The deposited definition
Grounding is not an intrinsic property of the representation, not a global property of the AI performer, and not permanent once established; it is a transition-specific condition of contact that the work system establishes and maintains through anchors, verification, renewal, and passage controls.
Grounding · Appendix A, Revision 6.0.1 · DOI 10.5281/zenodo.22072938
Gloss. The commentary on this page is plain-language explanation. The controlling definitions are in Appendix A of Revision 6.0.1; do not treat the glosses as substitute definitions.
Before an AI-shaped representation is allowed to become consequential, the workflow must bind it to the relevant operational reality. Where grounding is required, a Verification Gate evaluates whether the required forms hold before permitting passage.
Four forms, four questions
The deposit distinguishes four forms, each answering a different question:
- Source grounding (against approved documents and reference material) answers “What governs?”
- State grounding (against actual current conditions of the work system) answers “What is true now?”
- Dynamics grounding (against validated relationships, physical constraints, and operating limits that describe how the system changes in response to action) answers “What happens if?”
- Work-as-done grounding (against the situated knowledge of the people who perform the work) answers “How does the work actually happen?”
Anchors are the mechanism
A specific, measurable point where a representation binds to something real that can confirm, contradict, correct, or stop it.
Anchor · Appendix A, Revision 6.0.1
Anchors produce grounding only while set and valid. An unset anchor holds nothing. A stale anchor produces the appearance of grounding while the work moves on. Anchors must also be independent: multiple representations descending from one source are not multiple anchors, and apparent corroboration across screens can share a single ancestor.
Renewal
Contact is never permanent once established. For each anchor, the work design must specify the events, conditions, or elapsed time that invalidate it. For each consequential transition, it must specify when contact must be renewed before passage. Yesterday's grounding is not today's.
False grounding
False grounding is the condition in which the workflow appears to have established the required contact because anchors are present, but one or more are stale, inapplicable, invalid, circular, or defeated. It is more hazardous than absent grounding because it can pass the checks designed to catch the loss of contact.
Anchor Access
Where verification is automated, the workflow must supply Anchor Access: the anchors the AI performer's function requires, reachable at the positions where its output crosses consequential transitions, with reaching them enforced by the harness or orchestration layer rather than left to the model. Access is necessary. It is not proof that grounding was established.