Human in the Ground: A New Paradigm for Contestable Machine Reasoning

In 1958, Michael Polanyi made an observation that has shaped how we think about expertise ever since: experts know more than they can tell. From this he concluded that expert knowledge cannot be transmitted. He was right about human-to-human transmission. He was not thinking about what a machine might do with the output of that knowledge. This paper argues that machines do not need the full knowledge of an expert. They need the output of that knowledge, written down at the object the expert examined, anchored with an identifier someone else can independently find, with the reason stated and the judgment open to contestation. Put simply: the machine does not need to understand how the expert knew. It needs to know which expert concluded what, and where. I call this "human in the ground," the human who lays the foundation before the machine reasons over grounded expert judgments, rather than the human who watches the machine and catches its mistakes. Three conditions make this work. The judgment must be addressable: anchored at an identifier a qualified party can independently reach. It must be inspectable: the reason stated, not just the conclusion. And correction must be a permanent, traceable record at the same anchor, not a discarded footnote. Without all three, grounding becomes annotation with extra steps. This is not annotation. Annotation enriches an object. Grounding makes reasoning contestable before the machine acts. Human expertise is tacit and slow. Machines are fast but blind without ground. Grounding is the connective layer that converts human expertise into machine-readable judgments. Contestability is the mechanism that makes those judgments more accurate over time. And scale and speed are what transform the whole from an individual act of knowing into a collective, self-correcting knowledge system. I have built multiple systems on this mechanism. Two of them offer the kind of evidence this paper needs: one at institutional scale, one as a design demonstration. The paper shows how they work, where they fail, and why the distinction between grounding and annotation matters for anyone who needs AI to be accountable. This paper establishes a mechanism and offers a partial, living demonstration of it, not a completed empirical result. The imperfections in the systems described here are part of that demonstration, not exceptions to it. This is the fourth paper in a series on classification, contestability, and machine reasoning. See the related identifiers for the earlier papers: The Information Problem Was Never the Real Problem (doi:10.5281/zenodo.19202082), Exploring the Space Between Peirce and Turing — Paper 2 (doi:10.5281/zenodo.20624714), and Peirceiving Turing's Dream (doi:10.5281/zenodo.20822717).