Design strategy

Why this document exists

The platform began as the SFIA9 Skills Assessor — one AI interview, one report, one assessor dashboard. It is now becoming a Universal Capability Measurement and Growth Platform: multi-instrument, multi-stakeholder, multi-sector, serving government, private enterprise, community organisations, education, healthcare, and professional bodies.

This document is the design spine. It sets the four pillars every design decision is measured against — Information Architecture, Component System, Voice & Tone, and Accessibility & Trust — and the principles that govern how they fit together. It is decisive on purpose. The stakes are too high for design to be improvisational.

The deep operational references live elsewhere: the voice doctrine and anti-pattern catalogue at /docs/voice, the seven sector packs at /docs/sector-packs, the SFIA 9 alignment at /docs/sfia-partnership, and the trust-and-safety commitments at /trust. This page is the design strategy that frames them.

---

The five design principles

These are the platform's design tests. Every feature, every component, every line of copy must be defensible against all five.

1. One canonical per concept. Every domain concept — evidence, skill level, attribute, instrument, status, programme step, capability, consent — has exactly one canonical representation. When in doubt, extend the canonical; when there is no canonical, build it first.
2. Token-driven end-to-end. No hardcoded colour, no hardcoded duration, no hardcoded type size. Everything flows through a layered token system: reference → semantic → component → sector. This is what makes sector adaptability a runtime concern and white-label real.
3. Power-aware by construction. The platform measures people. The structural asymmetry between organisation, practitioner, and individual is the most consequential design constraint. Every surface is designed against the question: what does this look like from the position of the person being measured?
4. Accessible by default. Accessibility is a property of every primitive, not a stage in a release. Focus rings, keyboard interaction, ARIA, contrast, reduced motion — wired into Tier-1 components, not bolted on per use.
5. Truthful. We do not lie about what the platform is, what the AI is, who reads what, or what the rating means. Trust earned by transparency is durable; trust earned by concealment is borrowed against the day the concealment fails.

A feature that ships in violation of these is design debt before it is anything else.

---

The four pillars

The platform is designed along four dimensions. Each is a complete strategy. Each draws on the others. None is optional.

Pillar	What it answers
Information Architecture	What does the platform contain and how does a user find their way through it?
Component System & Design Tokens	How does the platform look, feel, and adapt — across sectors, themes, and audiences?
Voice, Tone & Content	How does the platform talk — to participants, to practitioners, to organisations?
Accessibility, Trust, Safety & Cultural Inclusivity	How does the platform honour the people it measures?

The pillars are not separable in practice — a voice decision constrains a component decision, an IA decision constrains a trust decision, a cultural decision constrains everything. They are separated below for clarity, with explicit cross-references where they interact.

---

Information architecture & audience modes

Why the IA spine matters

The platform's surface area has outgrown the mental model that produced it. We began with one job — running an SFIA interview and producing a report — and have accreted multi-instrument programmes, simulations, integration sessions, longitudinal pre/post measurement, and sector overlays for government, private, community, education, and professional bodies. Without an IA spine, every new feature becomes another sidebar item and the cognitive load is paid by the user.

The spine answers three questions every feature must satisfy before it ships:

1. Which pillar does it serve? Measurement, sense-making, growth, or governance?
2. Which audience mode is it for? Individual, Practitioner, or Organisation — and is it a handoff between them?
3. Which sector posture applies? Government accountability and probity, private competitiveness, community equity, education developmental framing, professional-body certification rigour.

If a feature cannot answer all three crisply, it is not ready to be built.

The four-pillar spine

The platform's value chain reduces cleanly to four verbs:

• Measure — generate signal about capability (instruments, interviews, simulations, programmes).
• Understand — turn raw evidence into shared meaning (reports, comparisons, calibration, integration sessions, analytics).
• Grow — convert understanding into deliberate change (development plans, goals, recommendations, longitudinal progress).
• Govern — make the measurement credible and the data safe (sign-off, consent, instrument health, sector posture, audit).

Every surface belongs to exactly one of these. The sharpest boundary is between Measure and Understand: producing evidence is Measure; reading a finished assessment is Understand. The other sharp boundary is between Understand and Grow: saying "what is true" is Understand; saying "what next" is Grow.

The three audience modes

Mode is determined by account identity, not by a toggle. A Practitioner does not become an Individual by clicking a button; they switch context to view their own development.

• Individual. A candidate, participant, learner, member, volunteer. Primary verbs: complete, see, understand mine, plan, progress, share. Low-density, single-column, plain language. One primary action per screen. Never sees other individuals' results, the practitioner's calibration apparatus, or organisational aggregates.
• Practitioner. Assessor, organisational psychologist, coach, examiner. Primary verbs: design, conduct, calibrate, validate, integrate, sign off. High-density, multi-pane, professional terminology used precisely. Surfaces confidence intervals, evidence provenance, multi-assessor dissent.
• Organisation. HR leader, capability strategist, executive sponsor, board director, programme owner. Primary verbs: commission, configure, view aggregate, compare cohorts, report up, plan workforce. Executive density: summary-first, drill-down on demand. Never sees individually identifying data below threshold cohort sizes.

Cross-mode handoffs

The IA's hardest work is at the seams. Six handoffs deserve first-class design:

1. Commission → Design → Run → Integrate → Report. Organisation commissions a Programme; Practitioner designs the instrument mix; Individuals participate; reports flow aggregated to Organisation and individual to each participant.
2. Individual completes assessment → Practitioner reviews evidence → Expert signs off → Individual receives report. The participant does not see their report until governance has cleared it.
3. Practitioner identifies development need → Individual accepts goal → Longitudinal re-measurement triggers. The development plan is a live object, not a static PDF.
4. Organisation views cohort gap → Practitioner proposes intervention → Individuals enrol → Pre/post deltas reported back. Closes the loop from organisational strategy to individual growth.
5. Individual disputes a finding → Practitioner re-reviews → Governance records the appeal. Procedural fairness for government and professional-body sectors.
6. Anonymous interview (/i/[token]) → Portal claim → Persistent Individual account. The platform's two access models converge — the token interview is the entry, the magic-link portal is the home.

Navigation pattern

Top-level: pillars. Four destinations — Measure, Understand, Grow, Govern — always visible. The spine does not change with mode.

Left rail: per-pillar. Within each pillar, a left rail lists the pillar's destinations filtered by mode. Practitioner inside Measure sees Programmes, Assessments, Profiles, Simulations. Individual inside Measure sees only their active assessment.

Breadcrumbs. Always Pillar / Destination / Object — never deeper than three.

Sector theming. The organisation's identity (logo, accent colour, sector posture label) sits in the top bar; sector-specific language reskins headings and empty states — but the pillar names themselves are stable. A government department and a university see the same four pillars; they see different vocabulary inside them.

Surface map

The table below is the canonical reference; the figure above is for skimming. The interview runner at /i/[token] is treated as a contract — its URL shape will not break.

Pillar	Destination	Audience
Measure	Programmes	Practitioner, Organisation
Measure	Assessments	Practitioner
Measure	Interview Runner (/i/[token])	Individual
Measure	Simulations	Individual, Practitioner
Measure	Profiles (Targets)	Practitioner, Organisation
Measure	Pre/Post	Practitioner
Understand	Report	Practitioner, Individual (released view)
Understand	Evidence	Practitioner
Understand	Calibration	Practitioner
Understand	Integration Session	Practitioner
Understand	Team Capabilities	Organisation
Understand	My Profile	Individual
Grow	Development Plan	Practitioner, Individual
Grow	My Goals	Individual
Grow	Recommendations & Partner Content	Individual, Practitioner
Grow	Progress (longitudinal)	Individual, Organisation
Govern	Sign-off	Practitioner (expert)
Govern	Consent Ledger	Organisation
Govern	Instrument Health	Organisation, Practitioner
Govern	Sector Posture	Organisation
Govern	Reference & Help	All modes

---

Component system

One canonical per concept

The platform sits on a sound technical base — shadcn/ui primitives, Tailwind v4 with OKLch tokens, dark-first theming, Geist for type, Lucide for icons. What we add is curation: every domain concept has exactly one canonical component. If you find yourself re-implementing an evidence card, you are doing it wrong. Find EvidenceCard or extend it.

Five operating doctrines follow:

1. One canonical per concept — extend the canonical; when there is no canonical, build it.
2. No hardcoded colour, ever — a hex value or a raw bg-blue-500 in a feature component is a bug.
3. Composition over configuration — components expose slots and variants, not thirty-prop kitchen sinks.
4. Accessible by default — focus rings, keyboard interaction, ARIA, contrast wired into the primitive.
5. Sector adaptability is runtime, not a fork — the same codebase serves an APS agency, an NFP, and a university.

Three-tier model

The narrowing is deliberate. Many surfaces (audience-specific compositions) build on few canonicals (one per domain concept) build on fewer primitives (shadcn / Radix). Promotion from surface to platform is a PR-reviewed move when a component appears in two or more audience modes; demotion happens when a platform component is used by exactly one surface for six months. Platform components consume primitives; primitives never graduate into domain-aware components.

The platform tier holds ~35 canonical components today — EvidenceCard, AttributeChip, SkillLevelIndicator, StatusBadge, ConsentSurface, ProvenanceTag, AcknowledgementOfCountry, CrisisReferral, AuthenticatedAppShell, FairnessMetricPanel, ProgrammeStepper, TimedTaskShell, and others. The catalogue is browsable in Storybook and is the single source for what to import — never re-implement a canonical.

Sector theming

Sector posture is a runtime choice, not a fork. The same codebase serves seven sectors via overlay packs that swap palette, density, motion, type, copy register, and policy defaults without divergent components. See /docs/sector-packs for the pack anatomy and the seven packs in the registry.

---

Voice, tone & content

The platform is a measurement instrument with a conversational surface. Voice is how we keep the interaction fair, legible, and dignified. The full doctrine — including the anti-pattern catalogue, sector tone variants, trauma-informed language guide, and AI persona spec for Ava — lives at /docs/voice. The summary here:

Five principles. Every line of copy must hold under all five.

1. Truthful — no misrepresentation of what the platform is, what the AI is, who reads what, or what the rating means.
2. Power-aware — name the asymmetry where it exists; give the individual the agency the design allows.
3. Plain — translate psychometric, SFIA, and HR language; use the technical word only when it carries weight, with a one-line gloss.
4. Warm without performing — friendly and respectful, never casual or cheerful by manufacture. A colleague, not a hype account.
5. Specific — concrete language over hedged abstraction. Evidence over impression. Name the thing.

AI disclosure is a position, not a hedge. The interview AI is disclosed up front by name (Ava), with a plain account of what is captured, who reads it, and what choices the participant has. The "NEVER mention you are an AI" pattern common in older assistants is retired. Hiding AI is a substitute for the agency a good consent surface already provides; we provide the agency and stop hiding.

Sector tone variants flex; the doctrine stays constant. The same idea renders differently across government, healthcare, community, education, finance, corporate, and professional-body deployments. Practitioners do not write sector copy ad hoc — overlays choose between pre-approved phrasings reviewed by sector practitioners.

---

Accessibility, trust & safety

This platform measures people. The stakes are the stakes of opportunity, identity, and dignity. We treat accessibility, trust, safety, and cultural inclusivity as design pillars — load-bearing architecture every feature is built on — not as a compliance stage at the end. The operational commitments and detailed mechanics live on the public trust page at /trust; the strategy here is the why and the what.

The four commitments

We make four commitments to every individual the platform measures.

1. Accessibility beyond compliance. Every individual can participate authentically, in a mode that works for their body, mind, and circumstances. Accommodations are designed-in, not bolted on, not opt-in-buried-in-settings. Stored preferences (high-contrast, large text, reduced motion, extended processing time) actually drive the rendered UI — not just sit in a database.
2. Power-aware design. The individual sees what is happening to them, has agency over what they share and how, and has a real path to act on disagreement. Observation is disclosed. Decisions are explainable. There is no shadow mode invisible to the participant.
3. Cultural respect. The platform does not impose a single cultural frame. It adapts to the individual, to the sector, and to the country in which it operates. It does not require people to translate themselves into a Western, English, individualist, deficit-framed business idiom to be seen. Acknowledgement of Country, inclusive identity fields, communication-norm awareness, sector overlays.
4. Trauma-aware as default. Pause, step away, and withdraw are three distinct one-click mechanics, available from every assessment surface. Distress signals trigger a clear choice (break, skip, end, talk to a person). Crisis-referral services are surfaced with locale-appropriate numbers. The platform is gentle by default, not as a sector flag.

Three load-bearing patterns

Pause / step away / withdraw. Three distinct mechanics, each one click from any assessment surface.

• Pause — save state, return when ready, no judgement.
• Step away — end this session, resume later. State preserved.
• Withdraw — end this assessment entirely. At withdrawal the individual chooses whether responses are retained, anonymised, or deleted. Deletion is the default offered.

Distress detection. The platform watches for distress signals (lexical cues, response patterns, explicit statements) and pauses the assessment to offer a clear choice — break, skip, end, talk to a person — with a crisis-referral component surfaced inline. No data about the disclosure is shared with the commissioning organisation unless the individual explicitly flags it for follow-up.

Layered consent. Consent is separable — to AI participation, voice / recording, anonymised use for norming, sharing with the commissioning organisation, cross-team visibility, anonymous benchmarking. Each can be agreed or declined independently. Mid-session change is supported. Withdrawal is one click. The full consent record is versioned and recoverable by the individual on request.

Sector safety overlays

Different sectors require different safety affordances. Each pack carries sector-tuned distress thresholds, crisis-referral routing, mandatory-disclosure handling, retention policy, and observation rules. Education adds child-safe + mandatory-reporting compliance; healthcare adds clinical-precision language + clinical crisis routing; government (Indigenous-led) adds mandatory Acknowledgement of Country and cultural advisor sign-off. See /docs/sector-packs for the per-sector requirements.

Power-asymmetry in the moment

• Before measurement. The individual sees plainly what is about to happen, who will read the results, what choices they have, and what each choice means. No surprise reveals.
• During measurement. Agency points at every stage — pause, skip, switch mode, ask a question, see what the platform has gathered so far.
• After measurement. The individual sees what was concluded, sees the evidence behind it, and has a clear one-click path to contest. Their input goes into platform improvement and is reviewed alongside assessor feedback, not as a footnote to it.

---

How the pillars connect

The four pillars are not separable in practice. Every meaningful design decision touches at least two of them.

• IA × Voice. Pillar names (Measure / Understand / Grow / Govern) are voice decisions as much as IA decisions. Their plainness is deliberate. The IA spine speaks the platform's voice every time someone reads a breadcrumb.
• IA × Component System. Each pillar has signature platform components — Measure is InstrumentCard + ProgrammeStepper + TimedTaskShell; Understand is EvidenceCard + ConsensusPanel + ComparisonView; Grow is DevelopmentPlanCard + GrowthChart; Govern is AuditTrailEntry + ConsentSurface + ProvenanceTag. The component library is the physical embodiment of the IA.
• Voice × Trust. The AI disclosure stance is a voice decision (we tell the truth about Ava) and a trust decision (the AI provenance ledger is what makes the disclosure verifiable). Voice without architecture is a marketing claim; architecture without voice is silent.
• Component System × Accessibility. Tier-1 primitives carry accessibility properties. Tier-2 platform components inherit them. Surface code cannot break what the primitive enforces. This is what "accessible by construction" means concretely.
• Sector overlays × everything. A sector overlay touches IA (which destinations show in which mode), the component system (tokens, density), voice (sector tone variant, terminology), and accessibility (healthcare requires trauma-aware default; education requires child-safe). The overlay system is what makes the platform plural without forking.
• Power-aware × everything. Of the five principles, "Power-aware by construction" is the one that most aggressively crosses pillars. It changes what is in the IA (consent surface, rights surface, defensibility bundle), what components exist (ConsentSurface, ProvenanceTag), what voice we use (the disclosure script), and what accessibility means (skip-without-penalty patterns). Designed-in power-awareness is the single feature that distinguishes the platform from every other AI assessment product on the market.

---

Governance & ownership

Pillar	Owner (now)	Owner (mature team)
Information Architecture	Founder	Head of Product
Component System	Founder + lead engineer	Head of Design (with engineering pair)
Voice, Tone & Content	Founder	Head of Content
Accessibility, Trust, Safety, Cultural	Founder	Trust & Safety Lead

Cross-cutting:

• Voice owner sign-off is required for every AI prompt change and every new sector overlay.
• T&S checklist is required for every user-facing PR (accessibility, consent, locale, distress pathway).
• Co-design panel input is required before any new instrument or major flow change ships.
• Ethics board is convened quarterly or for high-stakes decisions; composition is lived-experience + academic + regulatory.

A pillar without an owner is a slogan. The named ownership matters as much as the document.

---

How to use this document

If you are evaluating the platform for purchase. Read the five principles and the four commitments. Those are the architectural claims you can hold us to. Cross-reference with the trust page at /trust for the operational commitments, and the SFIA partnership page for framework fidelity.

If you are designing or extending a feature. Pick the pillar and audience mode from the surface map. Find the canonical components in Storybook (do not re-implement them). Write the copy against the voice doctrine. Test against the four commitments. If all four pass, the feature has the platform's design backing.

If you are commissioning a sector overlay. Read the sector packs page. Sector preset + sector tone variant + sector safety overlay compose to make a sector. None are written ad hoc — each follows a co-design process with sector practitioners.

If you disagree with something here. Open a revision PR. Disagreement is welcome and necessary. What we do not do is implement against the doctrine quietly — the document changes; the practice does not drift.