Deploy Graph Composition

How deploy graphs become the interface between infrastructure (primals) and products (gardens) — BYOB TOML, topological sort, and session-as-primal.

The gen4 Transition

In gen3, deploy graphs were test fixtures — static configurations that proved primals could compose. In gen4, deploy graphs become the product interface: the mechanism by which products ( helixVision, initioChem, esotericWebb) consume primals without source coupling.

BYOB Schema

Products declare their primal requirements in a TOML deploy graph:

[graph]
name = "helix-vision-composition"
version = "0.1.0"

[[graph.node]]
name = "barracuda"
binary = "barracuda"
order = 1
capabilities = ["tensor.multiply", "tensor.fft", "tensor.eigendecompose"]

[[graph.node]]
name = "toadstool"
binary = "toadstool"
order = 2
depends_on = ["barracuda"]
capabilities = ["dispatch.submit", "dispatch.status"]

[[graph.node]]
name = "coralreef"
binary = "coralreef"
order = 2
depends_on = ["barracuda"]
capabilities = ["shader.compile", "shader.validate"]

[[graph.node]]
name = "nestgate"
binary = "nestgate"
order = 3
depends_on = ["toadstool"]
capabilities = ["cas.store", "cas.retrieve", "cas.verify"]

The product’s launcher reads this TOML, performs a topological sort on dependencies, spawns primals in order, and injects a PrimalBridge for IPC. The product never imports primal code — it communicates via JSON-RPC over TCP sockets.

Sovereignty Guarantee

The deploy graph enforces a critical boundary:

  • No shared crates — products do not depend on primal source code
  • IPC-only interface — all communication is JSON-RPC over TCP
  • Independent evolution — primals can upgrade without product recompilation
  • Binary-only consumption — products consume ecoBin binaries, not source

This means a product can be written in any language, by any team, consuming primals it did not build. The deploy graph is the contract. JSON-RPC is the protocol.

Session-as-Primal

In interactive products (games, creative tools), sessions themselves become ephemeral biomeOS citizens:

This means the same infrastructure that manages a molecular dynamics simulation manages a game session. The patterns are isomorphic — both are workloads with lifecycle, provenance, and capability requirements.

Ecosystem Evolution Map

Deploy graphs create a feedback loop:

Product composition discovers capability gaps
    -> Gaps become tasks distributed to primals and springs
    -> Primals evolve new capabilities
    -> Springs validate new capabilities
    -> Products recompose with expanded capability set
    -> Next product discovers next gap

Each product composition stress-tests the ecosystem in a different domain. The NF case study is the gen5 exemplar: four products composing to serve a biological question that no single product could answer.

The GarageBand Moment

The deploy graph is sheet music. The primals are instruments. The product is the song.

A musician does not build a piano to write a concerto. They learn the instrument’s capabilities and compose within them. A product builder does not build barraCuda to do protein structure prediction — they compose a deploy graph that declares what GPU capabilities the product needs, and the ecosystem provides the instruments.

The “GarageBand moment” is when the deploy graph tooling becomes simple enough that a domain scientist — not a systems programmer — can compose primals into a product that serves their question. The BYOB TOML schema is the first step: declarative, human-readable, versionable, and self-documenting.

Deploy graphs are not configuration files. They are the interface between infrastructure and science — the boundary where sovereign compute meets domain questions, mediated by TOML and JSON-RPC rather than shared code and platform lock-in.