05 — Safeguarding and AI Ethics Appendix (v2)

Status: Stub. Boring and beautiful. Has to be substantive given the participant population (neurodivergent children).

Purpose

The original draft has one bullet about AI safety under Risks. That’s not enough. v2 needs an explicit appendix that:

Maps every operational decision to the relevant clauses of the NDIS Code of Conduct
Specifies the safeguarding posture for working with neurodivergent minors (the primary participant group)
Defines the AI ethics framework for any AI-powered tool deployed to a participant
Documents the data handling, consent, escalation, and exit story in detail sufficient for a plan manager to read and trust

Map each of the seven elements of the Code (see synthesis/ndis-context.md) to operational practice in this service. Concrete: how does the practice respect privacy? How does it provide supports safely? What’s the integrity story? How does it prevent and respond to harm?

Don’t restate the Code; show how the practice implements it.

2. Worker screening and clearances

NDIS Worker Screening Check status for James and Adrian (both should have it before launch)
State/territory Working with Children Checks (NT for James, TAS for Adrian)
Renewal and verification cadence
Disclosure to participants and families

Guardian consent: what is asked, when, how
Participant assent: age-appropriate explanation of what the tool does, what data it captures, what happens if they don’t want it
Right to withdraw without prejudice or financial penalty
How consent is documented
How assent is checked over time, not just at intake

4. Data handling

For every data type the service touches:

Participant assessment notes — where stored, who accesses, retention
Configuration files for deployed tools — same
AI tool interaction logs — this is the highest-risk category, especially if conversations with neurodivergent minors are stored
Video/audio recordings — default policy is don’t, but if exceptions, document them
Guardian access — how families view/correct/delete their child’s data

Reference: Australian Privacy Principles. NDIS providers handling personal information are bound by them.

5. AI ethics framework for deployed tools

This is the section that has to be substantive, not a wave at “Failure First.”

For any AI-powered tool deployed to a participant:

What the tool does and does not do (capability boundaries)
What data the tool sees (and doesn’t see)
What happens when the tool encounters something outside its design scope (escalation, fail-safe, who gets notified)
Logging and review cadence — how often does Adrian review tool behaviour? What triggers an out-of-cycle review?
Failure modes catalogued before deployment (the actual Failure First application)
Specific commitments around AI companion patterns: no fostering of parasocial dependency, no replacement of human connection, no collection of conversational data beyond what’s necessary for the participant’s stated goal

This section should reference SPARK’s existing safety architecture where applicable (see https://spark.wedd.au — privacy filter, QA gate, quiet mode, school-hours suppression, salience thresholding) but distinguish between SPARK as an existing personal project and any deployed configuration as a separate, reviewed instance.

6. Escalation and exit

What happens if a tool isn’t working for a participant — exit procedure
What happens if safety concerns arise — escalation path
What happens if a complaint is made — NDIS Code of Conduct complaints process (Adrian and James both need to know this cold)
What happens at end of engagement — data retention, decommissioning of deployed tools, family handover

7. Out-of-scope declarations

Explicit list of what this service is NOT:

Not a clinical mental health service
Not crisis support
Not a substitute for allied health assessment (OT, SLP)
Not a 24/7 monitoring service
Not behavioural intervention
AI tools deployed are not therapeutic devices and make no clinical claims

Out-of-scope declarations protect the practice and the participant.

Open dependencies

04-roles-and-scope.md (whose Worker Screening covers what)
06-spark-deployment-model.md (the AI ethics framework partly depends on what SPARK-as-deployed actually is)
External advice — at least one paid consultation with someone who understands NDIS safeguarding for neurodivergent minors
Verification of Failure First framework as a documented methodology with artefacts that can be referenced