The Tracking Minimum

The Tracking Minimum (TM) is the condition that a source can transmit testimonial knowledge that p only if the truth of p, or the evidential support for p, figures in the explanation of its producing the assertion that p, and figures there through the source’s own production process rather than through a proxy on which that process depends. I state and defend it in Assertion Without a Speaker, where I argue that the five leading accounts of testimony (Burge’s, Williamson’s, Goldberg’s, Faulkner’s and Lackey’s) all entail it.

Its strength is fixed by a case. A broken thermometer sits in a room held at constant temperature by a thermostat. Its readings are correct, for a stable and explicable reason, yet posting them transmits no knowledge of the temperature. A merely statistical condition, a high and non-accidental rate of truth, is too weak to exclude it. An internalist condition, production governed by states the source represents as evidence, is too strong and is not common ground: Lackey’s cases show that belief is not required. TM is explanatory, and the thermometer fails it, since nothing about the temperature explains what it displays.

TM can be satisfied originally, when the production process itself secures the connection (belief on adequate grounds, evidence-responsive production, or production governed by norms whose function is to constrain assertion to truth), or derivatively, when a source faithfully preserves a particular assertion whose source satisfies it originally. The clause about proxies is the endogeneity requirement, and it carries most of the weight.

Three further features matter. TM is neutral between safety and sensitivity: the thermometer in the thermostatted room is safe, and is excluded on other grounds. Its test is interventionist, not psychological: hold the evidence fixed and vary the proxies, then hold the proxies fixed and vary the evidence. And it is necessary for transmission, not sufficient. A working barometer satisfies it, so TM is not the line between testimony and instrument reading. That current language models satisfy it in neither way, on the best current evidence, is therefore a result about them rather than their category: they fail a condition ordinary instruments meet by construction.