The rule that keeps failing is a ratio, and that is the whole mechanism. At the R5 holdout the governed route sent 763.5 prompt tokens and the lexical route sent 110.0, which the author records as 6.94x against a frozen ceiling of 3.0x [7]. Passing would have meant fitting the same selection into 110.0 x 3.0 = 330 tokens, a cut of 57 percent on what governance actually sent [15]. Nothing in that arithmetic accuses the governed route of waste in absolute terms. 763 prompt tokens is not a large prompt.
The denominator moved for a defensible reason. R3's catalog was lexically tractable, so a keyword filter had real signal to match, scoring 0.660, then 0.737, then 0.631 mean F1 as the catalog grew [5]. Rebuilding R4 and R5 on opaque physical names removed that signal [6]. A route that matches nothing also retrieves nothing to describe, so its prompt shrank to 110 tokens, and under a ratio ceiling the denominator shrank with it [7]. My read: cost per correct answer is the metric this comparison wants, and it is undefined for a baseline scoring zero.
The other figure worth pricing is the direction of travel. Governed mean F1 went 0.780, then 0.632, then 0.447 across the three catalog sizes at the prespecified 0 percent classifier-miss condition [5], a relative fall of 42.7 percent from the smallest catalog to the largest [17]. That is the opposite of the scaling behaviour a catalog gets bought for.
Now what would have to be true for the sponsored number to transfer. The McKnight Consulting Group study this harness reproduces in structure, "Stop the Token Bleed" by Jake Dolezal and William McKnight, held the model constant and reported up to roughly 89x fewer tokens at scale with F1 1.000 against 0.29 to 0.66 ungoverned; it was sponsored by Informatica, a Salesforce company [11]. The baseline there is ungoverned stuffing, and this harness agrees on that comparison [10]. So the 89x is a claim about shops whose current practice is to shovel the schema into the window. Where a keyword prefilter already sits in the pipeline, that study did not run your experiment.
The earliest rejection was on quality rather than cost. At R2.1's 3,000-object holdout the governed route scored 0.24065 against a prespecified floor of 0.245533, short by about 2 percent, while the lexical prefilter scored 0.588, or 2.44 times the governed score [13][18][19]. The author discloses that he is employed by Salesforce and publishes the harness so readers can generate their own numbers instead of trusting his [12]. R3 and R5 ship public decision packs; R4's pack is held, and labelled as held [14].
