Frequently Asked Questions

No. We don’t train foundation models and we don’t compete with the labs. Models generate candidate language; Runcible tests whether that language can become institutional action. The runtime is model-neutral — commercial, private, or local models all work — which also means no vendor lock-in.

A closed, durable artifact produced when governed work finishes: what was claimed, what evidence was used and what was missing, which rules and authority applied, what passed and failed, whether action is warrantable, and who is accountable. A chat transcript shows what was said. A Decidability Record shows what your institution may do — built to the evidentiary standards of audit, appeal, and litigation.

Runcible doesn’t return accept/reject. Work closes in one of eight states — Certified, Survives, Repairable, Escalated, Blocked, Fails, Non-warrantable, or Undecidable — each saying exactly what the institution may do next. Repairable comes with a diagnosis; Escalated routes to human authority; Undecidable is the honest answer when evidence or authority doesn’t reach.

Guardrails filter what a model is allowed to say — tone, topics, phrasing. Runcible tests what an institution is allowed to do. The principle is constraint separation, not censorship: claims are tested separately for truth, reciprocity (does it dump cost or risk on someone who never agreed? — the deep form of is it permitted), possibility, and liability before they’re adapted for institutional use. A guardrail makes output safer-sounding. Adjudication makes it signable.

No — it protects them. AI participants get governed roles with explicit scope, prohibited tasks, and escalation triggers, so AI never drifts from support into unauthorized discretion. Where judgment is required, Runcible escalates to the human who holds the pen. Authority stays exactly where your institution put it; it just stops being undocumented.

The opposite. Deployment starts in shadow mode — Runcible reviews real work without touching decisions, measured against your human baseline. Then advisory mode, then human-approval mode. Scope expands only where the record proves value. Private tenants, redacted data, local models, and air-gapped architectures are available where the work requires them.

Institutions whose work product is liability-bearing language: insurers, lenders, healthcare administrators, compliance and legal teams, government agencies, and defense organizations. The common test: if a wrong sentence in your workflow is a lawsuit, a finding, or a headline — Runcible is for you.

The stack is fourteen years deep: a formal science of decidability developed at the Natural Law Institute since 2012, compiled into the Reality Description Language, run by the governance runtime, and deployed on the Oversing platform. The models are the newest — and most replaceable — part of the stack. The library goes as deep as you want, down to computational epistemology — and the strongest evidence is the artifact itself: read a worked Decidability Record and judge the machinery by its output.

From Edward Lear’s “runcible spoon” — literature’s most famous word that sounds meaningful but has no testable definition. That’s precisely the failure mode Runcible exists to end: language that feels true but can’t be adjudicated. The spoon in our logo is a runcible spoon. We gave the word a job.