Source-linked AI summary
Knowledge Cards: Structured Knowledge for AI Systems
Liliana Ferreira
TL;DR
Current documentation describes AI systems and their data but leaves undocumented the concepts, relationships, reasoning patterns, and limits behind consequential agentic decisions. The paper introduces ontology-grounded Knowledge Cards: expert-signed, provenance-linked artefacts for bounded concepts that agents can reason over and organisations can audit. Initial prototype cards cover energy and pharmaceutical settings, and the schema is released as a public draft.
Problem
Existing AI documentation omits the validated knowledge layer containing the concepts, relationships, reasoning patterns, and limits that agentic systems use when acting on conclusions.
Method
The paper defines Knowledge Cards as expert-signed, ontology-grounded artefacts that record entities, relationships, reasoning, boundaries, and provenance for one bounded concept.
Results
Initial prototype cards have been built for energy and pharmaceutical domains, while the Knowledge Card schema is released as a public draft.
Takeaways & Limitations
A shared library of Knowledge Cards can serve as an organisational semantic layer for agents and provide humans with auditable, bounded knowledge to review.
Abstract
from arXiv · showhide
AI systems whose outputs inform real decisions, and increasingly consequential ones, require something that current documentation practice does not provide: a structured, inspectable representation of the knowledge they need to ground, contextualize, and reason about those decisions, ideally reviewed and signed off by a domain expert. Established documentation artefacts already capture important aspects of an AI system. Model cards describe how a system behaves, data cards describe what it was trained on, and system cards describe the risks of a deployed system. None of them addresses the layer between inputs and outputs, more precisely, the concepts a system holds, the relationships it models, and the patterns of reasoning it applies. For pattern-recognition tasks this gap is tolerable. For agentic AI, where systems act on their conclusions, it is the step that most often separates a promising proof of concept from an operational solution an organisation can rely on. This paper introduces the Knowledge Card, a structured artefact that captures validated knowledge about a single bounded concept in a form that experts can review, organisations can audit, and AI systems can reason over. For one concept, such as a specific failure mode, a compliance obligation, or a process decision, a Knowledge Card records the entities and relationships involved, the reasoning that connects them, the conditions under which that reasoning no longer holds, and the provenance of every claim, all grounded in a formal domain ontology and signed off by a domain expert. Initial prototype cards have been built in the energy and pharmaceutical domains. The schema is released as a public draft for community engagement.
1 Introduction
Existing AI documentation covers system behavior, data, and deployment risks, but not the validated knowledge layer agentic systems use to reason and act. The paper introduces Knowledge Cards as ontology-grounded, expert-validated units for bounded reasoning tasks, designed for review, audit, and machine use.
- Model, data, and system cards document behavior, inputs, and deployment risks, but omit the concepts, relationships, reasoning patterns, and limits underlying system decisions.
- This omission is especially consequential for agentic AI, whose actions require organisations to inspect, validate, and govern the knowledge behind conclusions.The paper illustrates this with a wind-farm agent deciding whether a gearbox vibration alert warrants maintenance dispatch or observation.
- The Knowledge Card captures validated knowledge about one bounded concept, including entities, relationships, reasoning, applicability boundaries, and claim provenance.
- Knowledge Cards are grounded in a formal domain ontology and signed off by a domain expert, allowing LLM agents to reason over them while organisations verify bounded reasoning.
- The architecture separates a shared formal domain ontology from individual cards that instantiate it.
- The structure applies across failure modes, compliance checks, and process decisions, allowing a card library to function as an organisational semantic layer.
- The paper discusses agent memory, industrial diagnostics, and regulatory compliance, and releases the v0.1 schema as a public draft.
2 Related Work
Prior documentation artefacts describe systems and data, while retrieval and graph-based tools improve access to relevant material; neither necessarily supplies formally grounded, scope-declared, expert-validated reasoning knowledge. The Knowledge Card is positioned as a shared structure addressing that gap.
- 2.1 AI Documentation Artefacts: Model, data, system cards, and AI FactSheets established structured documentation for system construction, behavior, trustworthiness, and deployment risk.
- 2.1 AI Documentation Artefacts: Existing artefacts do not capture the specific concepts, causal relationships, and inference patterns a system applies to a particular case in machine-readable, formally grounded form.
- 2.2 Retrieval, Graph-Augmented Retrieval, and Context Graphs: GraphRAG and related tools organise corpora so language models can find and combine relevant passages more effectively.
- 2.2 Retrieval, Graph-Augmented Retrieval, and Context Graphs: Automatically constructed retrieval graphs may label nodes and edges with model-inferred strings whose meanings are not fixed against a formal ontology.
- 2.2 Retrieval, Graph-Augmented Retrieval, and Context Graphs: TrustGraph adds OWL-grounded types, PROV-O provenance, and temporal awareness, whereas OKF uses linked Markdown files with YAML frontmatter and leaves the content model to producers.
- 2.2 Retrieval, Graph-Augmented Retrieval, and Context Graphs: The comparison frames an agentic knowledge source as needing formal grounding, reasoning and verification, declared scope, and explicit expert validation.
- 2.2 Retrieval, Graph-Augmented Retrieval, and Context Graphs: A Knowledge Card contributes shared structure so the same concept retains the same meaning across systems that consume it.
3 The Knowledge Card
A Knowledge Card is a formal, expert-validated representation of one bounded concept, connecting ontology-grounded entities and relationships to reasoning, provenance, boundaries, and lifecycle. The wind-farm example shows how this structure supports precise, auditable decisions while deferring when its diagnosis does not apply.
- A Knowledge Card is a named individual for one bounded concept, using vocabulary from a formal domain ontology.
- Its tuple contains metadata, a concept anchor, entity register, relationship topology, reasoning patterns, boundary conditions, and a consuming-system link.
- Worked example: The wind-farm card represents outer-race spalling through typed entities and observable signals including vibration amplitude, oil temperature, and ferrous particle count.
- Worked example: The card turns a signal pattern into a calibrated judgement while specifying conditions that require the agent not to decide alone.
- Five defining properties: Five defining properties are ontology-grounded, provenance-linked, boundary-explicit, expert-validated, and versioned.
4 Schema and Reasoning
The schema serializes Knowledge Cards as portable, ontology-grounded artefacts, while the reasoning layer expresses the same inference from readable narrative through machine-checkable rules and optional probability. Explicit boundaries can override conclusions or route cases to human review.
- Schema: The schema follows the card tuple, with metadata, ontology concept, typed entities, relationships, layered reasoning, boundaries, and system linkage.The schema is specified as JSON Schema 2020-12 with a JSON-LD context.
- Reasoning: Cards use only the reasoning layers their knowledge supports, with plain language and formal rules sufficient for a complete useful card.
- Reasoning: Reasoning layers move from plain-language explanation to structured signal patterns and formal ontology-class rules.These layers preserve readability for experts while enabling systems to match evidence and verify conclusions.
- Reasoning: A probabilistic layer is added when calibration data support context-dependent signal weights and belief updates for conflicting evidence.The card’s causal structure can map to a Bayesian network, while partly known causal structure can be expressed with a Markov Logic Network.
- Boundaries: A card may combine formal rules, probabilistic reasoning, and human deferral, while boundaries remove its authority to conclude when conditions invalidate it.A card-invalid condition blocks positive conclusions regardless of supporting evidence; the wind-farm example escalates after recent re-greasing.
- Portability: The schema defines what card contents license but leaves inference computation to consuming systems, allowing LLM and symbolic reasoning over the same card.
5 Applications
Knowledge Cards address the accountability gap that emerges when agents move from retrieving context to taking consequential actions. The paper illustrates this scope through agent memory, wind-farm diagnosis, and pharmaceutical compliance.
- Agent Memory: In production-oriented agentic AI, organisations need validated, bounded, attributable knowledge rather than only retrieved text or assembled context.A Knowledge Card library identifies the expert-validated portion of agent knowledge that can be cited and audited.
- Wind-Farm Maintenance: For the wind-farm agent, the gearbox card combines formal matching, probabilistic confidence, and an engineer-readable rationale before a maintenance recommendation.The passage reports confidence of 0.89 given rising particle count alongside vibration, but the supplied text ends before the boundary outcome.
- Pharmaceutical Compliance: In pharmaceutical manufacturing, a GMP change-control decision requires current risk assessment, approval signatures, and revalidation evidence before a validated-process change may proceed.The example treats a compliance obligation as a bounded concept for Knowledge Card representation.
6 Conclusion
The paper presents Knowledge Cards as a standard artefact for making agent knowledge explicit, reasoning checkable, and limits known. Their shared structure supports organisationally consistent, auditable reasoning, while the schema is offered as a public draft.
- The paper identifies an undocumented knowledge layer comprising concepts, causal relationships, inference patterns, and limits of reasoning.
- A Knowledge Card gives this layer a formal ontology, expert sign-off, explicit boundaries, full provenance, and five named properties.The properties are ontology-grounded, provenance-linked, boundary-explicit, expert-validated, and versioned.
- A shared library of cards acts as a semantic layer so organisational systems can reason from the same validated knowledge.
- The contribution is a candidate standard specifying validated knowledge as an artefact independently of storage, retrieval, or serving.The schema is released as a public draft and invites community contributions.