Methodology

1 · Research Design

The programme is sequential and mixed-method. The two phases have different evidentiary status and are not interchangeable.

All Phase 1 material is secondary data. Every source used in the pillars — a regulation, a court judgment, a company disclosure, a benchmark result — already existed, produced by another party for a purpose other than this research. Within that material, sources are ranked by proximity to origin: the regulatory or legal text as published outranks a commentary on it, and a company’s own disclosure outranks a press report of it. Proximity to origin is a measure of source quality; it does not make the material primary data. The only primary data in this programme comes from the Phase 2 interviews.

2 · The Five Pillars — Phase 1

The literature is organised into five pillar studies. Each addresses one component of the applied layer and is written for the senior stakeholder who must act on it.

Source selection and weighting. Evidence is drawn in tiers and weighted accordingly. Tier 1 (load-bearing): peer-reviewed research, the management-science and IT-economics literature, regulatory and legal texts as published (EU AI Act, NIST AI RMF, ISO/IEC 42001), and inter-governmental sources. Tier 2 (supporting): named press, business-school research, and consultancy research with disclosed methodology. Tier 3 (context only): trade press and vendor engineering material. Vendor and self-reported figures are labelled at the point of use and, where they bear analytical weight, corroborated against an independent source.

A series-wide scope principle. Throughout the programme, “the model” is treated as a comparison between Western frontier models (OpenAI, Anthropic, Google) and Far East models (DeepSeek, Qwen, Kimi K2). Where a pillar's evidence base is Western-only, that boundary is stated explicitly rather than left implicit. For non-Western labs the source priority is: lab technical report → model card → peer-reviewed paper → benchmark submission → leadership talk. Western press characterisations of non-Western labs are treated as context only.

Claim grading. Because Phase 1 precedes the fieldwork, no pillar may state a conclusion the fieldwork has not yet tested. Every claim is therefore graded as evidenced (settled in the record), directional (the literature points this way but the finding is not settled), or hypothesis (a proposition Phase 2 exists to test). The grade determines the language the sentence is permitted to use. The programme's own central thesis is graded as a hypothesis.

3 · From Literature to Aims and Objectives

The relationship between the pillars and the objectives is iterative, not linear, and this is deliberate.

• First, a provisional set of aims and objectives was drafted to give the literature review its scope and direction.
• Second, the five pillars were written against that scope, as the literature.
• Third, what the literature actually established was fed back to revise the aims, the objectives and the questionnaire — a plan–do–check–act cycle rather than a one-way sequence.

The pillar text is therefore canonical: where the pillars and an earlier version of the framework disagree on the wording of an objective, the pillars win, because they carry the evidence.

The result is five aims and twenty objectives — one aim and four objectives per pillar. Each objective carries a permanent identifier (P1-O1 through P5-O4). These IDs are frozen: they are the spine of the traceability chain that runs through the questionnaire, the coding frame and the outcomes.

Objectives are framed as exploratory. They examine, characterise and test. They do not instruct the researcher to establish a conclusion decided in advance, because an objective that states its answer cannot be tested by the fieldwork that follows it.

4 · From Objectives to Interview Questions — the Phase 2 Instrument

Each objective is converted into interview questions. The instrument has two tiers, run in a single session.

Question-to-objective mapping. Every question and sub-prompt carries the IDs of the objectives it evidences. The mapping is many-to-many: a single question typically evidences several objectives, and an objective is typically evidenced by several questions. This is how all twenty objectives are covered within a session of realistic length. Coverage has been verified — no objective goes unasked.

Question design. Each granular module opens by asking the respondent what “good” looks like to them, before any maturity or capability question is put. This captures the respondent's own success criteria unprompted, and allows the maturity framework to be tested against their definition rather than imposed on it. Questions are paired: what enabled progress is asked before what was absent when progress slowed. Prompts are open and neutral, and are written to be capable of producing evidence that disconfirms the programme's thesis.

Sampling. Because each granular session goes deep on one pillar, coverage of all twenty objectives comes from the respondent pool rather than from any single interview. Sample composition across the five stakeholder groups is therefore load-bearing, and a minimum respondent count per pillar is set before recruitment.

5 · From Evidence to Outcomes

Interview responses are coded back to the objective IDs they evidence. Each of the twenty objectives yields one outcome: a stated finding, grounded in the field evidence, that answers the question the literature raised.

The coding frame captures both directions. Enabling conditions are coded, and so are their inverse — the conditions absent where progress slowed. Coding both allows later waves to measure movement over time rather than only current-state deficit, which is what makes the programme longitudinal rather than a snapshot.

Findings feed the next wave. Outcomes revise the aims, objectives and instrument for the following cycle, closing the loop that began with the literature.

6 · Scope and Limitations

• Phase 1 is documentary. It reports what the published record shows. It cannot describe what enterprises are actually doing, because it has not asked them. Findings about the field come only from Phase 2.
• Some evidence is modelled or self-reported. Cost figures are derived from public pricing and vendor disclosures rather than audited enterprise accounts. Benchmark results include lab-reported submissions. These are labelled directional at the point of use.
• The public case record is a selected sample. Documented incidents and course corrections entered the public record largely through litigation or press coverage. Cases where controls worked are, by construction, invisible. Phase 2 provides the corrective.
• The interview sample is qualitative and purposive, not representative. It is designed to produce depth across five stakeholder roles, not statistical generalisability to a population.
• This is the first wave. Claims about durability, progression or stall require longitudinal evidence the programme does not yet have. Wave 1 establishes the baseline against which later waves measure movement.

Editorial controls. All editorial changes made during the programme are recorded in a change ledger with the source location, the previous wording, the revised wording and a citation-impact check. No citation, figure, table, date, amount, benchmark result or legal finding has been altered in any editorial pass.