Skip to content

Release develop to master - #552

Merged
martyn merged 72 commits into
masterfrom
develop
Nov 20, 2026
Merged

martyn merged 72 commits into
masterfrom
develop

Conversation

@martyn

@martyn martyn commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Release v0.11.0 (develop -> master): promote unified_v11 default, publish technique-inventory.md + 66-cell scaling, document The Four Laws, and ship the new 98.52% PASS 100gaussians.gif hero GIF.

Release 0.11.0 Highlights

  • Default Unified 4-Law Formulation (DEFAULT_RECIPE = "unified_v11"): Strictly 0 torch.autograd.grad, 0 SVD/eigh/pinv across every forward, backward, penalty, transport, conditional fiber bank, prior, and optimizer path. Leads technique-inventory.md at 95/115 counted view requirements (27/27 Tier 1, 68/78 Tier 2), 40/69 full-suite PASS (6 FAIL, 11 not measured, 12 blocked), 50/68 stretch view assignments, 41/47 Tiers 1–2 discriminator_stability PASS (55/62 deduplicated unique runnable tasks PASS across 9 studies), and 90.24 ms/step median (1,372.60 MiB allocated VRAM) on matched 66-cell dcgan_performance_r256 (1.44× faster than empf-v8, 6.54× faster than rhbi-v12).
  • Minimal-Principle <900-LOC Fast Core (preset="minimal"): Self-contained 899-LOC zero-sibling-import core with zero autograd.grad, zero SVD/eigh/pinv, and zero Newton–Schulz (55.50 ms/step, 1,344.84 MiB on dcgan_performance_r256).
  • New 100gaussians.gif Hero GIF (98.52% PASS): Actual-training 0.11.0 default (unified_v11) + MoGParticlePrior(20000) (100/100 modes, 98.520% within 3σ PASS, 5/5 terminal checks, 4.19 MiB GIF; archived 0.9.0 ParticlePrior(20000) 42.115% FAIL preserved with commit/blob SHAs).

The Four Laws of Stable, Energy-Efficient GAN Convergence

Full derivation, causal knockouts, and verbatim receipts: https://github.com/255BITS/ParticleGAN/blob/develop/docs/four-laws.md

  1. Law 1 — Measure & Prior Topology: Streaming Laguerre–Pythagorean prototype quantization with Law-of-Total-Variance coalescence, quota-balanced Laguerre Monge commitment, and orthogonal pullback covariance isometry resolve discontinuous pushforward density bridges (t1-prior-v2, t1-coupling-v5, r4-occam-ablation-razor Arms 4–5).
  2. Law 2 — First-Order Critic Potential Geometry: 1st-order fwAD canonical/Householder/Murmur3-Rademacher tangents + live-endpoint BF16 micro-secants eliminate 2nd-order torch.autograd.grad double-backprop (1.48× faster and -34.8% VRAM with higher pass rate in r4-occam-ablation-razor Arm 7).
  3. Law 3 — Scale–Mass–Fiber Gauge Identifiability (The 3rd Missing Piece): Scale-invariant $k=4$ density-ratio orphan macro-transport, 1D Weibull batch-horizon scaling, and swap-invariant zero-eigh cross-batch conditional fiber operator banks resolve scale–mass heteroscedasticity, $42.4\sigma$ saddle shifts, and model_ddgan_gaussian_chain (rgdt-v3, cond-fiber-v1, unified-v11).
  4. Law 4 — Function-Space Two-Player Equilibration & Depth/Rail Trust: $10^{-38}$ dynamic scale-normalized peak_rms, Marchenko–Pastur operator scaling, realized generator pushforward trust ceiling delta_g_max, zero-pinv analytic direction_blend, and fused $Z=[X,Y]$ Quintic Tikhonov polar equilibration solve >600× Nash imbalances and 16-layer/rail architectures (t4-nash-v3, t4-arch-v3, unified-v11).
  • Occam's Razor Synthesis: Round-4's 8-arm causal knockout study (r4-occam-ablation-razor.md) shows Laws 1–4 collapse into Two Mathematical Halves (Measure/Potential Geometry + Function-Space Two-Player Dynamics) with depth-gated 2D Newton–Schulz strictly required only for deep skip-free MLPs (grid100_mlp16).

Honest Governance Disclosures

  • Forge screening calibration remains provisional (BLOCKED under configs/forge/calibration/criteria-v1.json; default_adoption: false). Remaining numerical failures of unified-v11 (staggered100, grid100_res16_tanh, five_word_joint_hold, model_ddgan_bimodal_chain, model_ae_ddgan_gaussian_chain, interp_img_patch_shift, interp_saddle_barycentric) and 12 unimplemented Tier-4 stretch hosts (BLOCKED) stay explicitly disclosed.

…itic

feat(t2-critic): 1st-order spatial-frame forward-tangent & secant critic (t2_secant)
feat(prior): Theory 1A Gersho-Laguerre pullback prior dynamics (t1_prior)
…-coupling

feat(t1-coupling): Track 1B deconflicted piecewise chart coupling and conformal transport
feat(rgdt): Track 3 — Solve GAN scale-mass gauge conflation & horizon starvation (RGDT)
feat(unified): Unified Synthesis — 1st-order fwAD/secant critic, isolated Pythagorean pullback prior, persistent chart coupling & conformal fiber deconfliction
…o feat/solve-gans-r3-conditional-ddgan

# Conflicts:
#	particlegan/recipes.py
#	particlegan/training.py
…ddgan

feat(cond-fiber): cross-batch conditional fiber operator bank — first-ever DDGAN Gaussian & Bimodal PASS (40/47 T1–2, 6/6 DCGAN @ 87.50 ms)
…-unified-v4

# Conflicts:
#	experiments/forge/api.py
#	experiments/forge/techniques.py
#	particlegan/__init__.py
#	particlegan/optim/dualnorm.py
#	particlegan/recipe_schedules.py
#	particlegan/recipes.py
#	particlegan/training.py
#	reports/forge/EXPERIMENT_MEMORY.md
#	reports/forge/compilation.json
feat(unified-v4): Round-3 Track 2 complete 4-Law zero-SVD / zero-autograd.grad unification (40/46 disc_stability, 6/7 t4-arch, 3/3 Nash, 3/3 14k endurance, 69.55-89.66 ms/step r256)
Round-4 Track 2: Occam's Razor systematic 8-arm ablation study (18 official Forge readouts)
…o feat/solve-gans-r4-minimal-core

# Conflicts:
#	experiments/forge/clockfree.py
#	experiments/forge/techniques.py
#	particlegan/recipes.py
#	particlegan/training.py
#	reports/forge/EXPERIMENT_MEMORY.md
#	reports/forge/compilation.json
#	tests/test_m4_it2_challenger2_stress.py
feat(minimal): collapse 4-law core into 899-LOC zero-NS minimal preset (R4 Track 1)
…d 4-Law + CondFiber unified_v11 recipe and 9 Forge readouts
…o feat/solve-gans-r4-solution-complete

# Conflicts:
#	experiments/forge/api.py
#	experiments/forge/clockfree.py
#	experiments/forge/techniques.py
#	particlegan/optim/dualnorm.py
#	particlegan/recipes.py
#	particlegan/training.py
#	tests/test_m4_it2_challenger2_stress.py
…plete

Round-4 Track 3: Solution-Complete 4-Law + Zero-eigh CondFiber Champion (unified_v11)
…tory, 66-cell scaling, and docs/four-laws.md
release(0.11.0): promote unified_v11 default, publish technique inventory, 66-cell scaling, docs/four-laws.md, and 98.52% PASS hero GIF
@martyn
martyn merged commit b86f037 into master Nov 20, 2026
22 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant