# Parametry — binding SLeeLa models to known or local sources **Parametry** is the arrangement that binds a model — a **SLeeLa LLM** or a **SLeeLa neural net** — to a concrete parameter set, resolving **either** a known/supported model (by name, from a registry) **or** a user's local model (by path). It is the glue that lets the rest of the subsystem size itself to a model without SLeeLa shipping any weights. ## The either/or A binding (`sllmp_binding`) names a `kind` (`llm` or `neural-net`) and gives a source: | Source | How it resolves | |---|---| | **known** (`name`) | looked up in the built-in registry; parameters come from that entry (caller may override) | | **local** (`local_path`) | the caller's declared parameters are used as-is (user-owned model) | Resolution (`sllmp_resolve`) fails closed: an unknown name with **no** local fallback -> `UNKNOWN_MODEL`; neither name nor path -> `NO_SOURCE`; a name whose registry kind differs from the binding -> `KIND_MISMATCH`; a local neural net with no layer shape -> `INVALID_ARGUMENT`. An unknown name **with** a local path falls back to the local model. ## Known models (registry) The built-in registry lists supported model **names** and their declared parameters only — **no weights**. Binding a known name still means the user supplies that family's weights. Mirror file: `known-models.json`. | Name | Kind | Parameters | |---|---|---| | `sleela-code-compact` | llm | 2 GB, ctx 2048 | | `sleela-code-standard` | llm | 4 GB, ctx 4096 | | `sleela-code-capable` | llm | 8 GB, ctx 8192 | | `sleela-code-large` | llm | 16 GB, ctx 16384 | | `sleela-nn-xor` | neural-net | shape 2-4-1 | | `sleela-nn-small-mlp` | neural-net | shape 16-32-16-4 | ## What each binding feeds - **LLM** -> a `CapacityProfile` (2–16 GB envelope) that sizes the reasoning budget (context/steps/backtracks), and pairs with the SLeeLa Code LLM (`sleela/`) and `SLLMR_SUBJECT_CODE`. - **Neural net** -> an `slnn_network_config` (the real `neural-nets/` API): the resolved layer shape builds an actual fully-connected net you can `predict` and `train_sample` on. ## Usage (C++) ```cpp using namespace sleela::ai; // A SLeeLa LLM bound to a KNOWN model by name: Parametry a = Parametry::known(SLLMP_KIND_LLM, "sleela-code-capable"); CapacityProfile cap = a.capacity(); // 8 GB -> capable tier // ... or to a LOCAL model by path: Parametry b = Parametry::local_llm("/models/my-sleela-coder.gguf", 6ULL*1024*1024*1024, 0, "gguf-q4_k_m"); // A SLeeLa neural net bound to a KNOWN shape, or a LOCAL one: Parametry c = Parametry::known(SLLMP_KIND_NEURAL_NET, "sleela-nn-xor"); Parametry d = Parametry::local_neural_net("/models/my-net.slnn", {8, 12, 3}); std::vector shape = d.layer_sizes(); // -> slnn_network_config ``` ## Files | File | Role | |---|---| | `parametry.h/.c` | C11 registry + resolver (`sllmp_*`, ABI v1) | | `parametry.hpp` | C++17 `Parametry` (known/local; -> CapacityProfile or layer shape) | | `known-models.json` | Declarative registry mirror | | `binding.schema.json` + `bindings.example.jsonl` | User binding records + examples | | `parametry_selftest.c` | Resolution, overrides, fail-closed | | `parametry_demo.c` | Binds an LLM and a neural net (known + local); builds a real `slnn_*` net | ## What is NOT bundled No weights, no model files. Parametry ships the registry of supported **names** and their declared parameters, and the binding logic. The user supplies the actual model — a known family's weights, or a local model by path.