Hypothesis from Data
A derived rule or model proposed from captures — useful, testable, not yet ground truth
A hypothesis from data is a generalized claim — formula, law, model weights, business rule — proposed from tier-one captures by analysis. It is not an L0 observation of the world. Typical placement: L2 compound aggregate (training pipeline, regression, rule engine) plus tier 6–8 (model-based / hypothesis / tested inference) until out-of-sample or prospective evidence accumulates.
Relative to the sample used to form it, the hypothesis adds no Shannon surprise: it is computable from (dataset, method). Relative to a reader who did not have it, transmitting the written rule is new information — on the communication channel from author to reader (channel capacity applies to that link). Those are different variables: DPI bounds re-processing the same past captures; Shannon transmission bounds telling someone who lacked the summary.
Predictive value is separate again. A validated kinematic law lets you forecast future positions from initial conditions without measuring every instant — enormously useful. The forecast is an L2 projection; each future instant still meets the Real only when new L0 measure captures arrive (or when a committed prediction event is later compared to outcome). The formula did not create mutual information with future ground truth from past rows alone; it compresses expectations until independent tests confirm the Real still behaves that way. Failed predictions demote tier; successful holdout and replication promote it.
Register honestly: HypothesisRegistered { rule, datasetVersion, method, author, at } as commit or assert; keep the causal chain from raw captures to published rule (Event Sourced Science).
Corpus stance
A1 — adopted on this site: Reconciles “this formula changed what we can do” with “this formula is not tier-zero fact until the world agrees.”