Implementation: SHADOW (logit readouts) · DESIGN (hidden-state heads) · Exposure: INTERNAL. See status labels. A readout is how the decision runtime gets a probability for each candidate from a component. The readout level says how far that probability can be trusted.

​
Architecture contract

Rules:
  • A result always reports the readout level served.
  • A readout is bound to an exact component revision, tokenizer, quantization, runtime, question and candidate set. Change any of these and calibration is invalid.
  • One universal head for every question is rejected.
A primitive matures in order: untrained, L0, labels accumulate, L1, an L2 head, monitored in production, and refit or invalidate on drift. Why levels: a component’s raw probabilities are often confidently wrong. Naming the level makes the difference between “the model said 0.9” and “0.9 is right about 90% of the time on held-out data” explicit.

​
Current implementation

The decision server on main reads next-token label probabilities from a local model over a constrained label set and renormalizes them over the valid labels. Sets with more than twenty options are scored hierarchically. When label probabilities are unavailable it falls back to sampled voting and marks the result as sampled. Forward and reversed debiasing exists but is off by default, because it did not help on the default local model. Temperature fitting is implemented. Hidden-state heads are not built.

​
Availability

​
Developer relevance

If your backend can expose label probabilities, the decision runtime can read them directly. If it cannot, report your results as uncalibrated; they will not be allowed to act.