Capability to regression graduation

Trace-detective lab (Module 8, L8.6): an eval that is right about everything and tells you nothing. EVAL-7 runs ten quarters from 12 to 94 percent and its score rises every single one -- while its gains fall from +14 a quarter to +2 and then +1 at the same $180 a run. The gain into Q8 is +8 against an established pace of 11, so the eval graduates at Q8, confirmed by Q9; Q9 and Q10 took 3 points for $4,680, which is $1,560 a point against $167 at Q5 (14x), and graduating at Q8 instead of Q10 saves $4,524. The level rule fails in both directions on the other traces: EVAL-3 tops out at 78 so a line at 90 never fires (3 idle quarters, $6,786) and EVAL-9 reaches 90 only at Q9, two quarters after its slope broke at Q7 ($4,524). One reading is not enough -- EVAL-5 breaks pace at Q5 and recovers with +14, and the confirmation reading saves 34 points of ranked progress. The ticket changes six dimensions at once (mechanism, task shape, purpose, owner, publication, cost), and the gate passes 3 of 4 candidates into 3 ties: insurance, not a ranking. In-memory only; scripted traces, 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.