NeuroBlocks

Definition

A NeuroBlock ships a neural function (for example collision loom response) together with identity: name, version, interface ports, and digests over the locked model or source bytes. Consumers depend on that identity, not on a notebook sitting on someone’s laptop.

What a block declares

• Inputs — named channels with types (vision events, heading deltas, …).
• Outputs — named channels the rest of the graph may consume.
• Runtime expectations — rate envelopes and deadlines the stack planner must respect.
• Provenance — where the module came from (connectome-derived, engineered, synthetic, …).

        inputs                  NeuroBlock                 outputs
   ┌────────────────┐      ┌──────────────────┐      ┌────────────────┐
   │ EventVision    │─────▶│  LoomGuard@1.0   │─────▶│ avoidance_vec  │
   │ …              │      │  digest · lock   │      │ trigger        │
   └────────────────┘      └──────────────────┘      └────────────────┘

Lifecycle

1. Author or derive the module and package it with an explicit version.
2. Publish or bundle it so Discover can list interface and provenance.
3. Compose locks the exact bytes into a stack lockfile.
4. Runtime loads only locked modules — no silent substitution.