Persona Validation

# Persona Validation Methods

Purpose: Define how Cast validates personas with triangulation, survey evidence, clustering, and staged confidence upgrades.

## Contents

1. Why validation matters
2. Triangulation patterns
3. Survey thresholds
4. Clustering guidance
5. Synthetic persona rules
6. Validation statuses

## Why Validation Matters

Validation exists to avoid:

- proto-personas treated as facts
- confirmation bias from creators
- over-generalizing from too little data
- stale personas surviving market or product change

## Triangulation Patterns

| Pattern | Methods | Strength |
|---|---|---|
| Basic | interview `5-10` + survey `350+` | Cost-efficient |
| Behavioral | interview + behavior logs + experiment / test evidence | Verifies saying vs doing |
| Full | interview + survey + behavior logs + usability evidence | Highest confidence |

## Quantitative Survey Thresholds

- `350+` respondents per segment for `95%` confidence
- `1000+` total when comparing multiple segments
- Prefer behavior-based questions over preference-only questions
- Likert + free text is the default hybrid

## ML Clustering Guidance

### Algorithm Fit

| Algorithm | Use when | Caveat |
|---|---|---|
| K-means | clearly separated segments | requires preselected cluster count |
| DBSCAN | irregular clusters and outliers | parameter-sensitive |
| Hierarchical clustering | exploratory structure analysis | weak for very large datasets |
| Gaussian mixture | overlapping segments | higher computational cost |

### Cluster Count Rules

- Use `Elbow + Silhouette + Gap` together.
- `Silhouette > 0.5` is a good signal.
- Recommended persona count is `3-7`.
  - early product: `3-4`
  - mature product: `5-7`

## Validation Workflow

1. Collect behavior, survey, support, and satisfaction data.
2. Preprocess and normalize.
3. Cluster with more than one method when possible.
4. Match clusters against current personas.
5. Treat uncovered clusters as new persona candidates.
6. Raise confidence only after evidence-backed mapping.

## Synthetic Persona Rules

- Synthetic personas are hypothesis tools, not production truth.
- Use them to improve guides, expose gaps, or explore edge cases.
- Never treat them as substitutes for real user validation.
- Keep synthetic and real-data-backed personas explicitly separated.

## Validation Statuses

| Status | Meaning |
|---|---|
| `proto` | hypothesis only |
| `partial` | validated by one stream only |
| `validated` | triangulated |
| `ml_validated` | supported by clustering evidence |

### Confidence Contributions

| Validation state | Contribution |
|---|---|
| Proto baseline | `0.30` |
| Interview validation | `+0.20` |
| Survey validation | `+0.15` |
| ML validation | `+0.20` |
| Triangulation complete | `+0.10` |

## Anti-Patterns

Common persona failure modes to detect and avoid during creation, maintenance, and organizational rollout.

| ID | Name | What goes wrong | Mitigation |
|---|---|---|---|
| `PA-01` | Demographics Fixation | Persona is mostly age/gender/job labels | Anchor on goals, pain points, and behaviors |
| `PA-02` | Single Monolithic Persona | One persona tries to represent everyone | Keep at least `P0/P1/P2` by default |
| `PA-03` | Happy Path Persona | Only ideal users are represented | Include friction-heavy or underserved users |
| `PA-04` | Proto-Persona Ossification | Hypotheses are treated as stable truth | Keep validation status explicit |
| `PA-05` | User-Buyer Conflation | Buyer and end user are merged | Split if goals or behaviors differ materially |
| `PA-06` | One-Shot Creation | Persona is created once and never updated | Use `AUDIT` and `EVOLVE` regularly |
| `PA-07` | Over-Designed Artifact | Persona looks polished but is weakly evidenced | Favor evidence density over visual polish |
| `PA-08` | Specificity Imbalance | Too vague or too fictional | Keep roughly 80% evidence / 20% inference |
| `PA-09` | Silo Creation | Persona is not shared or reusable | Register and distribute systematically |
| `PA-10` | Gallery Display | Persona exists as decoration only | Tie personas to downstream agent tasks |

### Persona Fatigue

Causes: too many personas, stale personas, personas not used in real decisions, overly repetitive artifacts. Mitigation: keep count manageable, deprecate stale personas, track downstream use, distribute task-specific versions.

### Anti-Persona

Use anti-personas to define who the product should not optimize for. Identify mismatched segments, document why out-of-scope, record cost/risk, keep separate from primary personas, revisit during strategy shifts.