Reading a distribution: bimodal means two regimes
How a CRO reads shape — the modes are the contingency.
A distribution with two peaks is not a confused forecast; it is an honest one. The two modes are two regimes the model cannot distinguish between, and the contingency plan has to cover both. A CRO who reads the shape reads the contingency; a CRO who reads only the mean misses the entire forecast.
A unimodal distribution says "here is the likely outcome, and here is how spread I am around it." A bimodal distribution says something different and more important: "there are two regimes I cannot distinguish between, and the outcome is probably one of them, not something in between." Reading the second as if it were the first — averaging the two peaks into a single mean — is the most expensive mistake a CRO can make reading a forecast, because the average is an outcome the model did not predict.
The mean of a bimodal is a lie
If a distribution has one peak at 3% default and another at 9% default, the mean might sit at 6%. But 6% is not a forecast; it is an average of two forecasts the model treats as separate regimes. The model did not say "around 6%." It said "one of two worlds, and here is the probability of each." Hedging for 6% when the actual outcome is either 3% or 9% is hedging for a world the model does not believe in. The shape — two peaks, a valley between them — is the forecast; the mean is a number that hides it.
What the modes are telling you
The two modes usually correspond to two causal regimes the sisters could not resolve: a base case and a tail case, or a policy-hike path and a hold path. The probability weight on each mode is the model's honest disagreement with itself about which regime is live. That disagreement is the signal — it tells the CRO which contingency to plan for and with what weight. A unimodal forecast hides the regime split; a bimodal forecast surfaces it. The Oracle does not collapse the modes; deino-oracle normalizes the distribution (INV-1) but preserves the shape, because the shape is the decision.
The mean of a bimodal is an outcome the model did not predict. The two modes are the forecast; the valley between them is the contingency the CRO has to plan for.
- Two peaks = two regimes the model cannot distinguish between — the contingency lives in both.
- The probability on each mode is the model's honest disagreement about which regime is live.
- The mean of a bimodal is a number that hides the forecast — never hedge to the average of two worlds.
Calibration is what makes the modes trustworthy
A bimodal shape is only useful if the probabilities on the modes are calibrated — if 30% actually happens about 30% of the time. That is the Gate 1b loop: shadow runs against a realized timeline, scored with Brier and log-loss via deino-eval, against a disclosed baseline. We are pre-revenue, building the POC; the distributions are produced, and the calibration that makes the modes worth reading is gated. An uncalibrated bimodal is a more sophisticated guess than a point estimate, not a better one. The shape earns its keep only after the loop earns the word "probability."
Read the shape, not the mean. The modes are the contingency; the valley between them is where the CRO's plan has to live. A forecast that hides the modes has hidden the entire decision.
— risk note