System Changelog.

Rewrote Inverse Kinematics solver in CUDA to use Damped Least Squares (DLS). Completely prevents singular matrix crashes during hyper-parallel domain randomization.

Deprecated synchronous `iacon.env.step()`. Replaced entirely with asynchronous `iacon.physics.step_parallel()` for batched GPU execution.

TensorRT edge compilation latency reduced by 14%. KV-Cache memory footprint reduced via aggressive INT8 quantization flags.

Implemented deterministic fallback controllers. If neural policy outputs a torque command exceeding thermal limits, the hardware bus instantly reverts to a rigid PD loop.

Added native support for rendering synthetic RGB-D camera feeds directly inside the simulation engine via Vulkan.

Initial migration from State-Space Models to Causal Transformers for the Large Behavior Model (LBM). Added infinite horizon memory buffers.

Initial release of the `iacon pull` CLI tool for fetching pre-trained baselines.

Full integration with MuJoCo 3.0.0 bindings. Improved contact-patch dynamics for multi-toed end effectors.

Added `iacon-humanoid-v2` dataset containing 4.2M expert rollout tokens.

Enabled multi-GPU distribution via NCCL. Synchronized gradient collection across 8x H100 nodes.

First public preview of the Python SDK with JIT-compiled reward functions.

Introduced Hyper-Parallel Simulation core. Support for 4096+ environments on single-node V100/A100 clusters.

Added hardware-level encryption for model weight transfers across nodes.

Initial public release of the Iacon OS (Arkenos). Basic URDF ingestion and rigid-body dynamic solvers.

Implementation of the immutable experiment ledger and hashed policy versioning.