Truth hierarchy

MEL's evidence model is explicit about source authority. Lower layers never override higher ones.

  1. ①
    Runtime evidencecanonical

    Typed deterministic records: ingest records, state transitions, audit events. Always wins on conflict.

  2. ②
    Bounded calculatorsdeterministic

    Heuristics and derived values with explicit inputs and documented uncertainty bounds.

  3. ③
    Assistive inferencelabeled

    AI assistance labeled non-canonical when present. Never overrides deterministic layers.

  4. ④
    Narrative and UI copyfollows

    Must not contradict layers above. Narrative follows evidence — not the reverse.

Full detail: Architecture overview (docs) · Architecture truth (product)

Component map

High-level runtime structure. Each layer has explicit boundaries and documented ingest constraints.

entry point

CLI / Daemon

cmd/mel — init, doctor, serve, demo, operator commands. Single binary entry point.

persistence

SQLite Store

Local operational database. All evidence, incidents, and audit events. Freshness is explicit.

transport

Ingest Workers

Serial/TCP/MQTT paths. Unsupported paths (BLE, HTTP ingest) stay labeled — no implied support.

api layer

Runtime API

REST API serves embedded console and external clients. /healthz, /readyz, /api/v1/status.

surface

Embedded Console

React+Vite UI compiled into internal/web/assets/. Served by mel serve — not this site.

Folder-level map for contributors: ARCHITECTURE_MAP.md

Transport and control boundaries

MEL observes what ingest and configuration make visible. It does not implement mesh RF routing as a product feature, and it does not collapse degraded or unknown posture into “all green.”

Control paths separate submission, approval, dispatch, execution result, and audit — each is a distinct state with attribution. See Trust & privacy and the docs on control-plane trust.

Operational boundaries

  • MEL is not the RF routing stack — it persists evidence from what your ingest workers connect.
  • MEL does not claim RF coverage, mesh delivery, or propagation proof.
  • BLE ingest and HTTP ingest are unsupported — labels remain in UI and docs.
  • Assistive AI output is non-canonical — deterministic layers always win conflicts.
shell — open deep docs
git clone https://github.com/Hardonian/MEL-MeshEdgeLayer.git
cd MEL-MeshEdgeLayer && ls docs/architecture/