Synthetic persona modelling for public-sector decision support

The brief

Sector: Civic / public-sector. Need: the client — a policy-design organisation working with regional governments — wanted a structured way to anticipate how a proposed policy intervention would land across distinct constituencies before committing to public consultation. They had access to rich underlying data (longitudinal surveys, civic registries, voting record aggregates, position datasets from advocacy groups) but no operational way to turn that data into actionable stakeholder analysis at the speed of the policy-design cycle.

The hard constraint: every persona produced by the system had to be defensible against a "where did this come from?" challenge. A persona that influenced a policy decision and could not be sourced was unacceptable.

Workstream decomposition

• WS-1: Source-tier evaluation of persona inputs. Apply Zhianrui's A–E source tiering to every dataset that feeds the persona generation pipeline. Survey data with documented methodology = A; civic registry data = B; advocacy-group position datasets = C with explicit framing-bias notes.
• WS-2: Constrained generation. Build a persona-generation pipeline that takes (a) a policy proposal, (b) a target geography, (c) a demographic slicing, and produces a structured persona dossier — with every claim about the persona's likely position carrying its source row. No claim without a source.
• WS-3: Calibration against held-out adjudicated cases. Run the system retrospectively against historical policy proposals where the actual constituency response is known (post-implementation surveys, voting records on ballot measures). Measure the system's prediction accuracy by demographic slice; document the slices where it under-performs.
• WS-4: Reviewer interface for human escalation. Personas where the system's confidence rating is below a configurable threshold, or where source-tier inputs are predominantly C-or-below, are escalated to a human reviewer before being released into the policy-design workflow.

Method highlight

The non-trivial design choice was to refuse to generate personas without a clear "what would change my mind" section. Each persona dossier includes the specific evidence (e.g., a poll result, a registry statistic) that if it changed, would change the persona's likely position. This is not a novelty — it is the structural mechanism that prevents the system being used as a confirmation-bias generator. A reviewer comparing two personas with opposing positions can immediately see what evidence each is contingent on, and whether that evidence is itself contested.

The calibration work in WS-3 surfaced the system's most important limitation: it under-performed materially on demographic slices where the client's underlying datasets were thinnest (younger voters, recent migrants). This was documented in the open-questions register and presented to the client as a reason to constrain the system's authority on those slices, not as a reason to obscure the limitation.

Deliverable shape

• Persona generation pipeline with source-tier-aware constrained generation.
• Calibration report against ~30 historical policy episodes with held-out outcome data.
• Reviewer interface with configurable escalation thresholds.
• Documentation of the system's known under-performance slices and recommended use constraints.
• Research dossier: 41 source rows, including the source-tier rationale for every input dataset.

Outcomes

The client used the system to inform stakeholder analysis on six policy proposals over the year following deployment. In two of those, the system surfaced a constituency response the client's analysts had not anticipated — both were corroborated in subsequent public consultation. In one case, the system's open-questions register flagged that the source data for a critical demographic slice was too thin to support the persona; the client commissioned additional survey work before proceeding, which revealed a position distribution materially different from the system's first estimate.

The client adopted the source-tier-aware generation approach as a standard for two adjacent decision-support tools they have since built.