Genie Cortex

A knowledge layer that refuses to act on an ungrounded belief. Every non-trivial action stakes a falsifiable, citation-backed hypothesis — and the runtime rejects the ones it can't ground.

Cortex is in internal beta. The contract below is stable; the surface around it is still landing.

The one idea

Cortex is discipline as product, not data as product. The moat is a single mechanism — the hypothesis enforcer. An agent must commit to a claim, cite the specific facts that justify it, and accept that the outcome can falsify it. Ungrounded actions cannot fire.

typescript
// Before any non-trivial action, the agent stakes a hypothesis:
const h = await cortex.stakeHypothesis({
  claim: "Project uses Prisma, so adding a migration is the right edit",
  citedFactIds: ["fact_abc123"],          // must be real, unexpired facts
  action: "edit:packages/state/prisma/schema.prisma",
});

// The runtime REJECTS the action if the claim is ungrounded:
//   - no cited facts, a cited fact doesn't exist, or it's expired
if (!h.accepted) throw new UngroundedActionError(h.rejectionReason);

// The action runs, and its outcome is recorded as ground truth:
await cortex.recordOutcome(h.id, { success: result.ok, evidence });

What this gives you

Falsifiability — every action commits to a claim that an outcome can invalidate. No claim, no action.
Ground-truth signal — hypothesis success rate per skill, per persona, per fact-source becomes an eval signal. Degradation is detectable without humans in the loop.
Provenance — every change traces back through its hypothesis to the facts and sources it relied on. Auditors see exactly what the agent believed and why.
Anti-game ratchet — an agent that cites facts cynically eventually shows a low success rate; the aggregate exposes the shortcut.

Why it's a moat

Other systems have citation patterns; some attach “supporting evidence.” None enforce rejection of ungrounded actions or record outcomes as ground truth. The hypothesis log becomes an internal eval no competitor can replicate without rebuilding their runtime. That is the entire design — anything not described here is intentionally out of scope.