Read the explanation
The simulation begins with equal probabilities over the supplied entities. Shannon entropy sums minus p times log base two of p for positive probabilities. Our illustrative four-entity matrix therefore starts with probability one quarter each and entropy two bits. A balanced binary answer retaining two candidates leaves one half each and one bit. The probability segments use four hundred pixels per probability unit: four segments of one hundred become two of two hundred. These are toy hypotheses, not outputs of an external model or verified facts. The algorithm assumes the supplied feature values determine exact yes or no answers. Displayed uncertainty describes that authored matrix and those assumptions. The ranking computes expected entropy after a feature question, weighting each answer branch by its probability mass. In our equal four-entity illustration, a two versus two split has one bit in either branch, so expected entropy is one and information gain is one. A three versus one split has log base two of three bits in the larger branch and zero in the singleton branch. Weighted by three quarters and one quarter, expected entropy is about one point one eight nine. The gain is about point eight one one. The entropy bars use two hundred pixels per bit: two hundred for the balanced split and about two hundred thirty seven point eight for the unbalanced split. Higher gain sorts first, but this one-step ranking is not a proof of globally minimal questions for every dataset. Feature keys are taken from the first entity, so a feature missing from that key list is not automatically considered. Applying an answer uses a likelihood of one for matching binary features and zero for mismatches. The remaining probabilities are divided by their sum. For a hypothetical prior point two five each, retaining two gives a total point five and posterior point five each. The bars show those two normalization totals using four hundred pixels per probability unit: two hundred before normalization and four hundred after. Automatic best-question execution looks up the stored target entity and reads its feature; it is not discovering an unknown real-world answer. If a contradictory response eliminates every candidate, active entropy returns zero as well. Zero bits therefore must not be confused with unique identification without checking the active count. The saved matrix, answer history and exported probabilities need separate native checks; the D three cluster is a separate visual dependency. In the current offline test the missing D three library throws during initialization, before native control handlers bind. These equations describe the written implementation; working queries and exports are not verified in that environment.