A row that never moves is not a wasted measurement. It is the evidence that the rows which moved, moved for a reason.
the boring rows
The boring rows are load-bearing
Seven of our ten measured modes are controls. Here is the case for spending most of an eval on rows that should never move.
Our census measures ten failure modes. Three are ones a supplied source ought to fix. Seven are ones it should not touch at all.
Seven to three looks like waste. It is the only reason the three can be believed.
The status quo tax
Most evals measure only the thing they hope improved. Add retrieval, grounding goes up, ship it. Nothing in that design can tell you whether the retrieval informed the model — or just changed how it behaves in general.
The proposed shift
A passage about a company that improves ledger reconciliation did not inform anything. Perhaps it made answers longer and more careful. Whatever it did, it was not grounding — and the grounding result cannot be read as grounding either.
Three columns
Control modes reuse the same harness and the same run. The marginal cost is probes, not engineering.
Without controls, a general behaviour shift ships labelled as a targeted fix. It will not hold when the prompt changes.
A claim with flat controls under it survives a sceptical reviewer. A claim without them gets re-litigated every quarter.
The two objections
"Seven controls is overkill." Two would tell you less than you think — a single flat control can be flat by luck. Several flat rows, across different kinds of task, are what make "it only moved what it should" a finding rather than a hope.
"We'd rather spend the budget on more treatment cases." Then you will have a tighter interval around a number you still cannot attribute.
The low-risk first step
Next time you test a retrieval change, add one question the retrieved text has nothing to do with. Arithmetic works. If that answer changes too, you have learned something the headline number was hiding.
In your last eval of a retrieval or prompt change — what did you measure that was supposed to stay flat?
Every figure here describes something measured and committed. See the measurements · read the method