Grounding

Chandu, Bisk and Black (2021) observe that the two fields that use the term do not mean the same thing by it. Natural language processing applies it to "any linking of text to data or non-textual modality", which makes grounding a retrieval problem: the system is grounded when the right source has been placed before it. Retrieval-augmented generation (Lewis et al., 2020) is the canonical mechanism, and the industrial usage that speaks of grounding a model in a knowledge base descends from it.

Cognitive science defines the term more strictly. Following Clark and Brennan (1991), grounding is the process of establishing what mutual information is required for successful communication between two interlocutors. Grounding on this account is a coordination problem rather than a retrieval problem. It is achieved between parties, not supplied to one of them, and it is complete when both hold enough in common to proceed, not when a document has been read.

Chandu, Bisk and Black argue that the narrower usage has crowded out the wider one, and that the field should prefer the wider. They call the two regimes static and dynamic grounding. Static grounding treats common ground as given in advance, so that the agent need only link a query to an existing store. Dynamic grounding holds that common ground "is built via interactions including: requesting and providing clarifications, acknowledging or confirming clarifications, enacting or demonstrating". They judge the dynamic account "more general and more accurately matches real experiences".

The framework adopts the wider sense. What an agent loads is the material of grounding; it is not the whole of it.

The two terms are close enough in sound to be confused, and the literature separates them explicitly. Discussing the point at which an agent and a person have grounded a reference, Chandu, Bisk and Black note that the two need only reach consensus on truth values, "which need not be the same as the ground truth".

The distinction is one of purpose. Ground truth is an evaluative construct: reference data assumed correct, against which a system's output is scored. It exists to measure. What an agent loads in order to ground itself is not scored against anything, and is not assumed correct in the strong sense. A memory may be stale, a recipe may describe a command that has since changed, and a knowledge document may lag the code it documents.

What the framework asserts of these documents is weaker and different in kind. It asserts that an agent must prefer them to its own recall, because its recall is a statistical reconstruction and they are a record (Large language models). Preference over recall is a claim about provenance. Correctness is a claim about content. Only the first is being made.

Ground truth does have a proper place in this project, but a different one. An evaluation of a skill requires cases whose correct outcome is known independently, and those cases are ground truth in the strict sense (Verification). That is a construct for measuring the framework, not a construct within it.

The static component of grounding is discharged by loading, which classifies those documents an agent reads to inform work rather than acts upon (Regulatory Functions). Three types carry the weight of it, and they differ in the condition under which each is read.

Knowledge is read by relevance to the work. It states what a domain concept means, and an agent about to touch a concept is expected to resolve it here first.

Recipes are read on demand, when a command is needed. They state how an operation is performed in this system, which is frequently not how it is performed elsewhere.

Memories are read unconditionally. They record corrections that have already been made once, and their purpose is defeated if an agent consults them only when it suspects it needs to.

Loading answers what an agent knows. It does not answer what happens when the material does not exist, and that case is the common one on unfamiliar work.

The framework's answer is the rule that an unresolved concept becomes the first task rather than an assumption (Methods). An agent that cannot resolve a term to a document does not proceed on its reconstruction of what the term probably means. It stops and asks, or it writes the missing document and has it reviewed.

This is dynamic grounding in the sense Chandu, Bisk and Black intend. It is a request for clarification, and it treats the absence of common ground as a condition to be repaired rather than a gap to be bridged by inference. The corresponding move in the other direction is the record an agent leaves on a task, which returns what it established so that the next session and the human reviewer inherit it.

The asymmetry is worth stating plainly. An agent's reconstruction of an unfamiliar term is often plausible and occasionally correct, and the two cases are indistinguishable from the inside. Asking is therefore cheaper than it appears, and inferring is more expensive.

Chandu, Bisk and Black decompose common ground into four purviews, which they present as successive stages. The framework's document types correspond to them closely enough that the correspondence is worth recording, since it indicates where the catalogue is complete and where it is thin.

Personalised consensus is the one most easily overlooked, and the one memories exist to serve. A memory does not record a general truth. It records what this project decided, which is why a memory is neither knowledge nor a principle.

Clark and Brennan enumerate the constraints a medium places on grounding, among them copresence, visibility, cotemporality, sequentiality, reviewability and revisability. The list explains a structural feature of this project that would otherwise look like mere preference for documentation.

An agent and its principal do not share cotemporality. The person is frequently absent while the work proceeds, and reviews it afterwards. Simultaneity is likewise unavailable. What remains are reviewability and revisability: the medium permits either party to return to what was established and to correct it.

A framework grounded under those constraints must therefore write things down, because writing is the only channel that survives the absence of the other party. This is the reason the catalogue is documentary rather than conversational, and the reason a session that establishes something and does not record it has not finished grounding. It has grounded itself and left its principal ungrounded.

An earlier revision of the framework used grounding as a collective noun for the types implementing loading. That usage is withdrawn, for two reasons.

The first is a category error. Grounding names something an agent does. Knowledge, recipes and memories do not ground; an agent grounds itself by reading them. Naming the set after the activity invited the reader to treat the two as one thing.

The second is that the set had ceased to be well defined. When regulatory functions were extended across every document type, loading came to be implemented by most of them, including tasks, stories, sprints and designs. The collective noun continued to name three. A term that names three members of a set containing fifteen is a fossil of an earlier taxonomy rather than a description of the current one.

Where the set must be referred to, "the types implementing loading" states it exactly and in the vocabulary the model already supplies. The word grounding is reserved for the activity throughout.

The preceding sections support a single requirement, stated here for reference.

Before an agent acts on a piece of work, it resolves each domain concept the work touches to a document that defines it. Where no such document exists, producing one becomes the first task, and the agent does not proceed on an assumed meaning. Where the work involves a command, the agent consults the recipe rather than reconstructing the command. Memories are read at the opening of the session, unconditionally. What the session establishes is recorded before it closes.

The obligation is proportionate to the work. A method that fires on every trivial change trains people to route around it (Methods). The obligation binds where the cost of a wrong assumption exceeds the cost of checking, which is the condition it exists to serve.