Nucleoid

Nucleoid earns its place by being narrow and honest about it. The runtime does three things well. First, declarative logic with formal constructs: you write `class Human(name: str)` and `$Human.mortal = true`, and universal assertions are first-class rather than bolted on. Second, the logic graph. Every declared statement becomes a tracked relationship between variables, objects and functions, so the runtime can deduce answers like socrates.mortal == true without you writing the derivation. Third, persistence in natural logical form inside the built-in data store — no external database to stand up before your ontology prototype works. That combination is genuinely useful when you are building the reasoning half of a neuro-symbolic pipeline and you want the neural side to stay your problem. The trade-offs are equally clear. Your output quality is capped by the rules and ontology you write — the runtime will not invent domain knowledge, and it ships no bundled LLM, so model access and deployment stay on your side of the fence. The syntax is JavaScript-derived with a minimum-token flavour, which is friendly if you already write Node.js and alien if you come from Prolog or Datalog. And the audience is narrow by design: this is a code-first tool for researchers and engineers, not a builder for analysts. Where it fits best is early-stage neuro-symbolic prototyping — encode constraints alongside a model, check outputs for logical consistency, run rule-based validation over extracted entities before training, or produce auditable inference chains where each conclusion traces back to the rule that produced it. Where it doesn't fit is anything expecting a chat interface, a connector catalogue, or a system that trains the learning component for you. The cheap experiment is to port one small rule set — a domain ontology or a compliance constraint set — and see whether the logic graph saves you more time than learning the syntax costs.