Llebre turns operator intent into a GPT-5.6 architectural report, a typed Network Digital Twin and vendor-aware CLI - then independently verifies every artifact in Rust/WASM before it can reach the terminal.
No installation. No device credentials. No automatic deployment.
The hard problem is not generating commands. It is knowing when not to release them.
Network changes are stateful, vendor-specific and operationally risky. Llebre makes the model part of a visible engineering system instead of treating its answer as authority.
GPT‑5.6 Sol receives bounded operator intent and a selected vendor profile, then returns strict structured output: an architectural report, a typed Digital Twin and a candidate configuration artifact.
A browser-native Rust/WASM simulator checks topology integrity, addressing, subnet overlap, Layer 2 cycles, VLAN continuity and static-route completeness. A separate command gate evaluates the candidate CLI.
Intent, vendor, revisions, findings, CLI and change evidence are tied to the same SHA‑256 fingerprint. Change the intent or vendor and every downstream artifact becomes stale atomically.
A valid model response is not the same thing as verified CLI. Llebre preserves separate provenance for reasoning, topology simulation, command policy and the final review artifact.
Treated as untrusted input and bounded before the request reaches the model.
Performs contextual architecture reasoning and vendor-aware candidate generation.
Converts the proposal into a runtime-validated, inspectable network model.
Simulate topology and independently reject unsafe or unsupported CLI.
Release only when every artifact is current, consistent and verified.
The architecture, topology, candidate revision, deterministic findings, CLI gate and change evidence are visible surfaces - not hidden prompt machinery.
Click-first and keyboard-first navigation, local workspace state and an inspectable terminal release boundary.
Immutable accepted revisions, recoverable drafts, semantic patches and exact base fingerprints prevent prompt drift from silently rewriting history.
Entity-linked findings make every duplicate address, overlap, broken VLAN path or Layer 2 cycle inspectable.
Stable JSON and Markdown exports bind semantic delta, checks, risk, rollback posture and provenance into one review record.
The fastest path exposes both sides of the project: GPT‑5.6 reasoning and an independent local safety boundary.
Strict TypeScript, React, IndexedDB revisions, GPT‑5.6 Structured Outputs, Rust/WASM, native evaluation and a protected server-side gateway.
A guided entry, click and keyboard parity, visual topology, evidence surfaces, deterministic status and a live testable demo.
Network engineers can use generative reasoning without silently accepting generated commands as operational truth.
Llebre combines an AI architecture layer with typed network state and deterministic release gates in a browser-native Network IDE.
The hackathon work established the canonical model contract, persistence and revision boundaries, topology workbench, Rust/WASM ABI, vendor compiler contracts, immutable change packages, evaluation harness and security foundations.
When the hackathon ends and the remaining core capabilities are implemented, Llebre will be released as Open Source. Network engineers will be able to inspect the model, extend the vendor layer and audit the safety boundary together.
The project produces reviewable architecture, candidate CLI and deterministic change evidence. It does not store credentials, execute configuration, claim rollback execution or bypass the operator.
Open the live workspace, run the Safe Campus path and inspect exactly where GPT‑5.6 ends and deterministic verification begins.
