Definition
A86 §3: “two notions of translation quality. Fidelity as representational similarity: Does the translation resemble the original along particular observational dimensions? Fidelity as inferential sufficiency: Does the translation preserve the information needed to make the relevant distinctions about the communicative situation? The second is much closer to the functional character of translation.” Its measure is a task-restricted Blackwell deficiency δ_Q(V_1, V_2), “naturally directed”, and A86 §7’s decision distance D_Q(U, V) = sup_q |R_q(U) − R_τ(q)(V)|. A89: “fidelity is no longer defined as similarity to a presumed global semantic object. It is defined relative to the communicative distinctions a reconstruction preserves, the decisions those distinctions support, and the compatibility relations among the contexts in which they are observed.”
What it is not
Not a replacement for the similarity senses (TERM-021, TERM-015). The manuscript’s L = L_obs + λL_str + γL_dec keeps a similarity term, a structural term and this one (§6). The named contrast between the two kinds of fidelity is not in the manuscript. It is the sense CLAIM-050 is stated in.
The crystallized argument (§24) gave the decision term its sharp form, a worst case over tasks: D_dec = sup_(q∈Q) |R_q^o − R_τ(q)^t|. The manuscript’s L_dec “compares achievable decision risks for paired communicative tasks” (§6) and drops the supremum, so it no longer says whether fidelity is judged on the worst task or on an average.