reflexmesh

For actual user files, run reflexmesh workflow recipe.json --workspace /absolute/directory, or start the local workbench with --workspace. The workspace workflow contract covers registered processes, dependency verification, previews, cancellation and native receipts.

Released trained checkpoints are separate downloadable assets: reflexmesh fetch-models --destination models/checkpoints --version v0.01.000 verifies the archive and every model digest before replacing files. Use only project release artifacts: some trained framework formats require trusted model files. The scientific pipeline remains available to rebuild the checkpoints from owned data.

An event-driven software controller with a Rust authority core, Python tool adapters, learned policies and asynchronous language-model coordination. The library is the engine behind Neuraxis.

The controller maintains exact capabilities and resource versions separately from learned recurrent state. A policy proposes a bounded action or requests deliberation. The executor revalidates authority and state before dispatch, records effects durably, and preserves unknown outcomes for reconciliation. Model confidence never grants permission.

Install and execute

Python 3.11+ is required. Source builds require Rust and platform C/C++ build tools. Training and the local workbench server are optional dependencies.

python -m pip install '.[train,server,dev]'
reflexmesh doctor
mkdir workspace
reflexmesh workflow examples/verified-workflow.json --workspace workspace

The example performs a real file write and checks its SHA-256 postcondition through the native broker. It does not call an LLM. Use the Runtime SDK to register files, fixed process templates, explicit HTTP endpoints and application-specific tools. Use Controller to submit dependent goals, subscribe to events and coordinate a slow planner.

from pathlib import Path
from reflexmesh.effectors import WorkspaceFiles
from reflexmesh.runtime import Runtime

root = Path('workspace')
root.mkdir(exist_ok=True)
with Runtime(root, ('file.read', 'file.write')) as runtime:
    WorkspaceFiles(root).register(runtime)
    candidate = runtime.candidate('file.write', {'path': 'answer.txt', 'content': '42'})
    receipt = runtime.execute_candidate(candidate)
    print(receipt['status'])

Scientific pipeline

The owned workload suite contains 20 independent software-effect families, six fault variants and disjoint train/validation/calibration/test environments. It executes disposable real files, local HTTP services and child processes with independent postcondition checks. Twelve policy methods cover FSM, behavior tree, calibrated logistic/XGBoost scoring, GRU behavior cloning, LinUCB, masked PPO, learned-transition lookahead, direct LLM, asynchronous FSM+LLM, learned deferral and recurrent metacontrol.

python -m reflexmesh.pipeline collect --artifacts artifacts
python -m reflexmesh.pipeline train --artifacts artifacts
python -m reflexmesh.pipeline ppo --artifacts artifacts
python -m reflexmesh.pipeline gates --artifacts artifacts --model qwen3.5:4b
python -m reflexmesh.pipeline evaluate --artifacts artifacts --seeds 10

Model-dependent methods require a real configured local Ollama model. No unavailable planner is silently replaced with a rule policy. Raw traces, fitted checkpoints, lineage digests and aggregate statistics remain separate. External benchmarks have independent code/model/data license and environment provenance. Scores on designed workloads do not establish production or AGI performance.

Workbench service

reflexmesh serve --checkpoints artifacts/checkpoints --artifacts artifacts/evaluation --model qwen3.5:4b

The service binds loopback and prints a session token. Every executable endpoint requires that token and the same origin. The service executes registered disposable workloads; public static replays have no host-tool authority.

Evidence and limits

Fast/slow routing, recurrent policies, predictive control and learned deferral are prior art. The research hypothesis concerns outcome-grounded deliberation allocation under asynchronous state change, resource-versioned proposals and persistent effect residuals. It requires matched held-out comparisons and component ablations. Implementation alone does not establish novelty or superiority.

This is a process-level coordination library, not an operating-system sandbox or a hard real-time system. A cancelled request does not imply rollback or cancellation of a remote provider. Hostile native executables require independent isolation. See runtime contracts, native architecture, contributing and security.

Authored source is Apache-2.0. External models, data and benchmarks retain their own licenses.