Tokenization for traceability. Kova P13 compliant.
| Token | Human name | Module |
| f0 | preprocess | train |
| f1 | train | train |
| f2 | sample | train |
| f3 | sample_medium | train |
| f4 | sample_anvil | train |
| f5 | curate | curate |
| f6 | snap_to_grid | grid |
| f7 | pack_horizontal | sheet |
| f8 | pack_grid | sheet |
| f9 | auto_select | device_cap |
| f10 | reprobe | device_cap |
| f11 | auto_sample | device_cap |
| f12 | best_available | device_cap |
| f13 | probe_cluster | cluster |
| f14 | cluster_generate | cluster |
| f15 | deploy_to_node | cluster |
| f16 | deploy_all | cluster |
| f17 | sync_models | cluster |
| f18 | sync_models_all | cluster |
| f19 | train_combiner | combiner |
| f20 | generate_scene | combiner |
| f21 | load_combiner | combiner |
| f22 | save_combiner | combiner |
| f23 | train_experts | expert_train |
| f24 | route | expert |
| f25 | save_experts | expert |
| f26 | load_experts | expert |
| f27 | cascade_sample | moe |
| f28 | train_judge | judge |
| f29 | load_judge | judge |
| f30 | generate_seeded | scene |
| f31 | generate_bootstrap | scene |
| f32 | encode_batch | scene |
| f33 | run (app) | app |
| f34 | quality_gate | discriminator |
| f35 | generate_with_gate | discriminator |
| f36 | acquire (gpu) | gpu_lock |
| f37 | release (gpu) | gpu_lock |
| f38 | detect_class_count | train |
| f39 | load_palette | palette |
| f40 | quantize | palette |
| f41 | lookup | class_cond |
| f42 | cond_tensors | train |
| f43 | detect_model_version | train |
| f44 | stage_cascade_sample | moe |
| f45 | detail_only_sample | moe |
| f46 | ingest_gemini | main |
| f47 | prep_stages | main |
| f48 | sign | nanosign |
| f49 | verify | nanosign |
| f50 | strip | nanosign |
| f51 | verify_or_bail | nanosign |
| f52 | sign_and_log | nanosign |
| f53 | detect_v_pred | train |
| f54 | seeded_noise | train |
| f55 | eps_to_clean | train |
| f56 | cfg_denoise | train |
| f57 | cfg_denoise_vpred | train |
| f58 | compute_skeletons_v2 | train |
| f59 | load_skeleton_v2 | train |
| f60 | skeleton_start | train |
| f61 | cond_tensors | train |
| f62 | sample_seeded | train |
| f63 | sample_seeded_cfg | train |
| f64 | sample_medium_seeded | train |
| f65 | sample_medium_seeded_cfg | train |
| f66 | sample_anvil | train |
| f67 | anvil_sample | moe |
| f68 | cascade_with_gate | moe |
| f69 | anvil_with_gate | moe |
| f70 | stage_cascade_sample | moe |
| f71 | detail_only_sample | moe |
| f72 | four_dir_sheet | relight |
| f73 | compute_structure | relight |
| f74 | relight | relight |
| f75 | walk_pngs | lib |
| f76 | list_palettes | palette |
| f77 | model_ready | pipeline |
| f78 | download_model | pipeline |
| f79 | generate (pipeline) | pipeline |
| f80 | sigma_from_uniform | precond |
| f81 | sampling_sigmas | precond |
| f82 | EdmCoeffs::at | precond |
| f83 | EdmCoeffs::with_gamma | precond |
| f84 | Normalizer::load | normalize |
| f85 | Normalizer::save_sidecar | normalize |
| f86 | Normalizer::sigma_data | normalize |
| f87 | normalize_stats (CLI) | main |
| f88 | quality_check (CLI) | main |
| f89 | vulkan_tiny::train | vulkan_tiny |
| f90 | vulkan_tiny::load_candle_safetensors | vulkan_tiny |
| f91 | vulkan_tiny::save_candle_safetensors | vulkan_tiny |
| f92 | vulkan_backend::train_silo | vulkan_backend |
| f93 | palette_net::predict | palette_net |
| f94 | tiered_pipeline::generate | tiered_pipeline |
| f95 | class_router::pick_silo | class_router |
| f96 | sponge::retry | sponge |
| Token | Human name | Module |
| t0 | TinyUNet | tiny_unet |
| t1 | MediumUNet | medium_unet |
| t2 | AnvilUNet | anvil_unet |
| t3 | PixelForgeApp | app |
| t4 | TrainConfig | train |
| t5 | PackedDataset | train |
| t6 | Tier | device_cap |
| t7 | DeviceProfile | device_cap |
| t8 | Backend | device_cap |
| t9 | ClusterState | cluster |
| t10 | NodeProfile | cluster |
| t11 | WorkUnit | cluster |
| t12 | SlotGridTransformer | combiner |
| t13 | CombinerTrainConfig | combiner |
| t14 | MicroClassifier | judge |
| t15 | ExpertHead | expert |
| t16 | ExpertSet | expert |
| t17 | ExpertStage | expert |
| t18 | SceneGrid | scene |
| t19 | Biome | scene |
| t20 | SwipeStore | swipe_store |
| t21 | Swipe | swipe_store |
| t22 | SwipeStats | swipe_store |
| t23 | Discriminator | discriminator |
| t24 | DiscriminatorTrainConfig | discriminator |
| t25 | ClassCond | class_cond |
| t26 | ClassConditioner | tiny_unet/medium_unet/anvil_unet |
| t27 | VerifyResult | nanosign |
| t28 | CascadeConfig | moe |
| t29 | WorkUnit | cluster |
| t30 | NodeDef | cluster |
| t31 | LightDir | relight |
| t32 | EdmCoeffs | precond |
| t33 | Normalizer | normalize |
| t34 | VulkanTinyParams | vulkan_tiny |
| t35 | VulkanTinyTrainConfig | vulkan_tiny |
| t36 | MicroUNet | micro_unet |
| t37 | PaletteNet | palette_net |
| t38 | TieredConfig | tiered_pipeline |
| t39 | SiloRouter | class_router |
| Token | Human name | File | Params | Size |
| m0 | Cinder | pixel-forge-cinder.safetensors | 1.09M | 4.2MB |
| m1 | Quench | pixel-forge-quench.safetensors | 5.83M | 22MB |
| m2 | Anvil | pixel-forge-anvil.safetensors | 16M | ~64MB |
| Token | Module | Purpose |
| M0 | tiny_unet | Cinder model — channels [32,64,64] |
| M1 | medium_unet | Quench model — channels [64,128,128], self-attention |
| M2 | anvil_unet | Anvil model — channels [96,192,192], XL |
| M3 | train | Training loop, sampling, data pipeline |
| M4 | app | egui GUI — mobile + desktop |
| M5 | device_cap | Device detection, tier selection, benchmarks |
| M6 | cluster | Distributed generation across SSH nodes |
| M7 | combiner | SlotGridTransformer — scene composition |
| M8 | judge | MicroClassifier — quality scoring |
| M9 | expert | MoE expert heads — shape/color/detail/class |
| M10 | expert_train | Expert training on frozen Quench |
| M11 | moe | Cascade pipeline — Cinder → Quench + Experts |
| M12 | scene | 8x8 SceneGrid, biome generation |
| M13 | swipe_store | Tinder-style swipe data, ring buffer |
| M14 | lora | Rank-4 LoRA adapters for TinyUNet |
| M15 | discriminator | Quality gate — binary classifier |
| M16 | palette | Color palettes + quantization |
| M17 | grid | Snap to grid, tile alignment |
| M18 | sheet | Sprite sheet packing |
| M19 | curate | Data curation pipeline |
| M20 | gpu_lock | File-based GPU lock for training |
| M21 | plugin | Kova plugin interface |
| M22 | poa | Proof of Authorship signing |
| M23 | pipeline | SD pipeline (optional, desktop) |
| M24 | class_cond | Hybrid class conditioning (super-cat + tags) |
| M25 | relight | 4-directional sprites from SDF + normals |
| M26 | quantize | f32 → f16 model quantization |
| M27 | nanosign | BLAKE3 model integrity signing |
| M28 | relight | 4-directional sprites from SDF + normals |
| M29 | precond | EDM preconditioning math (Karras 2022) + Min-SNR-γ |
| M30 | normalize | Per-channel z-score sidecar manifest (mandatory) |
| M31 | vulkan_backend | any-gpu MicroUNet (silo) training path |
| M32 | vulkan_tiny | any-gpu TinyUNet (Cinder) training path |
| M33 | palette_net | ~100K MLP — class → palette colors |
| M34 | tiered_pipeline | silo → PaletteNet → Cinder-detail pyramid |
| M35 | micro_unet | ~97K-param per-class silo (shape specialist) |
| M36 | class_router | Reads class_config.toml; picks per-class silo |
| M37 | sponge | Sponge Mesh: auto-retry on NaN/plateau (max 3) |
| Token | Command | Description |
| c0 | (none) | Launch GUI app |
| c1 | sprite | Generate sprite from text |
| c2 | tileset | Generate seamless tileset |
| c3 | train | Train a model |
| c4 | generate | Generate from trained model |
| c5 | curate | Curate training data |
| c6 | swipe | Record swipe feedback |
| c7 | train-judge | Train quality judge |
| c8 | judge | Score a sprite |
| c9 | scene | Generate a scene |
| c10 | train-combiner | Train scene combiner |
| c11 | train-lora | Train LoRA adapters |
| c12 | pipeline | Run SD pipeline |
| c13 | cascade | MoE cascade generation |
| c14 | train-experts | Train expert heads |
| c15 | auto | Auto-detect and generate |
| c16 | probe | Probe device capabilities |
| c17 | cluster-deploy | Deploy to cluster nodes |
| c18 | cluster-sync | Sync models to cluster |
| c19 | stage-cascade | Cinder-sil → structure → Quench-detail |
| c20 | ingest-gemini | Slice Gemini sprite sheets into tiles |
| c21 | prep-stages | Extract structure maps for conditioning |
| c22 | relight | 4-directional sprite sheet from SDF |
| c23 | skeletons | Precompute per-class skeleton images |
| c24 | normalize-stats | Compute per-channel mean/std → sidecar manifest |
| c25 | quality-check | Per-image stats + cross-image pairwise diversity |
| c26 | train-cinder-vk | Train Cinder/TinyUNet on Vulkan via any-gpu (AMD-friendly) |
| c27 | train-silo | Train per-class MicroUNet silo (~97K params) |
| c28 | train-fleet | Multi-node distributed training across IRONHIVE |
| c29 | tiered | Tiered generation: silo → PaletteNet → Cinder-detail |
| c30 | train-palette-net | Train PaletteNet (~100K MLP) on K-means class palettes |
| c31 | prep-silo-cond | Coarsen training images (32→16→32) as conditioning hints |
Replaced integer class IDs with hybrid super-category + binary tags.
See src/class_cond.rs for full taxonomy (108 class dirs mapped).
| ID | Name | Examples |
| 0 | humanoid | character, elf, knight, hero_* |
| 1 | creature | animal, bird, cat_*, slime |
| 2 | monster | enemy, dragon, demon, alien |
| 3 | weapon | sword, gun, staff_wand |
| 4 | equipment | armor, helmet, shield, key |
| 5 | consumable | potion, food, gem_treasure |
| 6 | terrain | terrain, floor, wall, water_lava |
| 7 | nature | tree, tree_*, bush_flower |
| 8 | structure | building, furniture_, vehicle_ |
| 9 | ui_fx | ui, effect, fx_*, misc |
| 10 | (null/CFG) | unconditional |
[alive, humanoid, held_item, worn, nature, built, magical, tech, small, hostile, edible, ui]
| ID | Name |
| 0 | stardew |
| 1 | starbound |
| 2 | snes |
| 3 | nes |
| 4 | gameboy |
| 5 | pico8 |
| 6 | endesga |
▸ THE COCHRAN BLOCK, LLC · CAGE 1CQ66 · UEI W7X3HAQL9CF9 · UNLICENSE · cochranblock.org