epistemic-edge

Subjective-logic guardrails for LLM-driven IoT actuation — an air-gapped, neuro-symbolic AIoT framework that places calibrated uncertainty quantification between your sensors and any LLM that can touch the physical world.

Companion package for:

Epistemic Edge: Subjective Logic Guardrails for LLM-Driven IoT Actuation. Muntaser Syed and Marius Silaghi. IEEE 27th International Conference on Information Reuse and Integration for Data Science (IEEE IRI 2026), Seattle, WA. To appear.

Full reproduction materials — experiment scripts, immutable result sets, and the pinned inference fork — live in the repository: https://github.com/jemsbhai/epistemic-edge

Architecture

Epistemic Edge orchestrates four tiers into a strict verify–decay–generate pipeline:

Locally deployed LLMs can act on conflicting or stale context. Epistemic Edge annotates every source with a calibrated Subjective Logic opinion, fuses them via Jøsang cumulative fusion, decays stale beliefs toward uncertainty, and verifies epistemic thresholds before any actuation is allowed. In the paper's ablation (seven locally-deployed LLMs, 2,800 controlled trials), removing this epistemic layer collapses threshold-guardrail accuracy from 1.00 to 0.40 for every model. On the BATADAL water-distribution benchmark, calibrated SL opinions reach AUROC 0.9004 with no trained weights and no labeled attack data.

Installation

pip install epistemic-edge                # core: transport + trust + memory
pip install "epistemic-edge[llm]"         # + local LLM inference (llama-cpp-python)
pip install "epistemic-edge[transport]"   # + MQTT / CoAP transports
pip install "epistemic-edge[all]"         # everything

Quick Start

import asyncio
from epistemic_edge import EdgeNode
from epistemic_edge.memory import DecayConfig

async def main():
    node = EdgeNode(
        node_id="gateway_alpha",
        llm_path="./models/bonsai-8b-1bit.gguf",
        decay=DecayConfig(mean_reversion_rate=1.5, threshold=0.2),
    )

    @node.guardrail(action="close_valve")
    def check_safety(state, intent):
        return state.max_uncertainty() < 0.15

    @node.on_actuate
    async def execute(intent, receipt):
        print(f"Executing: {intent.action} on {intent.target}")
        print(f"Audit trail: {receipt}")

    await node.start()

asyncio.run(main())

Citing

@inproceedings{syed2026epistemic,
  author    = {Syed, Muntaser and Silaghi, Marius},
  title     = {Epistemic Edge: Subjective Logic Guardrails for {LLM}-Driven {IoT} Actuation},
  booktitle = {2026 IEEE 27th International Conference on Information Reuse and
               Integration for Data Science (IRI)},
  year      = {2026},
  note      = {To appear}
}

Core Libraries

• jsonld-ex — JSON-LD 1.2 extensions with Subjective Logic, FHIR R4, PROV-O
• cbor-ld-ex — Compact Binary Linked Data for constrained IoT networks
• chronofy — Temporal validity framework implementing TLDA

License

MIT