Roadmap & Boundaries

The project has broad ambition, but the public site must preserve the current evidence boundary.

Current support boundary

The current source-backed implementation supports the TinyRustLM lane:

static browser shell;
no-crate Rust/WASM runtime;
custom .slm format;
deterministic smoke artifacts;
f32/q8_0/q4_0 routes;
model manifests and provenance display;
ADP1/ASP1/ALR1 adapter validation and apply;
selector registry route gates;
module-plan receipt gates;
assembly and adapter-family receipt gates;
browser smoke and route drift evidence.

Not claimed yet

The site should not claim:

general trained assistant quality;
production assistant accuracy;
broad GGUF model import;
remote model hosting;
hidden model download;
telemetry-driven production monitoring;
live autonomous model replication;
Mini Browser orchestration as a completed feature.

Future lanes

Lane	Future direction
GGUF import	Add only after parser, tensor routing, tokenizer, quantization, manifests, and browser smoke proof exist.
Trained models	Add with raw source manifests, provenance, runtime smoke, eval sidecars, and quality gates.
LoRA-style adapters	Extend ADP1/ASP1/ALR1 with trained factor evidence and runtime validation.
Multi-model routing	Introduce after one local runtime path remains stable under tests.
WebGPU	Compare after CPU/WASM correctness and evidence routes are mature.
OPFS and persistence	Add after privacy, storage, checksum, and recovery boundaries are explicit.
Workers	Add after current single-threaded WASM proof remains stable.
Agent integration	Keep the site agent-readable before exposing richer browser tools.

Public language rule

Use “planned,” “future,” “not claimed yet,” and “requires proof” where appropriate. That keeps the project credible and prevents ModelBreeder-style ideas from being mistaken for completed runtime behavior.