We are sharpening how we describe what we build. Maple is an accountable AI agent runtime — the durable authority layer that sits between any model and the real world. Not an operating system that replaces your host, not a model lab, and not a blockchain. A runtime that makes autonomous action safe to operate and trivial to audit.
The core loop is unchanged and uncompromising:
LLM proposes · Guard validates · Kernel authorizes · Runtime executes · WorldLine records · Receipts prove · Evals judge · Evolution promotes or rolls back · CreditChain optionally anchors, verifies, settles, and rewards.
Local-first inference with the Hermon engine
Maple no longer needs a cloud key to think. The Hermon engine is Maple's owned inference path, in two tiers:
- Hermon embedded — a pure-Rust, zero-dependency engine compiled into the runtime. It is always available, fully offline, and acts as the floor so there is never "no engine."
- Hermon server — a high-performance GGUF server for frontier-class open models, reached over a local OpenAI/Ollama-compatible API.
The router prefers external accelerators when present — Hermon server → llama.cpp → Ollama — and falls back to the embedded engine otherwise. Today we run Qwen2.5-7B-Instruct (Q5_K_M GGUF) locally as the default production model. The LLM only proposes: it drafts a typed action, Guard validates it, and the fail-closed Kernel authorizes it. If a model asks for something it shouldn't — wallet access, an unscoped tool — it is simply rejected, and the deterministic planner takes over. A bad draft triggers exactly one repair re-prompt before fallback. LLM proposes, kernel disposes — end to end, on your own hardware.
Every action carries a proof
Consequential actions don't just happen — they are sealed. Each one produces a signed, hash-chained receipt in the WorldLine. Receipts embed the producing node's public key, so any other node can independently re-execute and verify the work with no shared secret. Tampering is detected; forged attestations are ignored.
Proof of Contribution
Distributed AI work is only trustworthy if rewards follow verified contribution. Maple's Proof of Contribution (PoC) settles rewards only when an independent reputation-weighted quorum of verifiers agrees with a re-execution of the work. Honest verifiers gain reputation; rubber-stampers and false-failers are slashed. The rule is blunt: no verified contribution, no reward.
Blockchain-capable, not blockchain-captive
Maple boots and runs fully local. Trust backends — Local, Enterprise, Cooperative, or CreditChain — are pluggable; the chain is a feature-gated add-on for anchoring, settlement, and rewards, never a requirement. A local build links zero chain code.
Where this is going
Models will keep commoditizing. Governed, provable execution will not. That is the layer Maple owns — and it now runs on your laptop, offline, with a receipt for everything it does.