MK-SH-1

A decision model that reads a value from the state and answers a threshold question. It returns calibrated probabilities, a choice, and a refusal when its confidence is too low.

v1.2 · 140M parameters · T = 7.653133 · tau = 0.79 · loads from Gjusev/MK-SH-1

Use cases

Three interactive interface studies for queues that benefit from an explicit review lane. They illustrate the decision contract; the live checkpoint runs in the lab immediately below.

Swipe horizontally to inspect every lane at full, readable size.

Live decision lab

Every value below comes from the shipped formula: the encoder scores each option in one pass, softmax at T = 7.653 turns the scores into calibrated probabilities, and the frozen gate at tau = 0.79 sends anything below it to review instead of answering.

State, as JSON (the value lives in motor_temp_c)

The state the model reads. In this lab only the motor_temp_c field matters.

Threshold t, in degrees (question: is the reading at least t?)

The question's tide line. Changing it re-runs the decision.

After you run a decision: p(No) and p(Yes) — the calibrated probabilities softmax(logits / T = 7.653). They sum to 1; the larger one is the confidence the gate compares against tau.

The decision: the winning option as choice, or null with reason_code below_validated_threshold when the confidence sits below tau = 0.79. The rule check underneath compares it with the answer computable from the state.

Run a real checkpoint example

Examples

State, as JSON (the value lives in motor_temp_c)	Threshold t, in degrees (question: is the reading at least t?)