Fork semantics: what the child sees

Ablation lab (Module 9, L9.4): a fork decides what the child sees, and getting it wrong costs 115x and contaminates the grade. The same repair runs three ways: 16 tokens with a distilled brief in 1 step keeping 3 of 3 facts, 13 tokens with an anemic brief in 2 steps keeping 0 of 3, and 1,849 tokens inherited in 1 step -- 115.56x the bill, and contaminated, because the parent transcript carried the expected diff. Then the brief schema (task, scope, acceptance, artifacts), checked at spawn: the anemic brief is refused for missing three fields, and a 23-token brief that pastes in the answer still passes. Ablate the choice across 3 families x 2 fork modes -- inheriting succeeds on 3 of 3 and zero-context on 2 of 3, until the contaminated successes are struck out and zero-context counts 2 of 3 against inheriting 1 of 3, because a contaminated success does not count. Inheritance wins exactly one family, debugging, where the 4 dead ends are the evidence and no brief can carry them; it loses review, where the context holds the diff under review and its answer key. Five inherited levels cost 57,319 tokens against 80 in briefs (716.49x), and the deepest child carries 29,584 tokens, 1,849.00x one brief. In-memory only; scripted worker and brief, words as the token proxy, no model or provider call.

This Harness Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.