Winning formulation

Product repositories import one stamp from particle-sliders-core:

from particle_sliders import winning_formulation

stamp = winning_formulation()
stamp.require(stamp.as_dict())

winning_formulation() is the only product entry point. It is gmix architecture plus the current formulation overlay.

Depends on ParticleGAN develop. Pinning particle-sliders-core pulls particlegan (develop SHA in that package's pyproject.toml). Products call stamp.regularizer() / stamp.losses(); they do not vendor ParticleGAN excerpts. Optional helpers: particlegan_get_recipe(), particlegan_locked_shared() (make_b_cap / make_gan_loss).

Architecture: gmix. Routed particles and a global-mix critic (gmix_architecture()). This is the product game structure. It does not wait on ParticleGAN #38.

Formulation: the #38 winner, provisional until that search finishes. Caps, coefficients, learning rates, particle counts, schedules, and a later upsampler choice such as residual16 versus transpose come from ParticleGAN #38. The related search is ParticleGAN #39. The ultimate gate is the full live leaderboard: 9 trained toys and all 29 live bounds. A partial win, a 3-toy screen, or an EMA-only result does not count. stamp.formulation_provisional is true until that crown is plugged into the gmix architecture.

Products do not keep a second copy of the routed particle adapter, the global-mix critic, the paired-error losses, the noise schedule, locked_shared, or GradRegularizer. Cap / GAN loss / particle VIC are ParticleGAN develop API calls through this core (GradientPenalty, GANLoss, ParticleRegularizer); GradRegularizer is only a compatibility alias.

Architecture

Id	gmix (stamp.architecture_id)
Family	particle-gmix
Adapter	Routed particle MLP, rank 8, width 48, router width 16
Critic	Global-mix, 8 tokens, width 48, 1 layer, 4 heads, score bound 8
Game	Paired-error relativistic logistic GAN plus particle VIC

Current provisional formulation

Id	particle-gmix-1600-v2 (stamp.formulation_id)
Callable	particle_gmix_1600_v2()
Hub recipe name	anneal-routed-particle-error-yue2-v1
Particles	128 of dimension 4
Cap	b_cap coefficient 1, kappa 1, every 4 steps
Noise	Paired-edit whitening, start edit_rms / 0.28, decay over 1600 steps, hold 1.3 * edit_rms
Auxiliary MSE / FM / cover / pole / lyric hold	All zero

These parameters are the released Hub record (analysis/slider2d/yue2_gmix_v2_exam.V2_SPEC, UNI gates in merged #128, Anima's published particle release). They are the overlay until #38 crowns a full-board winner. They are not a second architecture.

V2_SPEC["propose_only"] is true so the 2D exam cannot flip the Music trainer's --lm_target. Products still call winning_formulation().

Named callables

Callable	Role
gmix_architecture() / gmix_recipe()	Fixed gmix architecture constants.
particle_gmix_1600_v2()	Provisional formulation overlay. Current CURRENT_FORMULATION.
locked_shared_recipe()	Endpoint recipe (cover and pole MSE, VIC and noise off, teacher faithful_guard_e). Not the product architecture.
winning_formulation()	Gmix architecture with CURRENT_FORMULATION applied.

When #38 crowns a winner

In packages/particle-sliders-core/src/particle_sliders/formulation.py, replace the formulation block. Leave gmix_architecture() in place:

CURRENT_FORMULATION_ID = "<crowned id>"
CURRENT_FORMULATION = <crowned_formulation>
CURRENT_FORMULATION_PROVISIONAL = False

Add the crowned parameter export in that file if it is not already particle_gmix_1600_v2. The overlay must not restate gmix architecture keys (critic stays gmix). Caps, coefficients, learning rates, particle counts, schedules, and an upsampler choice such as residual16 versus transpose belong in the overlay. Leave every product call site on winning_formulation(). Say so in the commit. Product repos then bump:

particle-sliders-core @ git+https://github.com/HyperGAN/particle-sliders.git@<new-commit>#subdirectory=packages/particle-sliders-core

A product does not fork the math while waiting for that bump. Model surfaces listed on stamp.model_surface_keys (learning rates, batch, data budget) may differ per backbone. Hub ids, Comfy class names, prompt cards, hooks, and sampling stay in the product.

What Krea2's self-contained rule gets wrong

krea2-particle-sliders currently treats a particle-sliders checkout as forbidden. Its README says a checkout is not required, and tests/test_docs_contract.py fails if PARTICLE_SLIDERS_ROOT appears in the README or REPRODUCE.md. Its FORMULATION.md says a winning recipe comes back into that repo as a local config.

That is not the long-term pattern. Krea2 should pin this package the same way Anima and Supra do, and call winning_formulation(). An ad-hoc PARTICLE_SLIDERS_ROOT path is a worse install than the git subdirectory pin, and banning the dependency entirely leaves the product re-implementing the game. Formulation toys still belong in HyperGAN/conceptmod, not in the product repo and not in ParticleGAN.

Product follow-ups

These land in the product repositories. This pull request does not rewrite their trainers.

1. anima-particle-sliders. Replace the pin concept-slider-core @ git+https://github.com/mikkel/sliders-conceptmod.git@beaffeb#subdirectory=packages/concept-slider-core with particle-sliders-core from HyperGAN/particle-sliders, subdirectory packages/particle-sliders-core, at the commit that introduces this stamp. Import particle_sliders.winning_formulation and call require() on the training config. Stop vendoring the reference implementation as the source of truth (core.lock.json still hashes the old concept_slider_core files). Backbone g_lr 2e-5 is a model surface. The published hold at 1.0 disagrees with this stamp's noise_hold_ratio 1.3; resolve that by training the stamp or by changing the stamp here, not by keeping a local schedule.
2. krea2-particle-sliders. Add the same git subdirectory dependency. Train through winning_formulation() (gmix architecture; today's provisional parameters are particle-gmix-1600-v2). #38 replaces the overlay, not the architecture. Delete the "checkout not required / do not depend on particle-sliders" contract, including the PARTICLE_SLIDERS_ROOT absence checks. Keep the Hub id jimmycarter/krea2-turbo-bbox, the Comfy node, the turbo sample numbers, and the prompt cards in that repo.
3. supra-concept-sliders. The opt-in trainer still lives in this repository as conceptmod/textsliders/train_lora_supra.py. Move that product entrypoint into the supra repo and depend on this core for the game. Keep Supra2-IMG ids, the Euler sample card, and Comfy (when it exists) in the product repo. Do not copy RoutedMLP or GradRegularizer into the move.