Audience construction: lessons from the reference workbench

Joint Claude/Codex synthesis, 25 October 2026. General methods are carried forward from prior segmentation and synthetic-audience work. Source identities, proprietary deck details and original paths remain in the private archive. This is a proposed process, not a claim that assets are cleared for reuse or a Trueness audience has passed validation.

The reference work linked research-based segment descriptions to persistent synthetic characters and a facilitated discussion interface. It demonstrated an audience experience and automated workflow. The inspected material did not establish prospective validity against a new, sealed human holdout. The implementation inheritance review records software patterns; the calibration protocol governs statistical details.

1. Define the decision

Specify the choice, user, population and consequential errors before writing personas. Standard scope: one audience, market, language and task family, with at most two priority subgroups. Reject an unrestricted claim to represent every customer for every question.

Output: signed scope, exclusions, decision thresholds and fieldwork feasibility.

2. Establish rights and provenance

Inventory surveys, interviews, segmentation reports, behavioural observations and public references. Record owner, permitted purpose, applicable consent or other basis, date, population, transformations and deletion requirements. Separate background material from data authorised for model processing.

Access to prior employer/client work does not transfer ownership or permission. General lessons can inform a new implementation; code, research and brand assets require a rights decision. No completed clean-room review is claimed.

Output: source register and permitted-use decision before ingestion.

3. Map constructs and variables

Map sources to the characteristics and outcomes being represented. Inspect missingness, wording, coverage, weights and incompatible definitions. Preserve contradictory evidence rather than smoothing everyone into a stereotype. Separate development inputs from reserved questions and stimuli; a holdout custodian controls reserved evidence.

Output: variable map, gaps and sealed validation specification.

4. Use an approved segmentation

Consume the client's documented segmentation where appropriate. New segmentation is a separate project requiring sample design, clustering choices, stability checks and interpretation. A language model's label does not validate clusters.

Output: eligibility criteria, audience definition and rationale.

5. Treat cross-market matching as a hypothesis

Models can compare segment descriptions, but matching labels or model-generated scores do not establish equivalent distributions or transferable predictions. The standard pilot stays in one market/language; expansion requires its own evidence. Do not pool unlike populations to rescue a failed subgroup.

Output: transfer assumptions and evidence needed for expansion.

6. Assemble an evidence-backed dossier

Document context, needs, behaviour, attitudes, constraints, decision drivers, cultural/media context and category relationships where relevant. Link supported statements to sources; label interpretations and invented illustrative details. Qualitative interviews add meaning but do not estimate population proportions. The standard build's 12 interviews and developmental quantitative sample inform this stage, not the final sealed test.

Output: versioned grounding material with provenance and uncertainty.

7. Separate characters from quantitative inference

A few recurring characters can support exploration. Four convincing voices are generated examples, not a representative sample. Label characters and generated quotations as synthetic. Specify the quantitative configuration separately: weights, aggregation, stochastic repeats and error estimation. Generating more answers does not increase the human sample size.

Output: exploratory and quantitative modes with distinct claims and permissions.

8. Automate the controlled workflow

Automate permitted ingestion, retrieval, questionnaire drafts, runs, status, evidence assembly and reports. Keep accountable human decisions on scope, participant treatment, design and interpretation. Moderator/challenger agents may find weak reasoning; agreement between models is not human validation.

Log provider/version, prompts, grounding, settings, retries and lineage. Use queues and idempotency keys to prevent duplicate studies or charges. Separate customer and research-operator permissions.

Output: reproducible runs and traceability. See SOPs and screen specifications.

9. Freeze before prospective evaluation

Develop and select the configuration using developmental evidence only. Freeze model, prompt, grounding, questionnaire, stimuli, metrics and acceptance rules before opening the holdout.

Per build: 12 developmental interviews plus 400 quantitative participants, divided into 160 developmental and 240 sealed validation respondents. Four validation concepts are separate from four development concepts. Audit: 240 fresh respondents evaluating a frozen audience/output set. Across two builds and one audit: 24 interviews and 1,040 quantitative participants, including 720 validation respondents across three separate engagements—not a pooled validation sample.

At n=240, a simple unweighted proportion near 50% has an approximate 95% margin of error of 6.3 percentage points before design effects or multiplicity. Subgroups are less precise. A proposed 10-point tolerance is a pilot hypothesis, not an industry standard or guaranteed pass. Preserve the person as the sampling unit when each evaluates multiple concepts.

Output: endpoint-level pass, fail or inconclusive findings; permitted scope recorded separately. Tuning on the same holdout cannot turn a failure into a valid independent pass.

10. Release within scope and maintain the record

An external reviewer examines Trueness-built audiences; a separate custodian controls the holdout. These reduce, but do not erase, builder conflicts. Release only supported capabilities; exploratory-only access or no release may be appropriate.

The customer receives a Calibration Record: population, task, dates, sources, version/configuration, human sample, uncertainty, subgroup results, failures, restrictions, changes and expiry. New studies show scope compatibility and whether new human evidence exists.

Monitor model changes, source age, client context and quarterly human pulses. A small pulse can flag drift but does not automatically renew validity. Material changes require a new decision and potentially a fresh holdout and price.

Output: a usable record and maintained service with stop conditions. See templates and the pilot plan.

Trueness retains persistent audience context, structured evidence, facilitated exploration and automated administration, while adding prospective testing, explicit scope and accountable change control. The paid pilot must show that customers value these disciplines enough to fund the human research.