How we built Sylva support

Sylva is a digital mental health platform for families of neurodivergent children and young people, aged 2 to 25. It delivers psychologically informed support through five AI mentors: Hazel, Luna, Fern, Rowan, and Phoenix.

Each mentor is designed to a rigorous clinical specification. This document details how these mentors were built and trained, the evidence base behind them, their functionality, and how we continuously verify their clinical accuracy and impact.

1. How the mentors were built

The Sylva AI mentors are not generic chatbots adapted for parenting content. Each is a configured Large Language Model (LLM) persona, governed by a versioned Skills and Content Specification that defines its therapeutic scope, micro-skills, mediator pathways, and crisis behaviour. The build follows a deliberate, audit-traceable pipeline rather than conventional machine-learning training.

1.1 The five-mentor architecture

The platform addresses five distinct areas of family wellbeing, each with a dedicated mentor and its own evidence base:

1.2 The build pipeline

Every mentor is built through five mandatory stages. No stage is closed until its Definition of Done (DoD) self-assessment is signed off by the CEO, Ben Cosh.

Stage 1: Two-phase evidence review (SM-DoD-LITREV-001). Phase 1 identifies therapeutic modalities with qualifying systematic-review or meta-analytic evidence. Phase 2 identifies the mediators through which each modality produces change, tiered as Well Evidenced, Evidenced, or Theoretical/Exploratory.

Stage 2: Skills and content specification. Each evidenced mediator is broken down into discrete micro-skills, each tagged with a mediator pathway reference. No skill can exist in production without a traceable link to a published mediator estimate.

Stage 3: Prompt engineering and content authoring. System prompts, dialogue templates, and crisis pathways are written to operationalise the specification. For life-threatening situations, the guiding principle is "smoke detector, not triage nurse": mentors detect risk and route to appropriate services but never attempt acute clinical management.

Stage 4: Version-controlled deployment. The deployed system comprises a version-controlled base LLM, prompt chain, content libraries, and retrieval indices. Per the Clinical Investigation Plan, this combination constitutes the manufacturing control under UK MDR 2002 Reg 16(8)(a). Releases are gated through CI/CD with model-version pinning.

2. What the mentors were trained on

"Training" means two different things here, and only one applies to Sylva.

Foundation-model training (not done by Sylva Minds). The underlying LLM is a third-party foundation model. Sylva Minds does not perform additional gradient-based fine-tuning on parent or child data. This is a deliberate safety and information-governance decision, consistent with the platform's regulatory posture.

Behavioural training (done by Sylva Minds). What Sylva Minds does control is mentor behaviour at inference time, through two versioned artefacts:

• ·The system prompt, defining the mentor's role, scope, voice, refusals, and crisis routing
• ·The retrieval and routing layer, rules governing which content blocks are surfaced for which user state, including safeguarding and crisis templates that override normal conversational behaviour when triggered

These artefacts are the only things that shape mentor behaviour in production. They are version-controlled, CEO-approved, and tied to release tags so that any user session can be reproduced from the live artifact bundle at that time.

What the mentors were not trained on: identifiable user data, safeguarding-flagged content, transcripts containing minors' disclosures, or any data outside the lawful basis declared in the research Data Protection Impact Assessment (DPIA).

3. The evidence base

The evidence architecture is deliberately layered. No single body of literature can support a five-mentor family-systems intervention on its own.

3.1 Modality-level evidence (does this approach work?)

Acceptance and Commitment Therapy (ACT) and Psychological Flexibility. Over 55,000 Randomised Controlled Trials (RCTs) across mental, behavioural and physical health domains, with replicated mediation evidence. Hayes et al. (2022), the most comprehensive umbrella review of replicated psychosocial mediators to date, found that more than half of all successful mediational findings are accounted for by psychological-flexibility processes, directly supporting Hazel and Fern.

Parent training for child externalising behaviour. Furlong et al. (Cochrane/AHRQ) reported a pooled effect of Weighted Mean Difference (WMD) -20.90 on Eyberg Child Behavior Inventory (ECBI) Intensity across 24 trials of behavioural and cognitive-behavioural parent training (Incredible Years, Triple P, PCIT). Single-trial effect sizes of d = 0.74 to 0.89 (McGilloway 2012; Sanders 2012) form the historical-control benchmark for Hazel's clinical investigation.

Emotionally Focused Therapy (EFT) and Gottman Couple Work. A multi-decade RCT base for couple distress, with attachment-shift mediation, forms the foundation for Luna.

Mentalisation-Based Therapy (MBT) / mentalising and reflective functioning. Reflective-function-mediated change for emotion regulation and relational difficulty forms the foundation for Rowan.

3.2 Mechanism-level evidence (why does it work?)

For each mentor, the Phase 2 review identifies the mediators with the strongest empirical indirect-effect evidence. Hazel's primary causal pathway illustrates the approach:

External stressors such as illness, school transitions, and family disruption are captured as time-varying covariates via NLP extraction from session digests. This reflects the reality that families engage with Sylva in bursts around crises rather than on a fixed weekly schedule.

3.3 Digital delivery evidence

Digital delivery is not a compromise on quality. Breitenstein et al. (2016) reported 85.4% module completion for tablet-based parent training versus 50.6% for face-to-face groups. The Scottish PoPP evaluation found 76% completion digitally versus 50 to 60% face-to-face. Father participation rises markedly when barriers of stigma, transport and childcare are removed.

Regulatory and adoption precedent is established by ZOE (4.6M consented users), Sleepio/Big Health (12 RCTs, first NICE-guided digital therapeutic), Woebot (FDA Breakthrough Device), and Wysa (FDA Breakthrough Device, NHS NELF, n=36,070).

No AI LLM-personalised parenting tool currently has published full clinical evidence globally. Sylva is first-in-class.

3.4 Honest framing of comparators

Sylva benchmarks against published RCT data rather than a parallel internal control. Because Sylva is publicly available on app stores, allocation concealment within the same recruitment channel is structurally impossible. The Medicines and Healthcare products Regulatory Agency's (MHRA) May 2025 draft guidance on real-world data and external control arms recognises this design where mechanism evidence is strong and the novelty lies in delivery. NHS Talking Therapies real-world data (Gyani et al., 2013: 40.3% reliable recovery, 6.6% reliable deterioration across 19,395 patients) provides a separate calibration point for what "good" and "safe" look like in routine digital psychological provision.

4. What the mentors do

Each mentor performs exactly four functions. Anything outside this scope is either redirected within the mentor's remit, handed to another mentor, or escalated to crisis pathways.

• ·Detect. Identify what the parent is bringing, whether a meltdown post-mortem, a school-refusal episode, a couple argument, or a self-critical spiral, and match it to the relevant micro-skill cluster.
• ·Personalise. Draw on the user's screening and monitoring profile (via the UKCA-marked Class I screening module) and session history to tailor language, examples and pacing. Personalisation means content selection, not diagnosis or treatment recommendation.
• ·Deliver active ingredients. Offer the relevant evidence-based technique: defusion, values clarification, and behavioural activation for Hazel and Fern; softening or de-escalation scripts for Luna; reflective-function prompts for Rowan; psychoeducation; and coping plans for predictable trigger events.
• ·Detect risk and route. Surface a CEO-approved crisis template and signpost to 999, NHS 111 or Samaritans on detection of suicidality, self-harm, child safeguarding concerns, domestic abuse, or acute capacity loss. The mentor is a smoke detector, not a triage nurse.

5. How the mentors work

The software structure operates as follows:

Burst engagement is a feature, not a flaw. Parents typically engage with Sylva around stressors such as illness, transitions, and school escalations, rather than on a fixed weekly schedule. The session-digest layer carries state forward across these bursts, so the mentor always responds in continuity with what the family is actually living through.

6. How we know the mentors faithfully deliver what we designed

Fidelity, meaning faithful delivery of the designed intervention, is what distinguishes a defensible digital tool from a marketing claim. Sylva addresses this through six independent, mutually reinforcing controls.

6.1 Specification-to-output traceability. Every micro-skill carries a mediator tag and every production output must map back to a skill. If it cannot, it is logged as a device deficiency under ISO 14155 §6.4.2 and investigated accordingly.

6.2 Replicability test. The Definition of Done requires that another research team could replicate Sylva from the input documentation. This is the standard NIHR and MRC reviewers expect for complex interventions (Skivington et al., 2021).

6.3 Version-controlled manufacturing control. Under UK MDR 2002 Reg 16(8)(a), the version-controlled LLM, prompt chain, and content library together constitute the manufacturing control. ISO 62304-compliant CI/CD release gates and model-version pinning are mandatory before any cohort enrolment crosses a release boundary.

6.4 Adverse-event and device-deficiency adjudication. An ISO 14155 and DCB0129-compliant pipeline operates throughout. Causality is adjudicated by an independent DSL who is not the CEO.

6.5 Dose, mediator, and outcome triangulation. Hazel's SAP (SM-SAP-HAZL-001) requires that any non-inferiority claim against the parent-training benchmark is corroborated by a monotonic dose-response in engagement and a concordant mediator response (CompACT, PSOC, PSI-SF). A favourable outcome without concordant mediator movement is reported but flagged as not supporting the causal claim.

6.6 Independent process evaluation. Per MRC/NIHR (2021), every causal pathway requires a process-evaluation question with a method to test it. Qualitative exit interviews verify whether the mechanisms parents describe match those the specification claims.

7. Status and next steps

Wellness tier (live). Mentors are deployed under the wellness intended-use statement, with Tier-1 NICE DHT evidence including UK clinical experts in design, a Bangor feasibility study, SUS and accessibility testing, DCB0129 clinical safety, and a registered DSL.

Class I Medical Device (in preparation). The Hazel CIP draft against UK MDR 2002 Annex X and ISO 14155:2011 is complete. The SAP is locked at Rev 0.1 (n = 134 completers, 12-week ECBI primary, PSI-SF confirmatory secondary, Furlong AHRQ pooled WMD as benchmark). A Phase 0 MHRA Innovation Office meeting is the next gate, alongside umbrella ethics under the ZOE/KCL model with the University of Birmingham.

The programme's intent is to move Sylva's mentors from defensible wellness deployment, through Medical Device investigational evidence under MHRA notification, to NICE technology appraisal and NHS commissioning, with parallel preparation for US TRICARE and FDA pathways. The same architecture, specification, mediator-tagged micro-skills, version-controlled build, and ISO 14155-grade fidelity controls are the prerequisite at every step.