We ship cones, not forecasts

The word "forecast" invites a number: the rate will be 4.2, the load will peak at 18 GW, the churn will be 6 percent. A number is the one output a CRO cannot defend, because the moment it is wrong there is nothing left to point at. A cone is different: it is a distribution that widens with time, carries its own entropy, and says "here is the central path, here is the tail, here is how much I disagree with myself about each." The product surface is the cone. The number inside it is a derivative, not the deliverable.

Why a cone and not a path

A path is a single trajectory; a cone is a family of them, weighted. The Oracle normalizes the sisters outputs into an Ensemble in exactly one place: probabilities sum to one, scenarios sorted descending, entropy in nats. The cone is that Ensemble rendered over time — the spread is the entropy, the modes are the regimes, the tails are the contingencies. A CRO reads the shape to decide how much to hedge; a point estimate gives them nothing to read.

A forecast is a number you cannot hedge. A cone is a shape that tells you how much to hedge. The product is the shape.

The cone is reproducible, not just drawn

A cone that cannot be replayed is a chart, not an artifact. The manifest records the exact NodeVersions the query read, the prompt versions, the agent trace, and the output bytes — so the cone is reproducible bit-for-bit, years later. The shape is the deliverable, and the substrate that produced the shape is the evidence. We are pre-revenue, building the POC; the cones are produced, and the calibration loop (Gate 1b) is what will earn the word "probability" on them.

• The cone carries its own doubt: entropy in nats, modes as regimes, tails as contingencies.
• Normalization happens once (deino-oracle): sum ≈ 1.0, sorted descending — a guarantee consumers rely on.
• The manifest seals the cone: the NodeVersions and agent trace that produced it are part of the record.

What the cone does not promise

The cone does not promise accuracy. Accuracy is a calibration loop in progress against a disclosed baseline, gated by Gate 1b before any side is positioned as calibrated. A beautiful, uncalibrated cone is a more sophisticated lie than a point estimate, not a better one. What the cone promises is honesty about its own spread — and that is the one property a regulated decision can actually use.

We do not ship a number and call it a forecast. We ship a shape and call it a cone. The number is a derivative; the shape is the product; the honesty is the spread.

— product surface note