Nabla Infinity: An Epistemic Engine for Intelligence Platforms

Most intelligence platforms are maximum-likelihood machines. They tell you what they think. Nabla Infinity (∇∞) is different: it also tells you how it knows, what would change its mind, and which parts of its answer are load-bearing. That is epistemic reasoning, and it is the difference between an analyst tool and an opinion generator.

Axioms, not heuristics

∇∞ is defined by a small set of axioms that every component of Prismatic must honor. Three are doing most of the work:

1. Provenance is mandatory. Every claim carries the source(s) that produced it. A claim with no provenance is not a claim; it is a guess.
2. Uncertainty is decomposed. Epistemic (reducible by more data) is separated from aleatoric (irreducible noise). A single "confidence" number is a bug.
3. Retraction is cheap. When new evidence contradicts an earlier claim, the older claim is retracted, not overwritten. The history survives.

These axioms are enforced by mix nabla.verify, which walks the call graph and fails CI if any decision path produces a claim that violates them.

The shape of an epistemic claim

%EpistemicClaim{
  statement: "Entity X is a subsidiary of Entity Y",
  confidence: 0.78,
  uncertainty: %{epistemic: 0.14, aleatoric: 0.08},
  provenance: [
    %{source: "czech-ares", ts: ~U[...], tier: :t1},
    %{source: "cross-shareholding-analysis", ts: ~U[...], tier: :t2}
  ],
  contradictions: [],
  retractable: true,
  retraction_trigger: "ARES update contradicting ownership"
}

Notice what is not there: no boolean "true/false". Intelligence is not boolean. An analyst who treats it as boolean will eventually be wrong in a way that matters.

Retraction in practice

When a new signal arrives, the engine runs a contradiction check. If it contradicts an existing claim strongly enough, the claim is retracted and a new one takes its place — with a back-pointer to the retracted claim:

%EpistemicClaim{
  statement: "Entity X is NOT a subsidiary of Entity Y",
  confidence: 0.91,
  retracts: "claim_01HX...",
  retraction_reason: "ARES snapshot 2026-04-09 shows direct ownership dissolved"
}

The UI then surfaces both: the current claim and the retracted one. Hiding retractions is how platforms lose trust.

Why this matters for AI

LLM outputs are a firehose of confidence-free claims. Running them through ∇∞ forces them to wear their provenance on the outside: which source grounded which span, which spans had no grounding, which claims were retracted by later evidence. That single constraint eliminates most hallucination downstream — not because the model stops hallucinating, but because ungrounded claims never get past the envelope seal.

Where to go next

• Academy: Nabla Infinity Guide — the full axiom set + runnable examples
• Academy: Monte Carlo Fundamentals — how uncertainty is estimated
• Glossary: Nabla Infinity, Epistemic Reasoning, Epistemic Validation, Evidence, Confidence

Know what you know. Know what you don't. Track both. Everything else is noise.