Skip to content

release(0.11.0): promote unified_v11 default, publish technique inventory, 66-cell scaling, docs/four-laws.md, and 98.52% PASS hero GIF - #583

Merged
martyn merged 10 commits into
developfrom
feat/release-unified-v11-0110
Nov 20, 2026
Merged

martyn merged 10 commits into
developfrom
feat/release-unified-v11-0110

Conversation

@martyn

@martyn martyn commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

Release 0.11.0 Prep: Promote unified_v11 Default, Publish Technique Inventory + 66-Cell Scaling, Document The Four Laws, and Ship New 98.52% PASS Hero GIF

Summary

  • Phase 0 (Sync master CI workflows): Cleanly merges origin/master (75cb78b6d: auto PyPI Trusted Publishing in .github/workflows/release.yml + scripts/publish_github_source.py) into the release branch while preserving develop's max_bytes = 1_200_000 wheel cap in .github/workflows/tests.yml.
  • Phase 1 (Technique Inventory & 66-Cell Scaling Publication): Registers unified (unified-v11) and minimal (minimal-v2) in configs/forge/trainer-families.json & configs/forge/family-documentation/, pins unified-v11 from clean source commit d438ade70 in configs/forge/selections/family-current-v1.json (preserving default_adoption: false and archiving the prior selection), refreshes all 66/66 clear dcgan_performance scaling cells (11 families × 6 resolutions on uncontended NVIDIA RTX A6000), and regenerates reports/forge/technique-inventory.md, EXPERIMENTS_BY_TIER.md, problems.md, EXPERIMENT_MEMORY.md, and compilation.json.
    • Unified · unified-v11 leads the canonical leaderboard: 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 (4/4 T1, 46/48 T2), and 90.24 ms/step median (1,372.60 / 1,402.00 MiB alloc/res) on matched 66-cell dcgan_performance_r256 (1.44× faster than empf-v8 129.72 ms/step, 6.54× faster than rhbi-v12 590.59 ms/step).
  • Phase 2 (Default Promotion & 0.11.0 Version Bump): Promotes DEFAULT_RECIPE = "unified_v11" in particlegan/recipes.py (leaving particlegan/recipe_compat.py and all historical reports/forge/records/readout-*.json receipts untouched byte-for-byte), bumps pyproject.toml to 0.11.0, updates CHANGELOG.md, README.md, docs/api.md, docs/releasing.md, RESEARCH_GLOSSARY.md, examples/pytorch_loop.py, .github/workflows/tests.yml, and updates all default-recipe & Forge readout tests.
  • New Main README Hero GIF (100gaussians.gif, 98.52% PASS): Publishes the actual-training 0.11.0 default (get_recipe() = unified_v11) + MoGParticlePrior(20000, z_dim=2, sigma=0.025) 100-Gaussian hero GIF (100/100 modes, 98.520% within 3σ PASS, 5/5 terminal checks 99.085% / 98.870% / 99.390% / 98.635% / 98.520%, Mode TV 0.04555, Sliced W1 0.08353, 4.19 MiB GIF) while preserving the archived 0.9.0 ParticlePrior(20000) 42.115% FAIL cohort with exact Git commit/blob SHAs (1bc01a7bf) in reports/readme-100gaussians/README.md.

The Four Laws of Stable, Energy-Efficient GAN Convergence

Canonical guide: docs/four-laws.md

  1. Law 1 — Measure & Prior Topology (Discontinuous Pushforward Resolution): Continuous $G_\theta(z)$ on a connected prior forces inter-mode density bridges and Jacobian cliffs. Resolved via streaming Laguerre–Pythagorean prototype quantization with Law-of-Total-Variance coalescence, quota-balanced Laguerre Monge commitment, and orthogonal pullback covariance isometry (t1-prior-v2, t1-coupling-v5, r4-occam-ablation-razor Arms 4–5).
  2. Law 2 — First-Order Critic Potential Geometry (Zero-autograd.grad Born–Infeld Secants): Replaces expensive 2nd-order create_graph=True double-backprop with 1st-order fwAD canonical/Householder/Murmur3-Rademacher tangents and live-endpoint BF16 micro-secants (t2-secant-v2/v5, r4-occam-ablation-razor Arm 7: 1.48× faster and -34.8% VRAM vs 2nd-order EBI with higher pass rate).
  3. Law 3 — Scale–Mass–Fiber Gauge Identifiability (The 3rd Missing Piece): Even with $G = \text{nn.Identity}()$ (0 generator params) and a smooth $C^\infty$ 1-Lipschitz critic, scalar $D(x)$ + single Euclidean horizon $\ell_t$ still fails 0/2 across BCap/EMPF/RHBI. Resolved via scale-invariant $k=4$ density-ratio orphan macro-transport, 1D Weibull batch-horizon scaling, and a swap-invariant zero-eigh cross-batch conditional fiber operator bank (rgdt-v3, cond-fiber-v1, unified-v11: 2/2 causal isolation PASS, model_ddgan_gaussian_chain PASS).
  4. Law 4 — Function-Space Two-Player Equilibration & Depth/Rail Trust: Parameter-space Adam suffers $\Theta(\alpha L \sqrt{W},\eta)$ function-space impedance mismatch (>600× player asymmetry, -30-logit hostile tilts) and $10^4\times$–$10^8\times$ 16-layer Jacobian skew. Resolved via $10^{-38}$ dynamic scale-normalized peak_rms, Marchenko–Pastur operator scaling, realized generator pushforward ceiling delta_g_max, zero-pinv analytic direction_blend, and fused $Z=[X,Y]$ Quintic Tikhonov polar equilibration (t4-nash-v3 3/3 PASS at 39.68 ms/step, t4-arch-v3 7/7 PASS).
  • Occam's Razor & <900-LOC Core (preset="minimal"): Round-4 Track 2'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 irreducible only on deep skip-free MLPs (grid100_mlp16), while get_recipe("minimal") (minimal-v2, 899 LOC, zero Newton–Schulz) runs r256 at 55.50 ms/step / 1,344.84 MiB.

Honest Governance, Calibration & Failure Disclosures (AGENTS.md)

  • Provisional Screening Selection (default_adoption: false): Forge screening calibration remains BLOCKED under configs/forge/calibration/criteria-v1.json (reports/forge/studies/r3-calibration-holdout.md). DEFAULT_RECIPE = "unified_v11" is an owner-directed research default, not a calibrated superiority claim. Tuning budgets (11 hand-declared unified-v1..v11 cards) are explicitly disclosed in technique-inventory.md.
  • Remaining Numerical Failures (unified-v11): staggered100 (blocking staggered100_14k), grid100_res16_tanh, five_word_joint_hold (23/25 checks, endpoint_passed=True), model_ddgan_bimodal_chain, model_ae_ddgan_gaussian_chain (blocking model_ae_ddgan_bimodal_chain), interp_img_patch_shift, and interp_saddle_barycentric, plus 12 unimplemented STRETCH_VECTOR_HOSTS tasks explicitly marked BLOCKED.

Local Verification Gates

  • python3 -m experiments.forge validate: PASS (662 candidates, 42 family docs valid)
  • python3 -m experiments.forge history --check & compile --check: CURRENT (11,855 files valid)
  • python3 -m experiments.forge.scaling_publication --check: 66/66 clear records
  • python3 ops/refresh_forge_readouts.py --check --reports-only: Status: CURRENT (stale stages: none)
  • python3 scripts/check_lint.py: 0 new findings
  • Historical readouts modified: 0; particlegan/recipe_compat.py untouched
  • Wheel build: particlegan-0.11.0-py3-none-any.whl = 851,669 bytes (< 1,200,000), isolated wheel import verified (DEFAULT_RECIPE == "unified_v11")
  • Local pytest: test_docs_links.py, test_readme_animation.py, test_forge_scaling.py, test_forge_scaling_publication.py (46 passed), plus full Forge/readouts/docs/recipe/formulation/example suites PASS.

@martyn martyn changed the title release(0.11.0): promote unified_v11 default, publish technique inventory, 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 Nov 20, 2026
@martyn
martyn merged commit a40767e into develop Nov 20, 2026
11 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.

2 participants