Semantic Knowledge Technologies: what the Semantic Web lost sight of, and what it never had

• A machine checks a statement against declared constraints and existing knowledge, then reports a verdict with a reason.Source: §2 · p. 2
• A machine connects a statement by resolving its terms to entities and relations already held.Source: §2 · p. 3
• A machine derives consequences using a recorded inference method and traces them to their premises.Source: §2 · p. 3
• A machine acts by invoking computations bound to the statement’s terms.Source: §2 · p. 3
• A machine delimits its knowledge by reporting coverage, currency, and confidence, and refusing what falls outside.Source: §2 · p. 3
• Passing the five tests measures architectural understanding, not intelligence in a wider sense.Source: §2 · p. 3

• Because these omissions concern conditions and competence, adding more triples alone cannot produce reasoning.Source: §3 · pp. 3–4
• Context-free triples omit claim conditions, scope, method, evidence, and degree, preventing judgments of applicability and whether claims contradict.Source: §3 · p. 3
• RDF and OWL give terms denotational meaning but do not bind them to operations, so holding a statement confers no ability to act on it.Source: §3 · pp. 3–4
• Without declared coverage, the open-world assumption prevents a system from distinguishing missing data from an unrepresented fact or recognizing its content boundary.Source: §3 · p. 4

Layer 1. Qualified claims

• Qualified claims record conditions, provenance, evidence category, and lifecycle status, enabling applicability judgments and distinguishing contradictions from observations under different conditions.The condition schema is domain-parameterized; choosing the wrong dimensions risks under-reporting contradictions.Source: Layer 1. Qualified claims · pp. 4–5

Layer 2. Executable terms

• Executable terms bind vocabulary to typed operations and invocation details, so statements can support actions such as validation, conversion, resolution, computation, and retrieval.The proposed binding is minimal and task-driven; this layer is to be built.Source: Layer 2. Executable terms · p. 6

Layer 3. Declared coverage

• Declared coverage describes the base’s content regions, sources, releases, currency, and asserted completeness, allowing systems to refuse unsupported questions and route them elsewhere.Coverage must be asserted under the open-world assumption, and the useful granularity of regions remains an open empirical question.Source: Layer 3. Declared coverage · pp. 6–7

Layer 4. Inference

• Inference combines deduction, graph-based induction, and candidate-generating abduction or analogy, while labelling each derivation with its method, confidence, and trace.No principled semantics is offered for fusing deductive verdicts with embedding scores; the profile records both separately.Source: Layer 4. Inference · pp. 7–8

Layer 5. Learning and memory dynamics

• Learning and memory dynamics acquire, consolidate, revise, supersede, retract, and age claims, with feedback signals improving extraction and related components.Belief revision at scale remains unsettled, and consolidation, decay, and the full feedback loop are among the least mature parts.Source: Layer 5. Learning and memory dynamics · p. 8

Layer 6. Discovery

• Discovery produces gaps, contradictions, and candidate hypotheses rather than retrieval alone, using declared coverage and recorded conditions to identify meaningful gaps and disagreements.Its success is measured by expert-judged yield per candidate reviewed, since weak candidates consume expert attention.Source: Layer 6. Discovery · pp. 8–9

Layer 7. Presentation and exploitation

• Presentation delivers answers with supporting claims, conditions, checks versus inferences, alternatives, confidence, and uncovered areas, linked to underlying claim identifiers.The mechanisms include readable derivation traces, constrained language interaction, comparison of conflicting claims, interactive refinement, and programmatic access; presenting scope without overwhelming readers is the hard part.Source: Layer 7. Presentation and exploitation · p. 9

5. How the layers depend on one another

• Without conditions, inference cannot reliably judge applicability or contradiction; ungrounded terms make action unverifiable, and learning without declared limits risks undetected drift.Discovery also needs coverage declarations to distinguish an empty region from an unexamined one, while presentation needs provenance and traces to avoid becoming assertion.Source: §5 · p. 10
• The enabling layers are not ends in themselves: a base that only answers lookup queries has paid their cost without reaching the intended capabilities.The architecture is meant to support judging applicability, acting on terms, and recognizing where knowledge ends.Source: §5 · p. 10
• FAIR concerns access to and processing of data, whereas this programme concerns what machines can do with content; the two are complementary and neither implies the other.FAIR’s stated principles are findability, accessibility, interoperability, and reusability, with emphasis on machine-actionability.Source: §5 · p. 10

6. How it is used

• A machine agent queries a base, invokes bound operations, receives labelled derivations, and refuses when declared coverage does not extend to its question.This is one of the four stated modes of use.Source: §6 · p. 10
• A scientist can ask in ordinary language and receive an answer with its basis, conditions, and gaps, then move from the rendered answer to the underlying claims.The answer is intended to remain connected to its supporting claims.Source: §6 · p. 10
• Independent knowledge bases exchange claims with conditions, provenance, and evidence attached, allowing each to decide what to admit under its own policy.This makes federation between institutions more than schema matching.Source: §6 · p. 10
• Learned models can be trained or grounded on curated claims with provenance for filtering and weighting, while an autonomous agent can build and maintain its own base from declared sources.The agent studies those sources, grows the base it answers from, and returns annotations that enrich them.Source: §6 · pp. 10–11

7. Hypotheses

• The programme’s claims are framed as falsifiable propositions rather than an untestable general aspiration.The paper says the programme is stated as claims that can fail.Source: §7 · p. 11
• Qualified claims, executable terms, and declared coverage are hypothesized to enable applicability and contradiction judgments, multi-step tasks, and correct refusal and routing.For refusal and routing, the proposed comparison is against additional content or a larger model, neither of which is expected to provide those capabilities.Source: §7 · p. 11
• The cognitive-layer hypotheses predict benefits from labelled multi-method inference, maintained bases under new evidence, and discovery judged by expert-confirmed gaps and contradictions.The proposed outcomes are higher usable accuracy than deduction alone, longer correctness than a static curated release, and a discovery rate that justifies review effort.Source: §7 · p. 11
• Delivery is hypothesized to change user decisions, while the system-level test compares correctness, traceability, and correct failure against the same model given the same content as text.Construction and maintenance costs are to be measured and reported whichever way the comparison falls.Source: §7 · p. 11

8. Scope

• The programme includes a representation profile, vocabularies, qualified-claim authoring, validation, inference, memory, discovery, presentation, evaluation, and domain testbeds.Its work is infrastructure and method built over existing standards, with two candidate vocabularies.Source: §8 · pp. 11–12
• It explicitly excludes new logic or ontology languages, general theories of context or probabilistic combination, general-purpose belief revision, and AGI claims.Interface design beyond traceable presentation is also outside scope.Source: §8 · pp. 11–12

9. Sequence

• The first project combines qualified claims, declared coverage, selected acquisition and revision, and discrepancy detection in one mature semantic domain.Two systems study knowledge bases, ontologies, mappings, and literature; they admit validated claims, maintain coverage accounts, exchange claims, and undergo longitudinal specialist evaluation.Source: §9 · p. 12
• Later work adds executable term bindings, full labelled inference, memory consolidation and decay, hypothesis generation, presentation, institutional exchange, and additional domains.A model trained and constrained on this substrate comes only after the substrate exists at scale.Source: §9 · p. 12
• The sequence maps to G0 for the first project, G1 for testing additional domains, and G2 for exchange between bases.The programme is too large for one project, and the order follows dependencies.Source: §9 · p. 12

10. Risks and honest weaknesses

• The components are established research lines; the claimed contribution is their combination and the dependencies required for competence.The paper accepts the description as nanopublications plus tool use plus coverage vocabulary, but argues that the combination is essential.Source: §10 · p. 13
• The condition schema is the main intellectual risk, while binding scope, coverage granularity, immature memory operations, discovery precision, and evolving standards pose additional risks.Wrong or overly coarse condition dimensions would leave applicability and contradiction handling unimproved; portable fallbacks are specified for mechanisms relying on standards still in progress.Source: §10 · p. 13
• Whether traceable, conditioned, refusable answers justify their construction cost remains an empirical question the first project is designed to test.Source: §10 · p. 13

11.1 Large Knowledge Model

• A Large Knowledge Model uses qualified claims as training and grounding material, emits references to addressable claims or operations, and updates stored claims rather than retraining.Its outputs must resolve to admitted claims, recorded operations, proposals awaiting validation, or refusal; declared coverage supports self-report.Source: §11.1 · p. 14
• An LKM requires layers 1–5 and 7, and remains speculative because useful fluency and competitive training at feasible scale are unresolved.Its tests include the five understanding tests, auditable output resolution, and correct coverage self-report.Source: §11.1 · p. 14

11.2 Semantic Artificial General Intelligence

• SAGI names both a falsifiable position that general competence requires an external semantic substrate and semantic discerning agents that study declared sources and build their own knowledge bases.The position argues that checking, traceable derivation, and delimiting require explicit constraints, addressable premises, and coverage models.Source: §11.2 · pp. 14–15
• Generality is graded from G0 within one base, through G1 across domains and G2 across bases, to unproven G3 open-ended extension.G3 may not be reachable and is not required; a system without an external addressable store passing all five tests, including correct boundary refusal, would refute the necessity claim.Source: §11.2 · p. 15

11.3 SLKM: the agent-built memory

• An SLKM is the addressable knowledge base an agent builds from declared sources, containing what it learns, verifies, annotates, and derives alongside its coverage account.Entries are validated and carry conditions, provenance, and evidence; the structure reorganizes and remains human-inspectable and correctable.Source: §11.3 · p. 16
• The SLKM’s brain analogy is explicitly a design intuition, not a neuroscientific claim, and the terminology should travel with its definitions and mechanisms.The paper says the programme does not depend on the analogy being more than an analogy.Source: §11.3 · pp. 16–17

12. Naming

• SemanticKnowledge names the profile and architecture, while Semantic Knowledge Technologies (SKT) names the broader family.A conformant base is called a SemanticKnowledge base.Source: §12 · p. 17
• LKM, SAGI, and SLKM are reserved for the meanings defined in the relevant section.Source: §12 · p. 17
• In cognitive psychology, “semantic knowledge” refers to general world knowledge in human memory, an unrelated established meaning that motivates preferring the one-word name and family term in titles.Source: §12 · p. 17