Progressive Semantic Mechanization: Compiling Intellectual Work into Deterministic Programs with Typed Semantic Holes

Which intellectual work should be an executor call (a call to a model or a person)? We model work as a program with typed semantic holes. Each site is open or closed separately for decision, production, and acceptance, and separately per input scope. Progressive semantic mechanization moves work to deterministic mechanisms through an acceptance-controlled promotion. Under the stated independence assumptions, a share c of production routed to mechanisms gives the first-order horizon ceiling H_raw/((1 − c)λ_st + c·λ_pc), where H_raw is the ungated horizon, λ_st the shared blind-spot mass of Part I, and λ_pc premature-closure escape. Closure raises this ceiling only where λ_pc < λ_st. Because pointwise evidence cannot separate twins (mechanisms with the same evidence record), a confidence bound valid over the mechanism class can fall below a twin's fault mass only with the allowed error probability. Same-family evidence inherits the family's indistinguishability, not a general lower bound on the mechanism's error. Copies of outputs accepted by sound same-family gates inherit the floor in expectation; the corresponding claim for generalizing builders is an untested hypothesis. Further results give a validation-inclusive break-even volume, a composition bound under completeness and provenance assumptions, and two separate scale gains: from fault-free, rejection-free closure, and from bounded load under an untested hazard hypothesis. The results are elementary statements in the model, not universal laws. No modeled quantity is estimated from data. The simulation uses chosen parameters, the plan is a design, and the instance reports counts only.