Entropy is the honesty field
Spread as truth — the number that says how much the model disagrees with itself.
A forecast with no entropy is not a confident forecast; it is a lying one. Entropy — the spread of the distribution, in nats — is the field that tells you how much the model disagrees with itself. It is the honesty field, and it is the one field a CRO reads before reading anything else.
Every Ensemble the Oracle produces carries a number that is more honest than the forecast itself: the entropy. Entropy is the spread of the distribution over scenarios, measured in nats, and it is the field that tells you how much the model disagrees with itself. A sharp peak has low entropy — the model is confident. A flat spread has high entropy — the model is not. A forecast that reports a number without reporting its entropy is a forecast that hid the one field that says whether to trust it.
Why nats, and why it matters
Entropy is reported in nats — the natural-log unit — because that is the unit the normalization in deino-oracle produces (INV-1: probabilities sum to one, scenarios sorted descending, entropy in nats). The unit choice is not cosmetic; nats are additive under conditional decomposition, which means the entropy of a joint forecast decomposes cleanly into the entropy of its parts. A CRO can read the total spread and ask where it comes from. The unit is the honesty field's precision.
Low entropy is not the same as right
A model can be confidently wrong. Low entropy means the model disagrees with itself very little; it does not mean the model is correct. A forecast with low entropy and poor calibration is a confidently wrong forecast, which is the most dangerous kind. Entropy is the honesty field, but calibration is the field that makes the honesty worth something. The two are read together: entropy tells you the spread; calibration (Gate 1b, Brier via deino-eval) tells you whether the spread is trustworthy. One without the other is half a forecast.
A forecast that reports a number without reporting its entropy is a forecast that hid the one field that says whether to trust it. Entropy is the honesty field; calibration is the field that makes the honesty worth something.
- Entropy in nats: the spread of the distribution, the model's disagreement with itself.
- Low entropy ≠ right: a confidently wrong forecast is the most dangerous kind.
- Read entropy and calibration together — one is the spread, the other is whether the spread is trustworthy.
The field a CRO reads first
Before reading the top scenario, a CRO reads the entropy. High entropy means "the model is genuinely unsure — plan for breadth." Low entropy means "the model is confident — but check the calibration before you act on it." The ordering matters: read the spread, then read the calibration, then read the scenarios. A forecast that presents the scenarios without the entropy has presented the answer without the doubt, and the doubt is the part a CRO needs. We are pre-revenue, building the POC; the entropy field is produced by deino-oracle on every run, and the calibration that makes it worth reading is gated.
Entropy is the field that says how much the model disagrees with itself. A forecast without it is a forecast that hid its own doubt — and a forecast that hides its doubt is a forecast a CRO cannot trust.
— design note