KF-L001-B05 · Foundation Models and the Knowledge Sciences

Testing the Assumptions of Knowledge Systems

Evidence · Editorial Review · v0.4 · 27 min read

Question  →  Corpus  →  Evidence  →  Patterns  →  Candidate Insight

Purpose

This chapter critically examines whether foundation models fundamentally alter the assumptions underlying knowledge engineering, collective knowledge capability and knowledge governance as surveyed in B02, B03 and B04.

It is not a survey of AI technologies. It is an interrogation of the assumptions about knowledge systems that recent advances in artificial intelligence are alleged to have overturned.

Applying the common analytical methodology used throughout the Literature Landscape, the chapter evaluates which assumptions remain valid, which require modification, which become more or less important, and which genuinely new capabilities have emerged.

Its purpose is not to determine whether AI changes everything, but to determine precisely what it changes — and, equally importantly, what it does not.

Executive Summary

Foundation models have transformed artificial intelligence from a set of narrowly specialised systems into broadly capable computational partners. Knowledge can now be generated rather than only retrieved; reasoning can be approximated in natural language; and complex workflows can increasingly be delegated to semi-autonomous systems.

These developments have prompted a strong claim: that earlier approaches to knowledge engineering, knowledge management and governance are now obsolete.

This chapter evaluates that claim and finds it overstated. Examined assumption by assumption, foundation models do not eliminate the engineering, stewardship and governance capabilities matured by the traditions surveyed earlier in the Landscape. They relocate those capabilities across a new boundary — between symbolic representations, which are explicit and inspectable, and parametric representations, which are distributed and opaque.

Read through the analytical spine of the Landscape, the finding is continuous rather than revolutionary. B03 argued that every knowledge system governs representations, not knowledge itself; foundation models introduce a genuinely new representation, not a new knowledge. B04’s six Governance PrimitivesGovernance Primitive — one of the six operations by which warrant is conferred: evidence, authority, provenance, validation, versioning, reversibility (KF-1104). Vocabulary → — evidence/justification, authority, provenance/traceability, validation/quality, versioning/release, reversibility/audit — remain necessary under foundation models, but several of them (provenance and reversibility above all) become markedly harder to instantiate in a parametric substrate.

Consistent with the Literature Landscape methodology, whether this relocation constitutes a genuinely new paradigm, or an evolutionary extension of the traditions already surveyed, is reserved for C01; this chapter does not argue for any candidate methodology.

Orientation for the Standalone Reader

The three preceding evidence chapters examined knowledge from complementary perspectives.

B02 — The Evolution of Knowledge Engineering — traced how the object of engineering progressively expanded across successive generations, and identified a candidate gap around the cross-layer governance of knowledge change.

B03 — Knowledge Beyond Engineering — examined how knowledge becomes enduring collective capability across eight adjacent traditions, and introduced the Landscape’s recurring analytical lens: Scale → Capability → Time.

B04 — Governing Knowledge Change — examined how knowledge change becomes legitimate, and distilled six Governance Primitives together with the notion of a division of governance labour across traditions.

This chapter asks a different kind of question. Rather than surveying a further tradition, it tests whether the emergence of foundation models overturns the conclusions the previous chapters reached. The unit of analysis is therefore not a tradition but an assumption — a deliberate structural pivot explained in Section 4.

1. Introduction

Foundation models are among the most consequential developments in the history of artificial intelligence. Trained at scale on broad data and adaptable to many downstream tasks, they exhibit broad linguistic competence, in-context and few-shot learning, and behaviours often described as emergent reasoning (Bommasani et al., 2021; Brown et al., 2020).

These capabilities have prompted claims that traditional approaches to representing, organising and governing knowledge are no longer necessary. If a model can answer a question directly, why curate an ontology? If it can generate an explanation, why maintain a document repository? If it can act through tools, why engineer a workflow?

At the same time, persistent concerns — hallucination, opacity, provenance loss, model drift, and the difficulty of correction — suggest that many longstanding problems remain unresolved and, in some respects, are aggravated (Ji et al., 2023; Bender et al., 2021).

The purpose of this chapter is therefore to separate genuine discontinuities from evolutionary extensions of the traditions already surveyed. It approaches the question with the restraint applied throughout the Landscape: strong claims of novelty are treated as hypotheses to be tested against the literature, not as premises.

2. Central Research Question

The chapter is organised around a single guiding question.

What fundamentally changes under foundation models?

Supporting questions: Which long-held assumptions remain valid? Which require revision? Which become more important? Which become less important? Which genuinely new capabilities emerge? Which new governance challenges appear? Does AI change the nature of knowledge itself — or only its representation?

The final question is stated deliberately in its sharpened form. Section 6 returns to it directly, and the answer the chapter defends is anticipated by B03: knowledge systems govern representations, not knowledge; foundation models supply a new representation, not a new epistemology.

3. Analytical Lens

Each assumption is examined through six recurring questions. This is a chapter-specific lens, purpose-built for stress-testing assumptions; it does not displace the Landscape-wide Scale → Capability → Time device from B03. Section 6 reconnects the two explicitly, and Assumption Six is analysed using B04’s six Governance Primitives so that B05 remains commensurable with the chapters around it.

The “Relevance to Governed Knowledge Evolution” row records burden-of-proof implications only. It states what any AI-native knowledge methodology would have to account for; it does not claim that any particular methodology accounts for it. That argument is reserved for C01.

        The KF-L001 Method
    ───────────────────────────
         Research Question
                ↓
              Corpus
                ↓
          Analytical Lens
                ↓
        Comparative Analysis
                ↓
           Cross-patterns
                ↓
     Candidate Integrative Insight

Figure B05·1 — the recurring KF-L001 review method, applied uniformly across every evidence chapter of the Landscape.

4. Corpus of Investigation

The corpus consists not of technologies but of eight assumptions about knowledge systems that foundation models are said to challenge.

On the pivot from traditions to assumptions. B02–B04 each surveyed eight traditions. B05 surveys eight assumptions. The change is deliberate. The prior chapters asked what a mature field knows; this chapter asks what a field’s inherited commitments can still be relied upon to be true. Because foundation models cut across every tradition at once, an assumption-based corpus keeps the analysis integrative rather than fragmenting it into eight parallel technology reviews. The representative literature under each assumption is illustrative of where the assumption is contested, not an exhaustive survey.

4.1 Assumption One — Knowledge must be explicitly represented

Representative literature: symbolic AI and knowledge representation (Brachman & Levesque, 2004); foundation models (Bommasani et al., 2021); studies of emergent internal structure and probing (Tenney et al., 2019).

Central question: Can implicit parametric representations replace explicit symbolic knowledge?

Traditional assumption. From expert systems onward, usable knowledge had to be made explicit — encoded in rules, frames, ontologies or graphs — so that it could be inspected, reused and reasoned over.

AI development. Foundation models store knowledge implicitly, distributed across billions of parameters. No sentence, triple or rule is written down, yet the model can produce fluent, often accurate, knowledge-bearing outputs across domains.

Continuing validity. Explicit representation remains indispensable wherever knowledge must be inspected, audited, contested or guaranteed. Parametric knowledge cannot be pointed at; a regulator, auditor or downstream system still requires an explicit artefact. B03’s principle holds without amendment: the system governs a representation, and a parametric weight matrix is simply a new — and unusually opaque — kind of representation.

New capability. Broad coverage without hand-authoring. A single model exhibits competence across domains that would each have required a dedicated, laboriously engineered knowledge base.

Remaining limitations. Implicit knowledge is not addressable, not selectively editable, and not reliably separable from error. Probing shows structure exists internally, but that structure is not an interface. Explicit and implicit representation are therefore complementary, not substitutes.

Relevance to governed knowledge evolution. Any AI-native methodology must operate across both explicit and implicit representations simultaneously and maintain a defensible mapping between them. Representation becomes plural, not obsolete.

4.2 Assumption Two — Retrieval is the primary mechanism for accessing knowledge

Representative literature: information retrieval (Manning et al., 2008); dense retrieval (Karpukhin et al., 2020); retrieval-augmented generation (Lewis et al., 2020); enterprise search practice.

Central question: Does generation fundamentally replace retrieval?

Traditional assumption. To access knowledge, a system locates and returns the relevant existing artefact. Access is finding.

AI development. Generation offers a second mode: the system composes an answer rather than returning a document. Retrieval-augmented generation (RAG) then re-couples the two — grounding generation in retrieved evidence at inference time.

Continuing validity. Retrieval does not merely survive; it is promoted. The most reliable AI knowledge architectures are retrieval-augmented precisely because generation alone lacks grounding and provenance. Retrieval supplies the evidentiary substrate that generation cannot.

New capability. Synthesis across sources, and answers to questions for which no single artefact exists. Generation interpolates where retrieval can only return.

Remaining limitations. Ungrounded generation fabricates; grounded generation inherits the quality and coverage limits of its retrieval layer. Generation shifts, but does not remove, the burden of having good sources — and it introduces a new failure mode in which fluent output masks an absent or misused source.

Relevance to governed knowledge evolution. The dichotomy of “retrieval vs generation” is false. The methodological requirement is to govern the coupling — to ensure generated claims remain traceable to retrieved, authorised evidence. This is the provenance primitive from B04 restated under generation.

4.3 Assumption Three — Reasoning requires explicit symbolic structures

Representative literature: chain-of-thought prompting (Wei et al., 2022b); ReAct (Yao et al., 2023); tool use (Schick et al., 2023).

Central question: What kind of reasoning emerges from foundation models?

Traditional assumption. Reasoning is the manipulation of explicit symbols under formal rules — deductive, sound, and inspectable.

AI development. Foundation models exhibit approximate, language-mediated reasoning: chain-of-thought decomposition, tool invocation, and interleaved reasoning-and-acting produce answers that look like multi-step inference without an explicit symbolic engine.

Continuing validity. Where soundness, verifiability or guarantees are required, explicit symbolic and formal methods remain necessary. Parametric reasoning is plausible, not provable; it is a heuristic approximation, and its steps are rationalisations that may not correspond to the computation that produced the answer.

New capability. Flexible, general-purpose reasoning over natural-language problems without task-specific symbolic engineering — a breadth classical methods never achieved.

Remaining limitations. No soundness guarantee; sensitivity to prompt phrasing; confidently wrong chains. The reasoning is probabilistic and non-deterministic, which sits uneasily with any domain requiring reproducibility. This motivates neuro-symbolic hybrids (Assumption Eight).

Relevance to governed knowledge evolution. Reasoning becomes a spectrum from probabilistic to formal. A methodology must place work on that spectrum deliberately, and must not treat plausible generation as though it carried the guarantees of deduction.

4.4 Assumption Four — Knowledge systems are largely static

Representative literature: catastrophic forgetting and continual learning (French, 1999; Kirkpatrick et al., 2017); model editing (Meng et al., 2022; Mitchell et al., 2022).

Central question: Can parametric knowledge evolve deliberately?

Traditional assumption. A knowledge base changes when a human deliberately edits an identifiable artefact; between edits it is stable and its state is known.

AI development. Parametric knowledge changes through retraining, fine-tuning and targeted model-editing techniques (e.g. locating and rewriting factual associations). Change is now possible inside the model, not only in the documents around it.

Continuing validity. Deliberate, addressable, reversible change remains the goal — and remains far easier over explicit artefacts than over weights. B04’s entire apparatus of versioning, release and reversibility presupposes that one can name a state and return to it; parametric substrates make that presupposition hard to satisfy.

New capability. Direct intervention on internal knowledge, and continual adaptation without full re-authoring.

Remaining limitations. Edits interact unpredictably, may not generalise, and can regress unrelated capabilities; catastrophic forgetting persists. Crucially, parametric knowledge is difficult to version and roll back with confidence — the state before an edit is not cleanly recoverable. Change became possible before change became governable.

Relevance to governed knowledge evolution. This is the sharpest point of contact with B02’s candidate gap and B04’s primitives. The controlled transition of a knowledge system from one justified state to another — B02’s framing of the gap — is precisely what parametric substrates make difficult. Any AI-native methodology must deliver versioning and reversibility over representations that natively resist both.

4.5 Assumption Five — Humans remain the primary knowledge workers

Representative literature: human–AI interaction (Amershi et al., 2019); hybrid intelligence (Dellermann et al., 2019); generative and LLM agents (Park et al., 2023).

Central question: How does AI redistribute cognitive labour?

Traditional assumption. Humans create, curate, judge and are accountable for knowledge; machines assist.

AI development. Foundation models and agentic systems now perform substantial portions of drafting, synthesis, retrieval and multi-step task execution, shifting the human toward specification, review and adjudication.

Continuing validity. Accountability, judgement under ambiguity, and the authority to authorise a knowledge state remain human. B04’s authority primitive does not transfer to a model: a system may generate a claim, but legitimation of that claim as organisational knowledge is an act of authority that remains with people and institutions.

New capability. Delegation of cognitive labour at scale, and continuous rather than episodic knowledge production.

Remaining limitations. Automation bias, deskilling, diffusion of responsibility, and the difficulty of holding a probabilistic system accountable. Redistribution of labour is not redistribution of accountability.

Relevance to governed knowledge evolution. Cognitive labour is redistributed; governance authority is not. A methodology must define the human–AI division of labour while keeping the division of governance labour (B04) anchored to identifiable human authority.

4.6 Assumption Six — Governance applies primarily to symbolic knowledge

Representative literature: Model Cards (Mitchell et al., 2019); Datasheets for Datasets (Gebru et al., 2021); NIST AI Risk Management Framework (NIST, 2023); ISO/IEC 42001 (2023); model risk governance (e.g. SR 11-7).

Central question: Can existing governance mechanisms govern parametric knowledge?

This assumption is analysed directly against B04’s six Governance Primitives, so that B05’s most governance-heavy section speaks the Landscape’s shared vocabulary.

• Evidence / Justification. Partially transferable. RAG and citation restore some evidentiary grounding, but a parametric answer’s true justification (its training-data support) is not recoverable. Weakened.
• Authority. Transferable in principle. Authority remains human and institutional; what changes is the object being authorised (see Assumption Five). Preserved.
• Provenance / Traceability. Severely strained. Training-data provenance is generally lost, and outputs cannot be traced to the specific inputs that produced them. This is the primitive most damaged by the parametric shift. Most weakened.
• Validation / Quality. Transferable but harder. Evaluation shifts from artefact inspection to behavioural/statistical testing; correctness becomes distributional rather than checkable per item. Adapted.
• Versioning / Release. Partially transferable. Model versions can be released, but fine-grained knowledge-item versioning within a model is not yet tractable (see Assumption Four). Partially weakened.
• Reversibility / Audit. Severely strained. Rolling back a specific piece of learned knowledge, or auditing why a specific output occurred, is largely unsolved. Most weakened.

Continuing validity. Governance is more necessary under foundation models, not less — the classic error this chapter corrects. Model cards, datasheets, the NIST AI RMF and ISO/IEC 42001 exist precisely because the primitives do not transfer for free.

Remaining limitations. Two primitives — provenance and reversibility — presently lack robust parametric instantiations. Governance frameworks for AI are emerging but immature, and largely procedural rather than technical.

Relevance to governed knowledge evolution. The burden is explicit: an AI-native methodology must instantiate all six primitives across both symbolic and parametric knowledge, with particular attention to the two that the substrate most resists. B04’s guiding principle carries over intact — knowledge systems remain trustworthy only because their changes are governed — and becomes harder, not obsolete.

4.7 Assumption Seven — Organisational memory resides primarily in documents

Representative literature: organisational memory (Walsh & Ungson, 1991); knowledge graphs (Hogan et al., 2021); vector databases and dense indexes (Johnson et al., 2019).

Central question: Where does organisational knowledge now reside?

Traditional assumption. The institution’s memory lives in documents and records, curated and preserved as explicit artefacts.

AI development. Memory is now distributed across documents, knowledge graphs, vector indexes and model parameters. Some organisational knowledge is effectively encoded into fine-tuned or adapted models.

Continuing validity. Durable, authentic, preservable records remain the backbone of organisational memory — the archival and LIS concerns from B03 do not lapse. A model is a poor system of record: it cannot attest to what it once “knew,” and cannot be preserved as evidence of a decision.

New capability. Memory that is queryable by meaning (embeddings) and partially internalised into behaviour (parameters), not only by document lookup.

Remaining limitations. Parametric memory is unstable across model updates, cannot be authenticated, and risks silent loss when a model is replaced. Memory that lives only in weights is memory the organisation cannot guarantee it still has.

Relevance to governed knowledge evolution. Organisational memory becomes multi-substrate. A methodology must preserve authentic records as the anchor while accommodating graph, vector and parametric memory — and must prevent model turnover from silently erasing institutional knowledge.

4.8 Assumption Eight — Knowledge Engineering becomes less important

Representative literature: neuro-symbolic AI (d’Avila Garcez & Lamb, 2023); GraphRAG (Edge et al., 2024); knowledge graphs (Hogan et al., 2021); enterprise knowledge systems.

Central question: Does AI reduce or increase the need for knowledge engineering?

Traditional assumption. If models can generate knowledge, the laborious engineering of ontologies, graphs and pipelines becomes unnecessary. The folk syllogism is simple: more AI, less engineering. It is this assumption, more than any other, that underwrites the claim that foundation models render previous knowledge traditions obsolete — and it is the assumption the evidence contradicts most directly.

AI development. In practice the opposite emerges. Reliable AI systems are overwhelmingly hybrid. Knowledge graphs are used to ground and constrain generation, so that a model answers from a curated structure rather than from its own unaudited parameters (GraphRAG). Neuro-symbolic approaches pair parametric flexibility with symbolic guarantees where correctness must be assured. And around the generative core sits an increasingly elaborate engineered apparatus: retrieval indexes and chunking strategies, grounding graphs and schemas, evaluation harnesses and test sets, guardrails and validators, orchestration layers, and agent scaffolding that decomposes tasks and checks intermediate results. The generative model is the cheap component. The reliable system around it is engineered, and engineered heavily.

Continuing validity. Knowledge engineering does not disappear; it moves. The task shifts from authoring knowledge artefacts by hand toward designing and orchestrating the system in which a generative core can be trusted. The skill that once produced an ontology now produces the grounding graph, the retrieval design, the evaluation regime and the governance wrapper. This is not less engineering; it is engineering relocated up a level of abstraction.

New capability. Less hand-authoring of content; far more architectural engineering — the deliberate composition of symbolic and parametric components into a system whose behaviour is more reliable than either part alone.

Remaining limitations. The hardest problems — grounding, provenance, evaluation, controlled change — are engineering problems, and foundation models make them more prominent, not fewer. Removing the model’s need for a hand-built knowledge base did not remove the system’s need for engineering; it merely moved the difficulty from content to architecture, where it is less visible and often underestimated.

The counter-narrative, stated plainly. The popular chain runs in one direction:

AI → less engineering

Figure B05·2 — the popular (and unsupported) inference.

The evidence supports a different chain:

AI → different engineering → more architecture → more governance → more orchestration

Figure B05·3 — the chain the evidence actually supports.

Each link is load-bearing. The engineering changes because the object is now a hybrid system rather than a symbolic artefact. It requires more architecture because reliability is a property of the composition, not of the model. It requires more governance because a probabilistic, opaque, continually changing core must be constrained by exactly the primitives B04 identified. And it requires more orchestration because the work is no longer to state knowledge once but to coordinate retrieval, generation, validation and human authority at run time, on every request. The naïve expectation — that capability at the centre would dissolve the need for structure around it — inverts what practice shows.

Bridge to B02. This is where B05 returns to B02 most directly, and deliberately so. B02 read the history of Knowledge Engineering as a progressive expansion of the object of engineering — from rules, to knowledge bases, to ontologies, to knowledge graphs, to the socio-technical systems that surround them. If the object has expanded once more — from knowledge artefacts to Hybrid Symbolic–Parametric Knowledge Systems that are simultaneously symbolic and parametric — then foundation models are not the end of knowledge engineering but its latest enlargement. That reading is available, and Assumption Eight makes it visible. What Assumption Eight does not do is settle it. Consistent with the Literature Landscape methodology, whether this enlargement constitutes a genuinely new stage in B02’s trajectory, or a redistribution of work within the existing one, is reserved for C01. B05 records the shift and its direction; it does not pronounce on its status. The restraint is intentional: the observation that engineering increases under foundation models is well-supported by the literature, whereas any claim that this marks a new epoch is an interpretation that must survive C01, not a premise B05 is entitled to assert.

Relevance to governed knowledge evolution. The burden this places on any AI-native methodology is architectural, not merely representational. It is not enough to represent knowledge, symbolic or parametric; the methodology must engineer the coupling, grounding, evaluation and governance that make a generative core trustworthy — and it must do so as a first-class concern rather than a wrapper bolted on afterward. If B02’s throughline is right that the object of engineering keeps expanding, then the demand foundation models place on knowledge engineering is the largest expansion yet: from artefact, to system, to governed hybrid system.

5. Comparative Analysis

The following matrix summarises how each assumption fares under foundation models. The Effect column uses a common scale: Holds (unchanged), Holds+ (more important than before), Revised (true but modified), Strained (partially undermined), Superseded (no longer primary).

No assumption is straightforwardly abolished. The dominant effects are Revised and Holds+ — the signature of redistribution rather than replacement.

6. Cross-Cutting Patterns

Reading down the matrix, a small number of tensions recur:

• explicit vs implicit representation;
• symbolic vs parametric knowledge;
• retrieval vs generation;
• deterministic vs probabilistic reasoning;
• human vs AI cognition;
• governance across hybrid systems.

In every case the pattern is the same: the capability is not eliminated but relocated across a new boundary. This is where B05 rejoins the Landscape’s shared lens.

Through Scale → Capability → Time (B03). Foundation models act most visibly on the Capability axis (what a system can do expands dramatically) and most problematically on the Time axis (governing change, provenance and reversibility over parametric substrates is unsolved). On the Scale axis they lower the cost of broad coverage while raising the cost of guarantees. The lens thus locates AI’s contribution precisely: large capability gains, unresolved temporal governance.

The sharpened question — does AI change the nature of knowledge itself? The chapter’s answer is no, and B03 supplies the reason: every knowledge system governs representations, not knowledge itself. Foundation models introduce a new representation — distributed, parametric, opaque — and thereby change the mechanics of engineering, retrieval, reasoning and governance profoundly. They do not change the underlying relations that make something count as knowledge (that it be justified, authorised, preserved, situated). Claims that AI changes “the nature of knowledge” mistake a change of representation for a change of kind.

The substrate changes; the epistemic obligations do not.

This proposition is offered as a recurring one for the Landscape — a companion to B03’s every knowledge system governs representations, not knowledge itself — and is the chapter’s most compact answer to the sharpened question. Later chapters (B01 on the object of knowledge, C01 on synthesis) may refer back to it directly: whatever new substrate a knowledge system adopts, the obligations to justify, authorise, preserve and govern its representations are unchanged. The synthesis names this directly: the human, symbolic, parametric and hybrid modes catalogued in this chapter are what C01 formalises as Computational Modality — the cross-cutting axis along which the operations of a knowledge system are redistributed, never removed.

6.1 Limitations of This Review

The review is constrained by the rapid pace of AI development, unstable terminology, limited longitudinal evidence, an emphasis on scholarly over commercial literature, and a deliberately selective corpus. Several cited capabilities are recent enough that their durability is not yet established — and the literature itself contests some of them (for instance, whether “emergent abilities” are real discontinuities or artefacts of measurement; Schaeffer et al., 2023). The chapter should be read as an interpretive synthesis of a moving field, not a settled account.

7. Existing Contributions

Setting interpretation aside, several AI capabilities can already be regarded as established contributions to knowledge systems:

• general-purpose parametric representations as broadly adaptable substrates;
• retrieval-augmented generation as a grounding architecture;
• agentic workflows and tool use;
• model editing as targeted parametric change;
• hybrid symbolic–parametric systems (neuro-symbolic, GraphRAG);
• an emerging AI-governance apparatus (model cards, datasheets, NIST AI RMF, ISO/IEC 42001);
• multimodal reasoning.

This inventory records what has demonstrably emerged, prior to any claim about its broader significance.

8. Candidate Integrative Insight

Advanced only after the evidence above, and marked as interpretation reserved for confirmation in C01.

Foundation models do not invalidate the mature capabilities developed by the traditions surveyed in B02–B04. They relocate those capabilities across the symbolic–parametric boundary:

• Knowledge becomes simultaneously symbolic and parametric (Assumptions 1, 7).
• Retrieval and reasoning are re-coupled to generation, not replaced by it (Assumptions 2, 3).
• Governance becomes more important, and unevenly transferable: authority and validation adapt, while provenance and reversibility are the primitives the parametric substrate most resists (Assumptions 4, 6).
• Engineering shifts from constructing artefacts toward orchestrating hybrid systems (Assumption 8).

The literature therefore suggests that foundation models do not replace previous knowledge traditions. Rather, they increase the interdependence between them. Engineering (B02), organisational capability (B03) and governance (B04) become more tightly coupled, because Hybrid Symbolic–Parametric Knowledge Systems require all three simultaneously: a generative core must be engineered into a system, embedded in organisational practice, and governed across its continual change, or it cannot be relied upon at all. Where the previous chapters could examine their traditions largely in parallel, foundation models force them into a single, mutually dependent stack.

The candidate insight can be stated in one line: foundation models change where knowledge capabilities reside, not whether they are needed. Consistent with the Literature Landscape methodology, whether this relocation amounts to a genuinely new paradigm, or an evolutionary continuation of the traditions already surveyed, is reserved for C01.

9. Burden of Proof

Consistent with the Literature Landscape methodology, this chapter defines evidence, not theory. It establishes what any methodology claiming to support AI-native knowledge systems would have to demonstrate, structured around the three standard questions.

Established (what is mature). The AI period has demonstrably delivered new capabilities: parametric representation, retrieval-augmented generation, agentic tool use, model editing, hybrid symbolic–parametric systems, an emerging AI-governance apparatus, and multimodal reasoning (Section 7).

Unresolved (what remains open). No assumption is abolished; capabilities are relocated across the symbolic–parametric boundary, and two Governance Primitives resist that relocation. Such a methodology would need to show how it:

1. governs symbolic and parametric knowledge together;
2. validates AI-generated knowledge against authorised evidence;
3. maintains provenance across hybrid systems, including where training-data provenance is lost;
4. supports deliberate, continual revision of parametric knowledge;
5. coordinates human and AI judgement while keeping authority human;
6. governs autonomous agents’ contributions to the knowledge state;
7. preserves organisational memory despite continual model evolution;
8. explains how evidence quality is evaluated when generation exceeds direct retrieval.

Implication (what the synthesis must explain). Mapped onto B04’s primitives, requirements 3 and 4/7 fall hardest on provenance and reversibility/versioning — the sharpest test any candidate methodology will face — and the newly added requirement 8 falls on validation/quality: once a system routinely answers beyond what any single retrieved source states, evaluating the quality of the evidence behind a generated claim becomes a distinct problem from evaluating retrieval, and one that C01 will need to address directly.

10. Relationship to Previous and Subsequent Chapters

• B02 examined the engineering of knowledge systems and the expanding object of engineering. B05 supplies the AI-era data point for that trajectory (Assumption 8) without adjudicating whether it is a new stage.
• B03 examined how knowledge becomes collective capability, and gave the Scale → Capability → Time lens B05 reuses in Section 6.
• B04 examined how knowledge change becomes legitimate, and gave the six Governance Primitives B05 applies in Section 4.6.
• B01 (object) and C01 (synthesis) remain. The overall gap, and the question of whether AI redistribution constitutes a new paradigm, are deferred to C01.

Rather than replacing the earlier traditions, AI is evaluated here as a potential reconfiguration of their established principles.

11. Guiding Principle

The organising principle is intentionally simple.

Foundation models change where knowledge resides. They do not remove the need to engineer, steward or govern it.

They require those activities to operate simultaneously across symbolic, organisational and parametric forms of knowledge — and they make two of B04’s Governance Primitives, provenance and reversibility, materially harder to honour. Or, in the single line this chapter proposes to carry forward:

The substrate changes; the epistemic obligations do not.

Having examined how AI changes knowledge systems, the Landscape can now ask the deeper question that underlies every preceding chapter: what exactly is the object that these systems manipulate? That is the subject of B01.

Editorial Note (v0.4)

Change from v0.3: one sentence added at the end of §6, immediately after the discussion of B05’s recurring line, naming the chapter’s human/symbolic/parametric/hybrid finding as what C01 formalises as the Computational Modality axis. The addition extends an existing forward-reference to B01/C01 rather than introducing a new one, so the evidence-then-defer discipline is preserved: B05 still records its finding and defers adjudication to the Synthesis. No other content changed. References remain inline pending the v1.0 reference/DOI pass applied to the finalised chapters.