(B^2 \cdot D_{\text{img}})$, preserving rare-mode transport mass while preserving exact row-aligned fiber pairing on conditioned inputs.
   • Pairs prototype-chord EBI directional gradient capping, zero-mean logit gauge (gamma_gauge), and state-driven smoothstep thermal cooling (EMPFStepRecord, serialized across checkpoints with zero module-level global state).
3. Latent Prior Regularizer with Differentiable Dead-Particle Revival (particlegan/_empf/prior.py):
   • Combines an initial-moment covariance tether, scale-free $k$-NN log-repulsion, a short-range one-sided discrete codebook non-collision hinge (N <= min_dof), and a smooth dead-particle revival tether (d_{1\text{nn}}^2 / (isolation_ratio * r_{\text{med}}^2) - 1) pulling stranded outlier particles back to active clusters.

Forge CUDA Readout (empf-v1, Seed 0, NVIDIA RTX A6000)

Executed commit: cb5e484b059d3474f7259c40beb7acb4b5bb8704 (readout-d46a41a1da2da54d67d154f4.json, 40 attempts, 0 INCOMPLETE, 1109/1109 attempt coverage):

Selected Recipe / Family	Tier 1 (27)	Tier 2 (78)	Tier 3 (10)	Total View Reqs (115)	Full-Suite 61 PASS	Tuning Budget
BISPF (bispf-optimized-quintic-v1)	22/27 (4 not run)	55/78 (4 not run)	4/10	81/115	18	19 cards
EMPF (empf-v1, new)	27/27	48/78	0/10 (3 blocked)	75/115	20	untuned (1 card, 0 searches)
BCAP (recipe-prior-refactor-bcap-pure-v1)	22/27 (4 not run)	49/78 (4 not run)	0/10	71/115	16	85 searches / 473 cards
GAN v3 / K3P / KA2 / R1-R2	20/27 (4 not run)	0/78	0/10	20/115	5	8–46 cards (tuned)
Atlas (atlas-particle-normal-v4)	18/27	0/78	0/10	18/115	8	5 cards
E22 (e22)	0/27	0/78	0/10	0/115	0	untuned (1 card)

Disclosure: BISPF and BCAP pinned snapshots predate the 4 newer VAE/AE Tier 1 tasks (4 not run).

Key Scientific Findings

• 11/11 Required Tier 1 PASS (27/27 Tier 1 view requirements) + Clock-Free Audit PASS (clockfree_audit_measurement_v1, 4/4 parity comparisons): Passes two_pole, gaussian1d_smoke, zero_tanh_smoke, ring16_acquisition, five_word_joint_smoke, ae_gan_hold, unused_token_hold, model_ae_gan_multimodal, model_ae_reconstruction, model_vae_categorical, and model_vae_hard.
• Rare-Mode Preservation (vector_unequal_mass): Mass-weighted Sinkhorn (w_k = N_k / sum_j N_j) retains 2.98% mass (122/4096) on the 2% rare mode with component_core_min_eigen_ratio = 0.0595 (15× BISPF's 0.004 and 6.6× BCAP's 0.009).
• DCGAN Scaling Throughput (dcgan-performance-cloud-1acd41cb5f4a, NVIDIA RTX A6000): Fastest update throughput and lowest high-res peak VRAM across all 9 families at B64, 128², and 256²:
  • 64², B64, W64: 39.19 ms (1,628.56 img/s) vs BISPF 90.24 ms (710.67 img/s, 2.3×) and BCAP 674.06 ms (95.20 img/s, 17.1×).
  • 256², B16, W64: 156.88 ms (102.03 img/s, 1,522 MiB) vs BISPF 342.33 ms (46.72 img/s, 2,119 MiB, 2.2× faster / 28% less VRAM).
• Clear Structural Bottlenecks for EMPF v2:
  1. grid100 / rotated100 / staggered100: K_max = 64 < 100 target modes coarsens fine 10×10 lattice centers (center_max_sigma ~ 1.72–2.15).
  2. gaussian1d_stability & mode_hold (hq=0.8411 vs 0.90): stationary EMA inertia lags abrupt target shifts and loses 3 of 8 modes over long holds.
  3. model_ddgan_gaussian_chain / model_ae_ddgan_gaussian_chain: unconditioned prototype pooling over concatenated transition pairs [x_t, x_{t+1}] blurs conditional transition fibers.

Verification

• python scripts/check_lint.py — 0 new findings
• python ops/refresh_forge_readouts.py --check --reports-only — CURRENT (stale stages: none)
• uv build — wheel 401,152 bytes (< 600,000 limit)
• pytest — test_empf.py, test_refresh_forge_readouts.py, test_forge_comparison_policy.py, test_forge_trainer_families.py, test_core_inventory_legacy_documentation.py, test_recipe_defaults.py all passing