Scientific State Architecture

Science needs state before it can become programmable.

AI-for-science systems are arriving before the substrate beneath them exists. Agents can read papers, draft hypotheses, plan experiments, and operate tools, but most of their conclusions disappear into context windows, private files, and incompatible databases. There is no shared layer where a field's compiled state can persist, be corrected, be queried, and be inherited by the next system.

The core architectural question is therefore not "which app should scientists use?" It is: what state object should every serious scientific app, agent, lab, funder, and registry be able to call? The answer cannot be the paper. Papers are source artifacts. The unit of state has to be the finding: a claim with evidence, provenance, uncertainty, typed relations, time, and revision history attached.

State, runtime, network.

The older metaphor still matters: sky, astronomy, observatory. But the external architecture is simpler and more useful when named directly.

The stack is inverted today. Runtime is where most AI-for-science energy is going, while the state layer underneath it remains weak. That forces every agent to recompile the same literature, every lab to preserve its own private map, and every application to rebuild a field from papers instead of inheriting from a live record.

The finding bundle is the atomic object.

A finding is not a sentence extracted from a paper. It is a state object whose claim remains attached to the evidence that produced it, the uncertainty around it, the things it supports or contradicts, and the history of how belief changed over time.

{
  "claim": "Mechanism X increases blood-brain barrier permeability in condition Y",
  "evidence": ["experiment", "trial", "replication", "negative result"],
  "provenance": ["paper", "dataset", "protocol", "lab run"],
  "confidence": {
    "level": "contested",
    "history": ["strengthened", "weakened", "replicated"]
  },
  "links": [
    { "type": "supports", "target": "finding:..." },
    { "type": "contradicts", "target": "finding:..." },
    { "type": "depends_on", "target": "finding:..." }
  ],
  "world_time": "when reality produced the event",
  "system_time": "when the record learned of it",
  "revisions": ["who changed what, when, and why"]
}

This is the difference between paper-level search and scientific state. Search returns containers. State returns the live object inside the container, with its dependencies and revision history preserved.

The company-scale object is the change in belief.

The strongest primitive is not only the finding. It is the state transition: a claim was believed at one confidence, evidence arrived, provenance was attached, dependencies updated, and the field's state changed. That is the event future scientific systems need to observe, audit, and act on.

This is why the architecture is more than a knowledge graph. A static graph can describe what was compiled yesterday. Scientific state has to remember how the map changed, who changed it, why it changed, and which downstream objects now inherit the update.

Keep the protocol open. Build the best hosted state layer.

Vela should be framed as the proposed open protocol for scientific state: finding bundles, typed links, evidence objects, provenance, confidence, time, revision, and writeback. It should not need to be the only compiler, the only registry, or the only application.

The commercial layer should be the reference infrastructure around that protocol: hosted registries, high-quality compilers, private and public corridors, APIs, SDKs, agent integrations, review workflows, trust systems, and enterprise-grade provenance. The protocol is the standard; the hosted layer is the product.

Prove finding-level state in one corridor.

The first product should not be a generic "AI scientist" or a beautiful graph of all science. It should be a corridor: one high-value scientific question compiled into durable, queryable state until it can answer questions that PubMed, Google Scholar, Semantic Scholar, and a pile of PDFs cannot.

A corridor should expose contested findings, replicated findings, failed hypotheses, missing experiments, confidence drift, dependency collapse, and the next tests that would most reduce uncertainty. The claim should be earned practically: finding-level state beats paper-level search in a real field, with real users, real errors, and real review.

If that works, the expansion path is clean: compile public science, compile private science, expose APIs, power agents, accept experimental writeback, track confidence changes, coordinate funding against gaps, and eventually settle attribution and capital against scientific state transitions.