Virtual Kernel — Executable Surrogate of OS Kernel Dynamics

Virtual Kernel is an executable, branchable surrogate model of OS kernel dynamics under configuration interventions. Rather than acting as a static point predictor, it provides an interactive simulation environment learned from intervention–response telemetry, enabling counterfactual rollouts, uncertainty estimation, and grounded planning without live kernel disruption.

Core Capabilities

• Latent Dynamics & Ensembles: State-transition models F(z, a) = z′ learned from intervention–response evidence with quadratic action skips to capture non-linear latency curves, backed by deep ensembles for calibrated uncertainty.
• Branchable Environment API: Fork counterfactual worlds with copy-on-write semantics (observe, intervene, branch, rollout, measure, uncertainty, compare).
• Budgeted Search & Planning: Beam screening for rapid breadth exploration and uncertainty-penalized UCT MCTS for robust sequential policy optimization under fixed oracle budgets.
• Multi-Layer Trust & Doubt: Continuous validation using conservation invariants, manifold OOD novelty detection, task-dependent fidelity metrics, and adversarial agent debate.
• Strict Safety Authority: The deterministic PolicyValidator is the sole actuation authority. The surrogate proposes and explores counterfactual branches, but never writes unverified parameters to a live kernel.

Repository Structure

Directory / File	Description
kernel/	OS policy layer: KIR (Kernel Intermediate Representation) schemas, Policy ABI, deterministic PolicyValidator, reference KernelSimulator, offline RL trainer, and runtime actuator.
virtual_kernel/	Surrogate model layer: datasets, latent transition dynamics, ensemble models, branchable environment, beam/MCTS planning, truth oracle, metrology, and active probe loops.
benchmarks/	Rigorous evaluation suites: certification (run_levels), oracle ladder comparisons (run_ladder), degradation tests, zero-shot splits, and compositional tasks.
experiments/	Step-by-step experimental pipeline (phase1_predictor through phase6_closed_loop).
tests/	Comprehensive test suite covering kernel mechanics, schemas, dynamics, environment branching, and trust metrics.
demo_virtual_kernel.py	Self-contained end-to-end demonstration of learning, branching, screening, debate, and validation.

Quickstart

1. Installation

pip install pydantic numpy pytest

2. Run Tests

python -m pytest tests -q

3. Run Demo

python demo_virtual_kernel.py

4. Run Benchmarks & Experiments

# Evaluate certified surrogate levels (L1–L5)
python benchmarks/run_levels.py

# Run closed-loop active learning
python experiments/phase6_closed_loop.py

# Run oracle ladder comparison
python benchmarks/run_ladder.py

(Append --full to any benchmark or experiment script for exhaustive evaluations).