Building the Delphi panel
The panel defines the entire argument space of a study: every argument is one panel member's response to one thesis, so the arguments a study can ever produce are theses × panel members. Design the panel well and its coverage — and its blind spots — are known before any data exists. This guide is the methodological centre of planning a Delphi study in the ecosystem.
Planning (DPE) chooses orientation and drafts the panel. Facilitation (Pronoia) is where the panel is finalised, validated and gated before collection. The panel is a designed epistemic space, not a recruitment convenience.
The panel as a designed epistemic space
A panel is not a random set of experts; its reach and gaps are deliberate and documented. The guiding principles the facilitator works with:
A homogeneous panel (same education, sector, culture) is an epistemic risk and is recorded as such — even if every member is individually excellent. The target is commensurability — enough shared ground for genuine dialogue — not full consensus. Disagreement that is well covered is a feature, not a defect. And watch for lock-in: a single perspective claiming universality, which can masquerade as dialogue simply because it is loud.
The two-axis matrix is primary
The panel's primary structure is a two-axis matrix. Its cells hold the roles that panelist behavior is grounded in. The facilitator ultimately defines the axes and their classification — the matrix is the design surface, and the templates below are only starting points.
The axes are context-dependent. The canonical choices:
| Panel type | Axis 1 (primary) | Axis 2 (secondary) |
|---|---|---|
| Expert panel (N/S/R/D) | Expertise (research / practice / administration) | Stakeholder affiliation (public / private / civil society) |
| Participatory panel (E) | Concerned-party position (resident / civil servant / decision-maker / NGO) | Expertise / theme area |
| Corporate strategy | Concerned-party group (leadership / staff / customers / owners) | Competence area (strategy / finance / operations / market) |
Two properties matter in practice. A cell may hold 0..n panelists — several voices in one role, or none. Empty cells are expected, and a systematically empty cell is a structural blind spot (not a random gap). Naming blind spots before round 1 is part of designing the space; they become candidates for supplementation later.
Orientation sets the panel type
The study orientation chosen in planning (N, S, R, E, D or D+) decides whether the panel is built as an expert matrix or an observer matrix, and which axes, selection logic and minimums apply.
| Orientation | Panel type | Selection basis | Practical requirement |
|---|---|---|---|
| N / S / R / D (intentional) | Expert matrix | Expertise, stakeholder representation, competence domain | At least 3×3; ≥3 stakeholder groups; ≥2 domains; power balance considered |
| E (emergent) | Observer matrix | Observation distance (inside / edge / outside), not expertise | At least 2×3; ≥3 observation distances; ≥3 "outside" roles; no decision-makers; each role names its position and "what it sees" |
| D+ (four-path) | Two parallel panels | Intentional and emergent in parallel | Both requirement sets apply independently |
Emergent studies are about where you stand to watch, not credentials — which is why decision-makers are excluded and observation distance replaces expertise as the selection basis.
Perspective: position + view (Hautamäki)
The optional but recommended perspective layer rests on a simple idea: perspective = position + view. The same phenomenon genuinely looks different from different positions — not because anyone is wrong, but because the phenomenon is multidimensional. The aim is to identify what each position can see and what it necessarily cannot.
Following Hautamäki's viewpoint relativism, the deeper point is the multiplicity of perspectives in general — a panel should span the perspective space, not just tick boxes. The familiar four archetypes (constructivist optimist, empirical realist, critical-conservative, eco-social holist) are one useful special case, not the whole space; add or vary perspectives as the phenomenon demands.
When the layer is active, each role carries three attributes that make its standpoint explicit:
| Attribute | What it captures | Values |
|---|---|---|
| Observation position | Where the member stands relative to the phenomenon | inside / edge / outside / below / above |
| Epistemic type | The kind of knowledge the member brings | formal / experiential / tacit |
| Value basis | The standpoint's underlying values | short free text |
The perspective layer complements the matrix cell: the cell sets what a member speaks to (its role); the perspective sets how it reasons.
AI panels (Route B / C)
A study can run wholly or partly with AI panelists — a full-AI perspective panel makes an end-to-end run possible without recruiting humans, which is valuable for testing the whole process and for human-validated AI-analysis development (AI proposes → human validates → κ/calibration → improve).
An AI panelist sits in a matrix cell (its role) and carries an explicit perspective profile that complements it. Because the AI acts on that description, it must be precise and sufficient — this is enforced by the AI Panelist Description Standard (docs/ai_panelist_description_standard.md): each AI panelist needs a filled cell (expertise × stakeholder), a distinct perspective label, the four Hautamäki dimensions (ontological / epistemological / axiological / temporal) as real descriptions, perspective_axes in [0,1], a bias direction and the model used; across the panel the labels are unique and the perspectives diverse. Route B/C also requires a GenAI disclosure, and humans stay in the loop: every AI response is reviewed before a round closes.
Reading the arguments: the epistemic lens
As arguments arrive, Pronoia reads each one on two dimensions — where the claim draws from (its knowledge base) and what it does to the thesis (its function).
Knowledge base — seven categories: EMP empirical (data, measurement) · KOK experiential (personal/professional, tacit) · NOR normative (values, ethics) · TEO theoretical (models, frameworks) · INT interest-based (stakeholder/strategic position) · EPV uncertainty (acknowledged ignorance) · IMA imaginative (vision, radical reframing).
Function — three categories: TUK supporting (reinforces the thesis) · HAS challenging (questions, weakens, refutes) · AVA opening (introduces a new viewpoint, widens the frame).
Seven knowledge bases × three functions give a 21-cell map of the argument space. Empty cells are the panel's blind spots — viewpoints that never appeared. The Epistemic Coverage Index is how many of the 21 cells are filled; a typical run fills roughly 9–14 of 21, and very low coverage signals a need to add roles that bring the missing viewpoints.
Panel diagnostics (coverage, PDI, lock-in, PHS)
Description quality is per panelist; coverage is a property of the whole panel. The ecosystem computes it deterministically — the computable core of MaxPerspective (M5), aligned with Hautamäki's viewpoint relativism, available live in the new-study wizard and as a tool over any panel:
- Matrix coverage — filled vs empty cells (Absent cell = structural blind spot) and fill rate.
- Dominance & diversity per axis — aspect dominance AD% per category, the dominance class (DOMINANT >50% / STRONG / MODERATE / WEAK / TRACE / ABSENT), and the Perspective Diversity Index PDI = 1 − Σ(ADᵢ/100)² (Herfindahl). Lock-in is flagged when one category exceeds 50%.
- Structural PHS (0–50) —
PDI×25 + (1 − maxAD/100)×25. The full Perspective Health Score also weighs awareness and openness, which are qualitative (the six-phase MaxPerspective analysis) and not inferable from a panel spec. - AI perspective spread — over
perspective_axes, the spread of the AI panel across the perspective space; low spread is perspective lock-in even when the matrix looks covered. - Supplement recommendations — which categories to fill (Absent blind spots), where to reduce concentration, and how to broaden perspective spread.
As a fast human-facing summary the facilitator may also record a PHS quick score (0–12) from four diversity dimensions — position diversity, epistemic diversity, value-basis variation and marginal voices, each 0–3: 0–5 diversity too thin (supplementation required) · 6–8 adequate with gaps (Hautamäki perspective theses recommended) · 9–12 excellent diversity.
Quality gates before collection begins
Before a panel is accepted it must pass orientation-specific gates that turn the principles above into checkable minimums:
| Check | Intentional (N/S/R/D) | Emergent (E) |
|---|---|---|
| Matrix size | ≥ 3×3 | ≥ 2×3 |
| Stakeholder groups / observation distances | ≥ 3 groups | ≥ 3 distances |
| Domains / outside roles | ≥ 2 domains | ≥ 3 outside roles |
| Decision-makers | Power balance noted | Exactly 0 |
| Per-role attribute | Expertise named | Observation position + "what it sees" named |
| Theses | ≥ 3 | ≥ 3 |
If a gate fails, the panel is not yet ready — the facilitator adjusts composition or theses and re-checks. The gates keep coverage and balance honest before any arguments are gathered.
Building the panel in practice
Ready-to-run templates give a validated starting point for each panel type — Classic foresight, Argument Delphi, Barometer, Participatory civic, Real-Time, Corporate strategy and a full-AI perspective panel — each a contract-conformant package you instantiate in one step (from the study wizard or by importing a planning package) and then edit.
A working sequence for the facilitator: pick a template (or start blank) → set the two axes for your phenomenon → place roles in the cells (leaving deliberate blanks) → for AI panels, write each perspective profile to the standard → read the live diagnostics (PDI, lock-in, blind spots, structural PHS, AI spread) and act on the recommendations → pass the quality gates → begin collection.
The panel is where a Delphi study's epistemic reach is decided. Everything downstream — the arguments, the analysis, the scenarios — is bounded by the space you design here.