MIE files are per-database YAML documents that TogoMCP supplies to a large language model at query time so it can compose SPARQL against the DBCLS RDF Portal. Ours had grown to eleven sections, semi-automatically generated for each of 36 databases and reviewed by hand. Each section had been introduced in response to a systematic query failure observed in use — sound practice, but it left open whether any section still earned its tokens once the other ten were present. We measured that, in eighteen ablation conditions across four families: is a section necessary, is a functional group necessary, is the whole document worth anything, and is any one group sufficient alone. No single section and no single group is necessary. Removing the entire document costs 0.9 points out of 20, and the query-construction group alone recovers 99% of that — the whole is worth roughly 2.7 times the sum of its parts, the signature of heavy redundancy. We rebuilt the format around that evidence, making the verified executable example the atomic unit: 36 files, 303 examples, each 29–65% smaller than the file it replaces. A pre-registered equivalence run over 100 benchmark questions finds v3 statistically indistinguishable from v2 in answer quality (+0.29/20, 95% CI [-0.09, +0.68]) while using 15% fewer input tokens, costing 15% less and running 6% faster, with the factoid-question score up a full point. We also report eight measurement traps that faked or destroyed signal, and one budgeting error worth more than the results: we spent the most on the least informative experiment, and say what we would do instead.