# Methodology

## Research posture

PIXIE uses participatory, accessibility-first design. People help define the problem, test proposed interactions, inspect collected records, correct interpretations, and decide whether their experience may be published.

No diagnosis is required to participate. Neurodivergent participants are contributors to the design—not objects from whom behavior is silently extracted.

## Evidence collection

### 1. Context interview

Ask the participant what they create, which devices and applications they use, what they need to find later, what they fear losing, and what kinds of assistance feel intrusive.

### 2. Consented task observation

Observe a bounded task such as saving a recording, locating a reference file, resuming an interrupted project, or distinguishing a final from a draft.

Record task events—not inferred feelings:

- prompts presented;
- choices made;
- repeated steps;
- reversals and corrections;
- pauses or abandonment;
- time to complete;
- requests for help; and
- the participant's explanation, if they wish to give one.

### 3. Participant interpretation

Show the event record to the participant. Ask what the system misunderstood and which moments felt helpful, confusing, tiring, or irrelevant. The participant's interpretation remains distinct from the researcher's interpretation.

### 4. Prototype comparison

Compare the existing workflow with a PIXIE-assisted workflow. Use the same task class, not necessarily the same file. Measure completion, recovery, error correction, and participant control.

### 5. Follow-up

Ask whether the learned rules remain useful after real use. Make it easy to reset, export, or withdraw them.

## Minimum viable study

- 8–12 independent creators across at least three disciplines.
- Include participants with different access needs and people who do not identify as neurodivergent.
- Test on-device workflows before attempting cross-device automation.
- Run at least two sessions per participant so learning and recovery can be evaluated.
- Do not present results as representative prevalence data.

## Data minimization

Collect the smallest event record that answers the research question. Do not capture file contents, private messages, unrelated applications, raw keystrokes, precise location, contact lists, or health records.

Use participant-selected pseudonyms in working data. Keep consent records separate from task data. Publish aggregates and de-identified excerpts only after participant review.

## Interpretation rule

PIXIE may report: “The participant reopened the location picker four times.”

PIXIE may not convert that event into: “The participant was depressed,” “The participant lacked capacity,” or “The participant refused to cooperate.”

Observed action, participant explanation, researcher interpretation, and system hypothesis must remain separate fields.

## Withdrawal and correction

Participants may stop a session, withdraw future use of their contribution, and request correction of public interpretation. Historical public releases cannot always be erased from distributed copies, so publication consent must occur before release and must explain that limit plainly.

## Review before clinical use

Any therapist-facing experiment requires a separate protocol, clinical leadership, security review, applicable legal review, and explicit patient-controlled sharing. It is outside the initial creative-industry pilot.