# moomoo Internal Persona Platform — Product Requirements Document (PRD)
**Document version:** 1.0
**Status:** Draft for internal review
**Owner:** User Research Team — moomoo Global
**Last updated:** 2026-06-02
**Audience:** Product, Research, Design, Data Science, AI Engineering, Compliance, Legal
---
## 1. Executive Summary
moomoo is a global multi-market trading app operating across the US, Hong Kong, Japan, Singapore, Malaysia, Australia, and additional emerging markets. Across these markets the user research team has accumulated hundreds of qualitative interviews, longitudinal surveys, VOC streams, App Store and Play Store reviews, customer support tickets, competitor research, and quantitative product analytics. Today this knowledge is fragmented across Dovetail-style repositories, shared drives, regional research wikis, and individual researchers' heads. Reusing it for new product decisions is slow, and most product teams default to "ask a researcher" or "ship and learn."
The **moomoo Internal Persona Platform** ("MIPP") is an internal, AI-powered system that converts these accumulated assets into **evidence-grounded persona agents**. Researchers, PMs, designers, and growth leads will be able to:
- Query a persona ("Singapore new investor") and see grounded responses with citations;
- Run synthetic persona panels against early concepts, prototypes, or PRDs;
- Auto-generate research guides, screeners, and usability tasks calibrated to real personas;
- Search across all research assets with persona-aware retrieval;
- Trace every claim back to its source.
The platform is explicitly **not** a replacement for real user interviews. It is a force-multiplier that surfaces what is already known, generates better hypotheses, and frees the research team to spend their time on the questions that genuinely require real users.
The MVP targets a 6-month build, beginning with three flagship markets (US, Hong Kong, Japan) and four anchor personas per market. Success will be measured by research cycle time, hypothesis quality, and downstream product impact.
---
## 2. Background and Problem Statement
### 2.1 Today's research operating model
| Aspect | Current state | Pain |
|---|---|---|
| Research storage | Dovetail + market-specific wikis + Lark/Feishu folders | Discoverable only if you know it exists |
| Cross-market reuse | Manual, ad-hoc | A Japan researcher rarely sees an HK insight unless someone forwards it |
| Persona artifacts | Static slides per project | Stale within 1-2 quarters, not queryable |
| PM/designer self-service | Almost none | Every concept question routes through researchers |
| VOC integration | Separated from interviews | Quantitative signal and qualitative narrative live apart |
| Competitor knowledge | Lives in individual decks | No persona-level competitor mapping |
### 2.2 Why now
- **LLM grounding has matured.** Retrieval-augmented generation, citation chains, and evidence-binding patterns are now production-ready.
- **Competitor pressure on research velocity.** Robinhood, Webull, SBI, Rakuten and regional incumbents are all shipping product faster; moomoo cannot afford slow research cycles for low-risk decisions.
- **Synthetic research tools are emerging.** GetMinds, Uxia, Synthetic Users, and Simile AI all exist — moomoo can either adopt them externally (with all the data and grounding limitations that implies) or build a financial-services-aware, evidence-grounded internal alternative.
- **Regulatory environment.** Financial product copy, risk disclosures, and product positioning need fast iteration with risk-aware review. A generic external tool cannot meet this bar.
### 2.3 Problem statement
> moomoo's product, design, and research teams need a fast, evidence-grounded way to access user knowledge across markets and product lines, generate calibrated hypotheses before real-user testing, and design better research — without diluting the rigor of real user research or exposing the company to compliance risk from ungrounded AI claims.
---
## 3. Product Vision
> **An internal AI-powered persona and research intelligence platform that transforms moomoo's real user research assets, behavioral data, survey data, VOC, and market knowledge into evidence-grounded persona agents, helping researchers and product teams simulate user reactions, design better research, and accelerate product decisions.**
### 3.1 Vision pillars
1. **Evidence first, generation second.** Every persona response must be grounded in or qualified against real research artifacts. Ungrounded generation is labeled and rate-limited.
2. **Researcher-augmenting, not researcher-replacing.** The platform makes researchers higher-leverage; it does not pretend to be them.
3. **Global by construction.** Market, language, regulation, and cultural context are first-class persona attributes — not afterthoughts.
4. **Financial-services-aware.** Every output passes through financial-content safeguards. No personalized investment advice, no fabricated regulatory claims, no fake quotes.
5. **Continuously learning.** Personas update as new research arrives; stale evidence is surfaced; synthetic predictions are scored against real findings.
---
## 4. Goals and Non-goals
### 4.1 Goals
| # | Goal | How we measure |
|---|---|---|
| G1 | Reduce time-to-hypothesis for new product concepts from days to <1 hour | Median time from PRD draft to first persona-grounded hypothesis report |
| G2 | Increase reuse of historical research assets | % of new research projects that cite ≥3 historical assets via the platform |
| G3 | Improve quality of research design (screeners, guides, tasks) | Researcher rating of AI-generated guides ≥4/5; reduction in interview-guide review cycles |
| G4 | Increase cross-market knowledge transfer | # of cross-market insight citations per quarter |
| G5 | Calibrate synthetic predictions against real outcomes | % of synthetic hypotheses validated/refuted after real testing, with confidence scores updated |
| G6 | Maintain compliance and grounding integrity | Zero ungrounded financial-advice outputs; 100% of synthetic quotes labeled |
### 4.2 Non-goals
| # | Non-goal | Rationale |
|---|---|---|
| NG1 | Replacing real moderated user interviews | Synthetic users cannot replicate lived investor behavior, especially around risk, money, and regret |
| NG2 | A fully automated UX research tool that ships decisions without human review | Decisions touching live trading UX must have human research validation |
| NG3 | A generic LLM chatbot | If it doesn't ground, it doesn't belong in this product |
| NG4 | External (customer-facing) deployment | Internal-only at MVP; external/partner exposure is a separate compliance scope |
| NG5 | Investment advice generation in any form | Strict guardrail — see §19 |
| NG6 | Quantitative market sizing or forecasting | Out of scope — different data stack |
---
## 5. Target Users and Stakeholders
### 5.1 Primary users (daily / weekly)
| User | Role | Top jobs-to-be-done |
|---|---|---|
| Global / regional UX researchers | Plan, run, synthesize studies | Reuse prior research; design better studies; share findings |
| Product managers | Define features, prioritize | Pressure-test concepts pre-research; understand cross-market user reactions |
| Product designers | Concept, prototype, iterate | Pre-flight UX review; comprehension checks on financial copy |
| Content / UX writers | Localize and de-jargon financial copy | Identify confusing wording per persona/market |
| Growth & lifecycle PMs | Onboarding, activation, retention | Understand drop-off rationale across user types |
### 5.2 Secondary users (occasional)
| User | Role | Top jobs-to-be-done |
|---|---|---|
| Business / strategy leads | Cross-market opportunity sizing context | Query "why" behind user behavior |
| Marketing & brand | Messaging resonance | Pre-test campaign concepts against personas |
| Customer support leadership | Spot pattern shifts | Cross-reference tickets with research themes |
| AI product team | Build / tune moomoo AI features | Understand AI-feature user mental models |
### 5.3 Governance stakeholders
| Stakeholder | Role |
|---|---|
| Compliance & legal (per market) | Approve guardrails, advisory boundaries, disclosure language |
| Data privacy / DPO | Approve PII handling, retention, access controls |
| Information security | Approve auth, RBAC, audit, model deployment topology |
| Research leadership | Approve persona quality standards, evidence thresholds |
---
## 6. Core Use Cases
| # | Use case | Primary user | Module(s) |
|---|---|---|---|
| UC1 | "Show me what we already know about US options users' attitude toward event contracts." | Researcher / PM | Knowledge Base + Persona Chat |
| UC2 | "Pre-flight this new membership page concept with 5 persona reviewers." | Designer / PM | Synthetic Panel + UX Feedback |
| UC3 | "Generate a real usability test plan for the new margin trading flow in Japan." | Researcher | Research Design Assistant |
| UC4 | "Why do Japan users hesitate to enable margin trading? Cite sources." | Business lead | Knowledge Base + Persona Chat |
| UC5 | "Compare how an HK high-AUM investor vs SG new investor would react to this AI-trading concept." | PM | Persona Chat (multi-persona compare) |
| UC6 | "I just finished 8 real interviews. Update my persona's confidence score and flag what was wrong." | Researcher | Persona Library + validation loop |
| UC7 | "Audit: every claim a PM cited from this platform in their launch decision deck." | Research lead / Compliance | Governance |
| UC8 | "Help me write a screener for Robinhood-migration users in the US." | Researcher | Research Design Assistant |
| UC9 | "What's the latest VOC theme for SG new investors in the last 30 days?" | PM | Knowledge Base (freshness-weighted) |
| UC10 | "Localize this risk disclosure for AU ETF investors — flag anything that sounds like advice." | UX writer + compliance | UX Feedback + Governance |
---
## 7. Competitor Analysis
### 7.1 GetMinds / Minds
| Aspect | Detail |
|---|---|
| Core positioning | AI persona panel platform for concept and message testing at scale |
| AI persona capability | Hundreds of pre-built or user-built personas can react to a stimulus simultaneously |
| Synthetic research workflow | Upload stimulus → select panel → receive structured & open-ended reactions, often as a survey-style report |
| How users create / query | Persona library + ad-hoc persona builder via prompt; query via concept upload or question battery |
| Strengths | Speed; large panel scale; rapid concept screening; useful early-funnel filter |
| Weaknesses | Personas are largely LLM-generated; limited evidence grounding; no native financial-services safeguards; outputs can over-converge or sound the same |
| Inspirations for moomoo | Multi-persona panel UX; concept upload flow; structured + open-ended hybrid reporting |
| Avoid | Treating outputs as evidence; using as a replacement for real research; building personas from prompts alone without grounded source data |
| **Key angle** | **GetMinds is strongest for AI persona panels and concept/message testing. Useful for simulating many personas at once, but its outputs should be treated as hypotheses, not final evidence.** |
### 7.2 Uxia
| Aspect | Detail |
|---|---|
| Core positioning | AI-driven UX testing for prototypes and live screens |
| AI UX testing capability | Synthetic usability testers walk through screens / flows, generate verbal think-aloud, identify friction |
| Prototype / screen testing workflow | Upload Figma / image / URL → select persona profile → receive structured UX review |
| Synthetic tester logic | LLM-driven simulation of user comprehension and likely behavior; some platforms add heuristic UX rules |
| Heatmap / visual UX | Some visual annotation, attention prediction, click-likelihood mapping |
| Strengths | Fast pre-test review; useful for catching obvious comprehension and labeling issues before real testing |
| Weaknesses | Cannot measure real task success; weak on emotional/financial trust signals; risk of false confidence; visual heatmaps from synthetic users are not behavioral truth |
| Inspirations for moomoo | Screen-level annotation UX; "what would a persona notice/miss" structured output; A/B variant comparison |
| Avoid | Claiming task-completion metrics; replacing real usability testing; presenting synthetic heatmaps as observed behavior |
| **Key angle** | **Uxia is strongest for AI-driven UX testing and prototype/screen review. Useful for early design-stage usability issue discovery, but cannot replace real task-based usability testing with actual investors.** |
### 7.3 Synthetic Users
| Aspect | Detail |
|---|---|
| Core positioning | Synthetic participants for AI-simulated interviews, surveys, and exploratory research |
| Synthetic participant logic | LLM persona generation calibrated via demographic data + survey priors + LLM world knowledge |
| Data sources | Mix of public demographic data, syndicated survey data, LLM world model; varies per provider |
| Interview transcripts | AI-generated transcripts with follow-up questioning; can run "interviews" at scale |
| Scientific framing | Often invokes representational alignment, neuroscience-inspired arguments for validity |
| Strengths | Fast exploratory research; useful for hypothesis-generation; good for early-stage discovery before fielding real studies |
| Weaknesses | Cannot validate behavior tied to lived experience (esp. money, risk, regret); representational alignment ≠ behavioral prediction; transcripts can pattern-match LLM training data; cultural and market-specific behaviors weakly captured |
| Inspirations for moomoo | Interview-style simulation UX; follow-up probing mechanism; structured hypothesis report; calibration vs real research |
| Avoid | Conflating representational validity with predictive validity; using transcripts as evidence in product decisions; trusting synthetic users on culturally / financially sensitive topics |
| **Key angle** | **Synthetic Users is strongest for simulated interview-style research and early hypothesis generation. Its neuroscience / representational-alignment arguments may improve credibility, but they do not prove synthetic users can predict real investor behavior.** |
### 7.4 Simile AI
| Aspect | Detail |
|---|---|
| Core positioning | Research-grounded "digital replicas" built from real interview data |
| Persona generation | Real interview transcripts are encoded into persona models, then queried |
| Difference from generic AI personas | Grounded in actual research artifacts rather than prompt-only generation |
| Strengths | Most credible evidence chain among synthetic tools; persona answers can be traced to interview source; closer to true augmentation of qualitative research |
| Weaknesses | Quality is bounded by quality and recency of underlying interviews; doesn't generalize well to populations outside source data; still cannot replace fresh research for new contexts |
| Inspirations for moomoo | Strong inspiration — research-grounded persona replicas are the right model; evidence chain to source transcript is core |
| Avoid | Over-claiming generalization to unstudied segments; allowing persona drift as new generations propagate without grounding |
| **Key angle** | **Simile AI is most relevant as inspiration for research-grounded persona replicas. moomoo should prioritize evidence-grounded persona agents built from real user interviews, survey data, VOC, and behavioral data, rather than purely prompt-generated personas.** |
### 7.5 Competitor comparison table
| Platform | Core positioning | Main use case | Input data | Output type | Best feature | Weakness | Relevance to moomoo | Product inspiration |
|---|---|---|---|---|---|---|---|---|
| **GetMinds / Minds** | AI persona panel platform | Concept & message testing at scale | Stimulus (concept, ad, copy), persona selection | Multi-persona reactions, scored & open-ended | Panel scale & speed | LLM-generated personas with weak grounding | Medium — useful for panel UX patterns | Concept upload + multi-persona side-by-side |
| **Uxia** | AI UX testing for prototypes/screens | Pre-test UX review | Figma / screen / URL, persona | Structured UX issues, attention/annotation | Pre-test screen review | Cannot measure real task success | Medium-High — directly informs UX Feedback module | Screen annotation & A/B comparison |
| **Synthetic Users** | Synthetic interview participants | Hypothesis generation, exploratory research | Interview brief, persona attributes | AI-generated interview transcripts, follow-ups | Interview-style simulation | Representational ≠ predictive validity; cultural blind spots | High — directly competes for "synthetic research panel" use case | Follow-up probing mechanism + structured hypothesis report |
| **Simile AI** | Research-grounded digital replicas | Augmenting qualitative research with replayable persona | Real interview transcripts, persona attributes | Grounded persona answers with source links | Evidence-grounded persona answers | Bounded by source quality and recency | **Highest — closest to moomoo's intended model** | Replica-from-interview pattern + source traceability |
| **Generic LLM DIY (ChatGPT / Claude direct)** | General-purpose LLM | Anything, including persona role-play | Free-form prompts | Free-form text | Zero setup cost | No grounding, no audit, no compliance posture, no team workflow | Low — already used informally; risk of ungrounded use in financial contexts | Demonstrates demand; informs why a governed internal tool is needed |
| **Dovetail / Great Question** *(adjacent reference)* | Research repository / panel ops | Storing transcripts, themes, tags; recruiting | Real research artifacts | Searchable themes, tags, video clips | Source-of-truth research repository | Not a persona/synthetic system; no AI simulation | High as a **data source / integration partner**, not a competitor | Tagging & evidence model; quote-linking; theme clustering |
### 7.6 Synthesis — what moomoo should build
- The **Simile model** (research-grounded replicas) is the right backbone.
- The **GetMinds UX** (multi-persona panel, concept upload) is the right interaction pattern for panel work.
- The **Uxia capability** (screen-level pre-flight review) is the right interaction pattern for design work.
- The **Synthetic Users mechanism** (probing follow-ups, calibration claims) is the right interaction pattern for interview simulation — but moomoo must avoid over-claiming validity.
- The **financial-services safeguards** (no advice, regulated-claim guardrails, traceable quotes, compliance review mode) are unique to moomoo and not present in any external tool.
---
## 8. Product Principles
1. **Evidence or it didn't happen.** A persona response without retrievable evidence is labeled low-confidence and visually distinguished.
2. **Synthetic ≠ real.** Every synthetic output carries a "synthetic" badge; every real-user quote is traceable to its source.
3. **Show your work.** Citations are first-class UI, not a footnote.
4. **Markets are not interchangeable.** A US persona's defaults do not transfer to JP / HK / SG / MY / AU without explicit cross-market evidence.
5. **No financial advice, ever.** Hard guardrail at model output; reinforced by post-generation classifier.
6. **Researcher in the loop.** Persona creation, validation, and major output review require human researcher sign-off.
7. **Recency matters.** Evidence has a freshness score; stale assets are flagged or down-weighted.
8. **Fail loudly.** When the platform cannot answer with grounding, it says so rather than confabulating.
9. **Build for audit.** Every output is reproducible and explainable on request.
10. **Internal only at MVP.** External exposure is a separate, larger compliance review.
---
## 9. Information Architecture
```
moomoo Internal Persona Platform (MIPP)
│
├── 1. Persona Library ── Browse, filter, create personas
├── 2. Persona Chat ── Single & panel Q&A with personas
├── 3. Synthetic Panel ── Run concept / prototype reactions
├── 4. Research Design Assistant── Guides, screeners, tasks
├── 5. UX / Prototype Feedback ── Screen-level AI pre-review
├── 6. Research Knowledge Base ── Upload, search, retrieve assets
└── 7. Governance / Admin ── Access, audit, compliance, validation
```
### 9.1 Cross-module primitives
- **Persona object** (shared across modules)
- **Evidence chunk** (shared across modules)
- **Project / workspace** (groups outputs by initiative)
- **Citation object** (binds outputs to evidence chunks)
- **Confidence + freshness scores** (rendered consistently)
---
## 10. Detailed Feature Requirements
Priority levels: **P0** = MVP must-have, **P1** = first major post-MVP release, **P2** = roadmap.
### 10.1 Module 1 — Persona Library
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| PL-01 | Persona creation wizard | P0 | Guided flow to create a new persona from selected research evidence | Researcher can pick ≥3 evidence assets, AI drafts persona profile, researcher edits & publishes |
| PL-02 | Persona profile page | P0 | Canonical view: demographics, behaviors, attitudes, market context, evidence list, version history | All fields render; evidence list links to source; last-updated visible |
| PL-03 | Market filter | P0 | Filter by US / JP / HK / SG / MY / AU / future | Filter returns only personas tagged for that market |
| PL-04 | Product usage filter | P1 | Filter by product surface (options, futures, ETFs, event contracts, margin, AI tools, membership) | Filter returns personas tagged with selected products |
| PL-05 | Trading maturity filter | P1 | New investor / casual / active / professional | Filter returns personas matching maturity band |
| PL-06 | Risk appetite tag | P1 | Conservative / balanced / aggressive / speculative | Tag selectable in profile and filterable |
| PL-07 | Asset level tag | P1 | Tiers (e.g., <10k / 10–100k / 100k–1M / 1M+ in local equivalent) | Tag selectable, filterable, market-aware |
| PL-08 | Competitor usage tag | P1 | Tag prior/current usage of Robinhood, Webull, SBI, Rakuten, IBKR, etc. | Multi-select tag, filterable |
| PL-09 | Research evidence links | P0 | List of evidence chunks bound to persona, with type (interview, survey, VOC, analytics) | Each link opens evidence preview with source |
| PL-10 | Persona version history | P1 | All published versions retained; diff view | Researcher can view diff between any two versions, revert if needed |
| PL-11 | Confidence score | P0 | 0–100 score with breakdown by attribute coverage and evidence quality | Score visible on profile; methodology one-click explainable |
| PL-12 | Evidence freshness score | P1 | Score based on age + asset weight; visible per persona | Score auto-recomputes when new evidence is bound or assets age past threshold |
| PL-13 | Persona retirement | P2 | Mark personas as deprecated / superseded | Deprecated personas hidden from default views; redirected to successor |
| PL-14 | Persona templates | P2 | Reusable scaffolds for common archetypes per market | New persona wizard offers template selection |
| PL-15 | Bulk import of legacy personas | P2 | Import from existing slide decks / Dovetail | Imported personas marked as "unvalidated" until reviewed |
| PL-16 | Persona similarity search | P2 | Find personas similar to a given one across markets | Returns top-N similar personas with similarity rationale |
**Anchor personas for MVP (12 personas: 4 per MVP market):**
| Market | Persona | Notes |
|---|---|---|
| US | Active options trader | High activity; multi-leg strategies; competitor: tastytrade, Robinhood |
| US | Event contract curious user | New use case; influenced by Kalshi, Polymarket exposure |
| US | Robinhood migration user | Migrated for better tools; price-sensitive; commission-aware |
| US | AI-feature power user | Early adopter of moomoo AI tools; influences product feedback loops |
| Hong Kong | High-AUM multi-market investor | HK + US + A-shares portfolio; uses moomoo for cross-border execution |
| Hong Kong | A-share cross-border investor | Northbound / Stock Connect focus; sensitive to FX and tax friction |
| Hong Kong | Young digital-first investor | 20s–30s; Telegram/小红书-influenced; comparison-shops with Futubull / Tiger |
| Hong Kong | IPO subscription enthusiast | High activity around HK IPO calendar; margin-financing aware |
| Japan | Margin-trading hesitant user | Risk-averse; influenced by SBI/Rakuten conservatism |
| Japan | SBI / Rakuten migration user | Price- and UX-driven migration |
| Japan | NISA-driven long-term investor | Tax-advantaged account focus; conservative product mix |
| Japan | US-equities-curious Japan investor | Increasing interest in US stocks; FX-cost and regulation aware |
**Phase 2 anchor personas (SG / MY / AU — for reference):**
| Market | Persona | Notes |
|---|---|---|
| Singapore | New investor | First-time investor; Telegram/forum culture influence |
| Malaysia | Islamic-finance-sensitive investor | Shariah-compliance consideration |
| Australia | ETF-focused long-term investor | Superannuation context; ETF-default mindset |
### 10.2 Module 2 — Persona Agent Chat
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| PC-01 | Ask one persona | P0 | Threaded Q&A with a single persona | Responses include citations; persona voice consistent within thread |
| PC-02 | Ask multiple personas as a panel | P1 | Send the same question to N personas; side-by-side responses | Up to 8 personas; rendered in parallel columns or grouped cards |
| PC-03 | Compare persona responses | P1 | Highlight similarities & differences across panel | Diff view; "what they agreed on / disagreed on" summary |
| PC-04 | Follow-up questions | P1 | Multi-turn within persona context | Context preserved per persona; can branch threads |
| PC-05 | Citations from source research | P0 | Every factual claim linked to source evidence chunk | Hover/click shows source quote and asset; copyable citation |
| PC-06 | Mark unsupported claims | P1 | Flag spans of response that lack grounding | Visual underline / badge; tooltip explains why unsupported |
| PC-07 | Uncertainty warnings | P0 | If retrieval confidence below threshold, show banner | Banner visible above response; downgrades persona confidence for that turn |
| PC-08 | Export response summaries | P1 | Export thread as PDF/MD with citations preserved | Export retains citations and labels |
| PC-09 | Save useful Q&A to persona profile | P2 | Pin a Q&A to the persona's "frequently-asked" section | Pinned items appear on profile; researcher review required |
| PC-10 | Persona "would not know" responses | P0 | Persona declines questions outside its knowledge scope | Out-of-scope detection; suggests alternative persona |
| PC-11 | No-advice guardrail | P0 | Hard block on personalized investment advice | Output classifier blocks + audit log entry |
| PC-12 | Multilingual response (per market) | P1 | Respond in market language (EN / JA / ZH-HK / ZH-CN / MS) | Researcher can toggle response language; persona voice preserved |
### 10.3 Module 3 — Synthetic Research Panel
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| SP-01 | Select target personas | P0 | Multi-select from library, including filter-driven cohorts | Cohorts saved; persona count visible |
| SP-02 | Upload concept / prototype / image / text / PRD | P0 | Multi-format stimulus upload | Accepts PNG/JPG, PDF, MD, Figma link; text excerpt; URL |
| SP-03 | Structured question batteries | P1 | Likert / multiple-choice / rank questions auto-aggregated | Aggregated chart per question; per-persona detail on click |
| SP-04 | Open-ended simulation | P0 | Persona reactions as free text | Each response cited; synthetic badge applied |
| SP-05 | Ranking / scoring tasks | P1 | Personas rank options or score concepts | Aggregate ranking with disagreement view |
| SP-06 | Compare reactions across persona groups | P1 | Group by market / maturity / risk tag | Group-level summary + drill-down |
| SP-07 | Hypothesis summary | P0 | LLM-generated synthesis with confidence | Explicit "hypothesis, not evidence" label; cites contributing personas |
| SP-08 | Auto-generate real-user validation questions | P1 | Convert synthetic findings into questions to validate with real users | Hand-off to Research Design Assistant |
| SP-09 | Synthetic-only labeling | P0 | All outputs visibly labeled as synthetic | Labels present on screen and on export |
| SP-10 | Panel templates | P2 | Reusable panels for recurring tests (e.g., "new-investor panel") | Saved templates loadable in one click |
| SP-11 | Diversity / coverage warnings | P1 | Warn if selected panel is too narrow (e.g., all US, no JP) | Warning banner with suggested additions |
| SP-12 | Calibration history | P1 | Link panel run to subsequent real-user findings for calibration | Calibration view shows what synthetic got right/wrong |
### 10.4 Module 4 — Research Design Assistant
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| RD-01 | Generate interview guide | P0 | From topic + persona + market → discussion guide | Researcher rates guide ≥4/5 in pilot; covers warm-up / core / probes / wrap-up |
| RD-02 | Generate usability test tasks | P1 | From feature + persona → realistic task scenarios | Tasks are persona-appropriate, market-localized, and behavior-anchored |
| RD-03 | Generate screener questions | P1 | From target persona → screener (with quotas) | Generates qualifying + disqualifying questions; quota suggestions |
| RD-04 | Generate follow-up probes | P1 | From draft question → suggested probes | Probes attached inline in guide editor |
| RD-05 | Identify potential bias in questions | P0 | Flag leading, loaded, double-barreled, jargon-heavy questions | Detected at draft time; suggested rewrite offered |
| RD-06 | Recommend target segments | P1 | Suggest personas to recruit based on research goal | Top 3-5 with rationale |
| RD-07 | Recommend sample quota | P2 | Suggest quotas balancing market / maturity / experience | Quota table generated; methodology explained |
| RD-08 | Generate discussion guide by market | P1 | Market-localized version (language + cultural framing) | Reviewed by local researcher; markers for local idiom |
| RD-09 | Generate research risks & validation plan | P1 | Highlight what synthetic missed, recommend validation approach | Risks list with severity; validation steps |
| RD-10 | Guide library | P2 | Save / share / fork research guides | Searchable; reusable templates per study type |
### 10.5 Module 5 — UX / Prototype Feedback Assistant
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| UX-01 | Upload screenshots or Figma links | P0 | Multi-screen upload; ordered flow | Order preserved; per-screen annotation possible |
| UX-02 | Select persona reviewers | P0 | Choose personas to "walk through" the screens | At least 1, up to 6 per review |
| UX-03 | Generate comprehension feedback | P0 | What persona understood vs misunderstood per screen | Per-screen, per-persona feedback with citations to comprehension evidence from real research |
| UX-04 | Identify unclear terms | P1 | Highlight jargon / unclear copy per persona | Term-level highlight; alternative phrasing suggested |
| UX-05 | Identify trust barriers | P1 | Flag what would make persona hesitate to proceed | Cited against prior trust-related research |
| UX-06 | Identify risky financial wording | P0 | Flag copy that sounds like advice, guarantee, or regulated claim | Flag escalates to compliance review queue |
| UX-07 | Identify likely drop-off points | P1 | Per-screen drop-off prediction (hypothesis only) | Labeled as hypothesis; not behavioral measurement |
| UX-08 | Suggest UX improvements | P1 | Specific rewrite / layout suggestions | Suggestions tied to identified issue |
| UX-09 | Generate real usability test tasks | P1 | Hand-off to RD-02 to test in real session | Tasks pre-populated in Research Design Assistant |
| UX-10 | Compare A/B versions | P2 | Side-by-side review across two versions | Per-screen diff; persona preference summary (hypothesis) |
| UX-11 | "Not real task completion" disclaimer | P0 | Persistent disclaimer on every UX report | Visible on screen and on export |
### 10.6 Module 6 — Research Knowledge Base
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| KB-01 | Upload research documents | P0 | PDF, DOCX, MD, transcript text, CSV (surveys), JSON (VOC) | Successful upload + parse confirmation |
| KB-02 | Auto-tag by market, product, persona, theme, source type | P1 | LLM-assisted tagging with human review queue | ≥80% precision in tagging pilot; reviewer can override |
| KB-03 | Evidence extraction | P0 | Extract evidence chunks (claim + quote + source span) | Chunks indexed for retrieval |
| KB-04 | Quote extraction | P1 | Pull verbatim user quotes with attribution | Quote retains speaker context and source location |
| KB-05 | Insight clustering | P2 | Cluster similar insights across studies | Cluster view with representative quotes |
| KB-06 | Search by persona | P1 | Filter all evidence by persona binding | Returns chunks where persona is bound |
| KB-07 | Search by product topic | P1 | Topic-based search (e.g., "margin trading onboarding") | Top-N relevant chunks with snippets |
| KB-08 | Search by market | P1 | Market-scoped retrieval | Returns market-tagged chunks |
| KB-09 | Ask questions across assets | P0 | Natural-language Q&A over the knowledge base | Cited answer with linked chunks |
| KB-10 | Evidence citation | P0 | Every retrieved answer cites source assets and chunks | Citations clickable; hover shows source snippet |
| KB-11 | Source freshness indicator | P1 | Per-chunk and per-asset freshness score | Visible in retrieval UI; stale results visually de-emphasized |
| KB-12 | Integrations: Dovetail, Lark/Feishu, internal data lake | P2 | Pull research / analytics assets natively | Integration health visible in admin |
| KB-13 | VOC ingest pipeline | P2 | Continuous ingest from VOC channels & app reviews | Daily refresh; per-market segmentation |
| KB-14 | Behavior data binding | P2 | Bind aggregate behavior signals to personas | Anonymized cohort signals only; PII stripped |
| KB-15 | Multilingual indexing | P0 | EN / JA / ZH-HK / ZH-CN / MS / native scripts | Cross-lingual retrieval works in pilot |
### 10.7 Module 7 — Governance, Risk, and Compliance
| ID | Feature | Priority | Description | Acceptance criteria |
|---|---|---|---|---|
| GV-01 | Clear labeling of synthetic outputs | P0 | "Synthetic" badge on all generated content | Badge persists through export |
| GV-02 | No-investment-advice classifier | P0 | Post-generation classifier blocks advisory outputs | Blocked outputs logged with reason |
| GV-03 | Unsupported-claim detector | P1 | Flag generation spans lacking retrievable grounding | Spans visually marked; user can request grounding retry |
| GV-04 | PII protection | P0 | Detect & redact PII in uploads and outputs | PII scan on every upload; redaction logged |
| GV-05 | Access control by team and market | P0 | RBAC + market-level data scoping | Users see only authorized markets/assets |
| GV-06 | Source permission management | P1 | Per-asset access control; inherited by derived outputs | Restricted asset access does not leak via persona answer |
| GV-07 | Audit log | P0 | Immutable log of queries, outputs, citations, edits | Searchable by user, persona, asset, date |
| GV-08 | Human review workflow | P1 | Flagged outputs route to reviewer queue | Reviewer can approve / reject / request edit |
| GV-09 | Evidence traceability | P0 | Every output → evidence chunks → source assets | One-click trace for any output |
| GV-10 | Confidence and uncertainty display | P1 | Show confidence per output and per attribute | Confidence visible; methodology explainable |
| GV-11 | Model output disclaimer | P0 | Per-module disclaimers, per-output where applicable | Disclaimer visible on screen and on export |
| GV-12 | Data retention policy | P1 | Configurable retention by asset type & market regulation | Retention enforced; deletions logged |
| GV-13 | Compliance review mode | P1 | Mode that surfaces only compliance-flagged outputs for review | Reviewer dashboard with backlog and SLA |
| GV-14 | Real-vs-synthetic calibration tracking | P2 | Track how often synthetic predictions match real outcomes | Calibration metric per persona; visible on profile |
| GV-15 | Quote authenticity guarantee | P0 | Real quotes always traceable; synthetic quotes always labeled | Quote-authenticity test passes in audit |
| GV-16 | Per-market compliance profiles | P2 | Different guardrails per regulator (SEC, HKMA, MAS, JFSA, ASIC, SC Malaysia) | Profile loads correct guardrails per market scope |
| GV-17 | Red-team test harness | P1 | Recurring adversarial tests for advice leakage, fabrication, jailbreaks | Weekly suite; results logged |
---
## 11. Persona Data Model
The persona is the central object in MIPP. It is intentionally richer than a traditional marketing persona because three downstream systems consume it: (a) the LLM-grounded chat agent, (b) the synthetic panel simulation, (c) the compliance / audit layer. Every field below exists for one of these consumers, and every field is bound to evidence.
### 11.1 Field-group overview
| Group | Purpose | Consumed by |
|---|---|---|
| `identity` | Stable IDs, naming, status, version | All |
| `market_context` | Market, regulator scope, language, locale norms | Chat, Panel, Compliance |
| `demographics` | Age, income, employment, education, household | Panel, Calibration |
| `life_stage` | Major financial-life events that shape decisions | Panel |
| `investing_profile` | Maturity, risk appetite, AUM, products, horizon | Chat, Panel |
| `behaviors` | Observed actions: trading frequency, channels, research habits | Chat, Panel |
| `information_diet` | Where they hear about products / form opinions | Chat, Panel |
| `attitudes` | Beliefs, trust posture, AI openness, fee sensitivity | Chat, Panel |
| `competitor_usage` | Current / prior / considered platforms with switching reasons | Chat, Panel |
| `product_mental_models` | How they conceptualize each product they touch | Chat |
| `pain_points` / `goals` / `jobs_to_be_done` | Motivational layer | Chat, Panel |
| `regulatory_exposure` | Account types, restrictions, sophistication classification | Compliance |
| `cultural_context` | Locale-specific framing, taboos, idiom | Chat, Compliance |
| `evidence_bindings` | Source of every claim in this persona | All |
| `scores` | Confidence, freshness, coverage breakdown | Calibration |
| `calibration_history` | Prior synthetic vs. real-research alignment record | Calibration, Governance |
| `governance` | Owner, validation cadence, review SLA | Governance |
| `version_history` | Edit log | Governance |
### 11.2 Full schema example (US Active Options Trader)
```yaml
persona:
identity:
id: persona_us_active_options_trader
display_name: "US Active Options Trader"
short_handle: us-active-options
status: published # draft | published | deprecated | under_review
version: 4
persona_archetype: active_self_directed # new | casual | active | professional | institutional_lite
tags: [options, multi-leg, high-frequency, competitor-aware]
market_context:
primary_market: US
secondary_markets: []
cross_market_relevance: low
regulator_scope: [SEC, FINRA]
language_preferences: [en-US]
locale_norms:
currency: USD
date_format: MM/DD/YYYY
number_format: "1,234.56"
time_zone_typical: America/New_York
risk_disclosure_tone: "direct, U.S.-style mandatory disclosure"
demographics:
age_band: 28-45
gender_distribution: {male: 0.78, female: 0.20, other: 0.02}
income_band: 100k-250k_usd
employment: employed_full_time
education: bachelors_or_higher
urbanicity: major_metro
household:
household_size_typical: 2
dependents_typical: 0-1
primary_breadwinner_likely: true
notes: "Tech-adjacent professionals; high digital literacy"
life_stage:
typical_life_events:
- id: ls_001
event: "Job change with vested equity grants"
relevance: "Triggers reallocation; surfaces tax-loss-harvest interest"
- id: ls_002
event: "First mortgage in past 24 months"
relevance: "Raises cash-management vs. options-margin tension"
five_year_horizon: "Wealth acceleration via options income strategies"
investing_profile:
maturity: active # new | casual | active | professional
risk_appetite: aggressive # conservative | balanced | aggressive | speculative
asset_level_band: 100k-1M_usd
primary_products: [options, equities]
secondary_products: [futures, event_contracts]
holding_horizon: short_to_medium
leverage_use: routine # never | occasional | routine | core
typical_position_count: 8-15
avg_trade_size_band: 2k-10k_usd
behaviors:
trading_frequency: multiple_per_day
session_count_per_day_typical: 4-8
primary_device: desktop
secondary_device: mobile
research_habits: ["uses screeners", "follows IV ranks", "Twitter/X finance", "uses options chain heatmaps"]
decision_style: ["data-driven", "speed-focused", "thesis-then-action"]
moomoo_features_used: [options_chain, level_2, watchlist, paper_trading_for_testing]
moomoo_features_avoided: [community_posts, copy_trading]
information_diet:
primary_signal_sources: ["Twitter/X cashtags", "tastytrade YouTube", "options-focused Discord", "Bloomberg headlines"]
peer_influence: medium
influencer_followed_examples: ["@unusual_whales-style accounts"]
moomoo_community_engagement: low_read_only
receptiveness_to_in_app_education: low_for_basics_high_for_advanced
attitudes:
trust_in_platform: high_conditional # contingent on uptime + execution quality
sensitivity_to_fees: high
openness_to_AI_tools: medium_high
skepticism_of_marketing: high
privacy_posture: pragmatic # max_privacy | pragmatic | indifferent
competitor_usage:
current: [moomoo]
prior: [tastytrade, robinhood]
considered: [IBKR, thinkorswim]
switching_drivers:
- "Execution latency on multi-leg"
- "Options-specific tooling depth"
- "Margin rate transparency"
switching_blockers:
- "Effort to move ACATs"
- "Loss of existing keyboard shortcuts / muscle memory"
product_mental_models:
options: ["thinks in greeks", "uses spreads, not just calls", "expects IV-rank context built-in"]
event_contracts: ["sees as adjacent to options", "questions liquidity and settlement clarity"]
margin: ["expects competitive intraday rates", "compares to IBKR tiered"]
ai_features: ["wants concrete, citable signals not vague summaries"]
pain_points:
- id: pp_001
claim: "Execution latency on multi-leg orders is a deal-breaker"
severity: high
frequency: routine
evidence: [evidence_chunk_id_001, evidence_chunk_id_017]
- id: pp_002
claim: "Margin rate disclosures buried two screens deep"
severity: medium
frequency: occasional
evidence: [evidence_chunk_id_042]
goals:
- "Maximize edge per unit of time"
- "Reduce friction in multi-leg execution"
- "Avoid execution surprises"
jobs_to_be_done:
- statement: "When IV pops, help me act in under 10 seconds"
situation: "Volatility spike during market hours"
motivation: "Capture mispricing window"
outcome_measure: "Order placed within 10 seconds of decision"
regulatory_exposure:
sophistication_classification: experienced_investor
options_approval_level_typical: tier_3 # market-specific (e.g., U.S. Level 1-4)
pdt_exposure: yes_account_above_25k
margin_account: enabled
self_directed: true
advisor_relationship: none
restricted_products: []
cultural_context:
communication_style_preference: terse_technical
humor_register_appropriate: dry_industry_in_group
taboos_to_avoid_in_persona_voice: ["personalized investment advice", "celebrity-stock hype"]
market_specific_idioms: ["the tape", "size-on", "theta decay"]
moomoo_brand_perception: "trusted execution + AI-curious"
evidence_bindings:
- asset_id: research_2025_q4_us_options_study
chunks: [chunk_001, chunk_017, chunk_042]
contribution_weight: 0.45
- asset_id: voc_us_2026_q1
chunks: [chunk_113, chunk_209]
contribution_weight: 0.15
- asset_id: analytics_us_options_cohort_2026
chunks: [agg_001, agg_002]
contribution_weight: 0.20
- asset_id: competitor_research_robinhood_2026
chunks: [chunk_004]
contribution_weight: 0.10
- asset_id: support_tickets_us_options_2026_q1
chunks: [theme_001]
contribution_weight: 0.10
scores:
confidence_score: 78
freshness_score: 84
coverage_breakdown:
demographics: 90
life_stage: 60
investing_profile: 88
behaviors: 85
information_diet: 70
attitudes: 75
competitor_usage: 70
product_mental_models: 65
regulatory_exposure: 95
cultural_context: 60
evidence_balance:
qualitative_share: 0.50
quantitative_share: 0.35
voc_share: 0.10
competitor_share: 0.05
calibration_history:
runs:
- run_id: cal_2026_03
synthetic_panel_run: panel_run_us_options_concept_2026_03
confirmed: 7
refuted: 1
partial: 3
unaddressed: 2
accuracy_score: 0.74
- run_id: cal_2026_05
synthetic_panel_run: panel_run_us_options_membership_2026_05
confirmed: 9
refuted: 0
partial: 2
unaddressed: 1
accuracy_score: 0.86
drift_alerts:
- last_real_research_date: 2026-05-04
max_recommended_staleness_days: 120
governance:
owner: research_lead_us
co_owner: product_owner_us_options
last_validated_by: researcher_001
last_validated_on: 2026-05-12
next_review_due: 2026-08-12
review_cadence_days: 90
access_scope: [team_research_global, team_product_us_options, team_design_us]
pii_compliance_state: clear
retention_policy: standard_36_months
version_history:
- {version: 1, created: 2026-01-10, by: researcher_001, summary: "Initial draft from Q4 2025 study"}
- {version: 2, created: 2026-02-22, by: researcher_001, summary: "Added VOC + analytics binding"}
- {version: 3, created: 2026-04-05, by: researcher_002, summary: "Refined mental models after IBKR comparison"}
- {version: 4, created: 2026-05-12, by: researcher_001, summary: "Calibration update after concept test"}
```
### 11.3 Field rules
- **Every claim-bearing field** (behaviors, attitudes, pain_points, mental_models) MUST have at least one entry in `evidence_bindings`. The platform refuses to publish a persona with unbound claims.
- **Sensitive fields** (`regulatory_exposure.sophistication_classification`, `pdt_exposure`, `restricted_products`) are write-restricted to compliance-reviewed personas only.
- **Cultural & locale fields** are mandatory for any persona used in chat or panel — they directly shape persona voice.
- **Calibration history** is append-only; it cannot be edited, only added to by Workflow 6.
- **Coverage scores** are computed, not manually set; researchers cannot override them.
### 11.4 Evidence chunk model
```yaml
evidence_chunk:
id: chunk_017
source_asset_id: research_2025_q4_us_options_study
source_type: interview_transcript # interview | survey | voc | review | support | analytics | competitor | market_report
source_market: US
speaker_context: "Participant P07, age 34, trades options daily"
text: "If I can't leg into a spread in 3 clicks, I'm out — that's why I left Robinhood."
themes: [execution_speed, multi_leg, competitor_switch]
bound_personas: [persona_us_active_options_trader]
collected_on: 2025-11-04
freshness_score: 88
permissions:
visibility: [team_research_global, team_product_us_options]
pii_status: redacted
```
---
## 12. AI System Design
### 12.1 High-level architecture
```
┌─────────────────────────────────────────────────────────────────┐
│ Application Layer (Web UI) │
└───────┬────────────────────────────────────────────────┬─────────┘
│ │
┌───────▼─────────┐ ┌───────▼─────────┐
│ Orchestration │ │ Governance │
│ & Routing │ │ Gateway │
│ │ │ (advice block, │
│ - Module logic │ │ PII scan, │
│ - Persona bind │ │ ungrounded │
│ - Multi-agent │ │ detection, │
│ panel │ │ audit log) │
└───────┬─────────┘ └───────▲─────────┘
│ │
┌───────▼──────────────────────────────────────────────┐ │
│ Retrieval Layer │ │
│ - Persona-scoped retrieval │ │
│ - Market-scoped retrieval │ │
│ - Freshness-weighted ranking │ │
│ - Permission-aware filtering │ │
│ - Cross-lingual retrieval │ │
└───────┬───────────────────────────────────────────────┘ │
│ │
┌───────▼────────────────┐ ┌──────────────────┐ ┌──────▼────────┐
│ Vector Index │ │ Structured │ │ Frontier LLM │
│ (evidence chunks, │ │ Persona Store │ │ (Claude / │
│ quotes, themes) │ │ + Tag Index │ │ Anthropic │
└────────────────────────┘ └──────────────────┘ │ Claude 4.7) │
└───────────────┘
```
### 12.2 Model selection
| Use case | Model class | Rationale |
|---|---|---|
| Persona chat (primary) | Frontier reasoning model (Claude Opus / Sonnet class) | Quality, nuance, persona consistency |
| Bulk panel simulation | Mid-tier model (Claude Haiku class) | Cost-efficient at scale; consistent across personas |
| Tagging / classification (PII, advice, ungrounded) | Specialized classifiers + smaller LLMs | Latency and reliability |
| Embedding / retrieval | Production multilingual embedding model | Cross-market retrieval |
| Translation / localization | Frontier model with native-language fine-tune review | Cultural fidelity in market language |
### 12.3 Prompt-engineering pillars
1. **Persona system prompt** is auto-assembled from the persona object (not hand-written per query).
2. **Evidence-injection contract**: persona responses receive top-K retrieved chunks with required citation tokens.
3. **Refusal contract**: persona must respond "outside my knowledge / weak evidence" when retrieval scores are low.
4. **No-advice contract**: persona never gives personalized investment recommendations regardless of question framing.
5. **Synthetic-labeling contract**: any quote generated by the persona is labeled `[synthetic-quote]`.
### 12.4 Prompt caching
Persona system prompts and evidence-binding skeletons are large and reused — prompt caching is required to manage cost and latency. Cache hit rate is a tracked operational metric.
### 12.5 Multi-agent panel orchestration
For panel runs (Module 3) and multi-persona compares (Module 2):
- Each persona is instantiated as an isolated agent with its own context.
- A meta-agent runs after all persona responses are collected to produce the cross-persona synthesis.
- The meta-agent is forbidden from adding any claim not present in a persona response.
### 12.6 Calibration loop
After real-user research is conducted on a topic previously simulated:
1. Researcher tags real findings against the prior synthetic panel run.
2. System computes agreement rate per persona (predictive calibration metric).
3. Persona confidence is updated; evidence bindings can be re-weighted.
---
## 13. Evidence Grounding and Citation Logic
### 13.1 Grounding requirement
Every persona-attributable factual claim must be one of:
- **Grounded** — supported by ≥1 retrieved evidence chunk above similarity threshold T.
- **Inferred** — extrapolated from grounded evidence; labeled `[inferred]`.
- **Refused** — claim cannot be made; persona declines.
- **Hypothesis** (synthetic panel only) — explicitly synthetic; labeled and counted toward "to validate" list.
### 13.2 Citation rendering
| Element | UI treatment |
|---|---|
| Grounded claim | Normal text + numbered citation `[1]` linking to evidence chunk preview |
| Inferred claim | Italic + `[inferred]` badge |
| Synthetic quote | Quote in box + `synthetic` badge + persona attribution |
| Real quote | Quote in box + `verified` badge + source asset link + speaker context |
| Unsupported span | Underline + warning icon + tooltip explaining absence |
### 13.3 Citation object
```yaml
citation:
citation_id: c_4f1a
output_span: "execution latency is the main reason this persona considers leaving"
bound_evidence:
- chunk_id: chunk_017
similarity: 0.84
contribution: primary
- chunk_id: chunk_113
similarity: 0.71
contribution: supporting
generated_at: 2026-06-02T10:14:22Z
persona_id: persona_us_active_options_trader
module: persona_chat
audit_id: audit_2026_06_02_00091
```
### 13.4 Failure modes & responses
| Failure | System response |
|---|---|
| Retrieval returns no chunks above threshold | Refuse; suggest research gap |
| Top chunks are stale (freshness < 40) | Respond with stale-evidence warning; recommend fresh research |
| Top chunks conflict | Acknowledge disagreement; surface both views |
| User question is about unstudied segment | Refuse; recommend recruiting / studying |
| Question elicits advice-shaped response | No-advice classifier blocks output |
---
## 14. User Workflows
> **A note on workflow diagrams.** Each workflow below is illustrated with a Mermaid flowchart and references the closest competitor pattern that inspired it. We deliberately borrowed from validated UX patterns rather than inventing flows from scratch. Where MIPP diverges from the competitor (usually for evidence-grounding or compliance), the diagram makes the deviation explicit.
### 14.1 Workflow 1 — Researcher creates a new persona
**Scenario:** A US researcher creates "US Event Contract Curious User" combining prior interviews, competitor research, survey data, and VOC.
**Competitor reference:**
- **Synthetic Users — Persona Builder** ([syntheticusers.com](https://www.syntheticusers.com/)): wizard-style persona creation from inputs.
- **Dovetail — Insights & Personas** ([dovetail.com](https://dovetail.com/)): real-research-grounded persona pages with evidence trails.
- **MIPP deviation:** evidence binding is mandatory and AI drafts are never published without a confidence + freshness score computed from real research weights.
```mermaid
flowchart TD
A["Researcher opens
Persona Library"] --> B["Click 'Create Persona'"]
B --> C["Pick market + seed traits
(maturity, behavior hint)"]
C --> D["System retrieves candidate evidence
from KB: interviews · surveys · VOC · competitor"]
D --> E{"Researcher reviews
each evidence chunk"}
E -- accept --> F["AI drafts persona fields
(every field cites evidence)"]
E -- reject --> D
F --> G["Researcher edits draft;
flags low-evidence fields as 'to-validate'"]
G --> H["System computes
confidence + freshness + coverage"]
H --> I{"Confidence
≥ threshold?"}
I -- no --> J["Persona stays in DRAFT
Researcher gathers more evidence"]
I -- yes --> K["Publish v1
Audit log records binding + scores"]
K --> L["Persona available in library
to permitted teams"]
J --> D
classDef ai fill:#fde2c4,stroke:#c2630e
classDef gov fill:#ddd5f0,stroke:#5039a8
class D,F,H ai
class K,L gov
```
**Step detail:**
1. Researcher opens Persona Library → "Create Persona".
2. Selects market = US, candidate maturity = casual–active, candidate behavior = "interested in event contracts but hasn't traded one."
3. System retrieves candidate evidence: 4 interview clips, 2 competitor (Kalshi/Polymarket) research notes, 1 survey segment, 3 VOC themes from the past 90 days.
4. Researcher reviews and accepts/removes chunks.
5. AI drafts demographics, behaviors, attitudes, pain points, and mental models — each field with attached evidence.
6. Researcher edits the draft, marks low-evidence fields as "to-validate."
7. System computes confidence (e.g., 62/100) and freshness (88/100).
8. Researcher publishes v1. Audit log records action, evidence binding, scores.
9. Persona is added to library, available to all permitted teams.
### 14.2 Workflow 2 — PM asks persona panel for concept feedback
**Scenario:** A US PM uploads an Event Contract concept page and asks 5 personas to react.
**Competitor reference:**
- **Synthetic Users — Multi-Participant Studies** ([syntheticusers.com](https://www.syntheticusers.com/)): runs the same prompt across N synthetic participants, returns per-participant transcripts + aggregate themes.
- **GetMinds — Mind Panels** ([getminds.ai](https://getminds.ai/)): the canonical "ask a panel" pattern with side-by-side persona responses.
- **MIPP deviations:** (1) every persona response is grounded in real research and cites it; (2) the platform refuses to produce a "decision" — only hypotheses + validation questions; (3) the auto-generated "questions to validate with real users" list is unique to MIPP and hands off to Workflow 3.
```mermaid
flowchart TD
A["PM opens Synthetic Panel
'New panel run'"] --> B["Select N personas
(from library or saved cohort)"]
B --> C["Upload stimulus
PNG · Figma · PDF · MD · URL"]
C --> D["Pick question battery
open-ended only (P0)"]
D --> E["Run panel
each persona responds in isolation"]
E --> F["Per-persona response
with citations + synthetic badge"]
F --> G["LLM hypothesis summary
labeled 'synthetic hypothesis'"]
G --> H["Auto-generate
'validate with real users' question list"]
H --> I["Export report
(citations + synthetic labels preserved)"]
I --> J["Hand off to Workflow 3:
build real usability test"]
classDef compliance fill:#fde2df,stroke:#b6261c
classDef ai fill:#fde2c4,stroke:#c2630e
class G,H ai
class F,I compliance
```
**Step detail:**
1. PM opens Synthetic Panel → "New panel run."
2. Selects 5 personas: US active options trader, US event contract curious user, US Robinhood migration user, HK young digital-first investor, US AI-feature power user.
3. Uploads concept page (PNG) and a 1-page concept brief (MD).
4. Selects question battery: open-ended ("first impression," "what's confusing," "would you try"). (Structured ranking is P1 — out of MVP.)
5. Runs panel. Each persona responds in isolation; system compiles results in ~60–120s.
6. PM sees per-persona responses and an LLM-generated hypothesis summary labeled "synthetic hypothesis."
7. System auto-generates a "questions to validate with real users" list.
8. PM exports the report and shares with research lead. Research lead schedules a real-user concept test.
### 14.3 Workflow 3 — Researcher generates a real usability test plan
**Scenario:** Using synthetic feedback on the Event Contract concept, researcher generates a real-user test plan.
**Competitor reference:**
- **Uxia — Research Plan Generator** ([uxia.ai](https://uxia.ai/)): generates interview guides + tasks from a stimulus.
- **Maze — Test Builder** ([maze.co](https://maze.co/)): task / probe / quota patterns.
- **Dovetail — Magic** ([dovetail.com](https://dovetail.com/)): AI-augmented guide drafting from prior research.
- **MIPP deviation:** the guide is seeded directly from a synthetic-panel run — every probe targets a specific synthetic hypothesis. This is something no current competitor does, because no competitor connects synthetic and real-user research as one loop.
```mermaid
flowchart TD
A["Researcher opens
prior synthetic-panel run"] --> B["Click 'Generate real
usability test plan'"]
B --> C["System emits draft plan:"]
C --> C1["Target personas + sample sizes per market"]
C --> C2["Screener + quotas"]
C --> C3["Usability tasks (anchored in persona behaviors)"]
C --> C4["Probes targeting each synthetic hypothesis"]
C --> C5["Bias check on every draft question"]
C --> C6["Risk plan — 'what synthetic likely missed'"]
C1 --> D["Researcher edits in guide editor"]
C2 --> D
C3 --> D
C4 --> D
C5 --> D
C6 --> D
D --> E["Publish guide"]
E --> F["Panel ops recruits via screener"]
F --> G["Real fielding"]
G --> H["Results returned to MIPP
for calibration → Workflow 6"]
classDef ai fill:#fde2c4,stroke:#c2630e
classDef loop fill:#d9efd6,stroke:#2c7d29
class C,C1,C2,C3,C4,C5,C6 ai
class H loop
```
**Step detail:**
1. Researcher opens prior synthetic run (Workflow 2 output).
2. Clicks "Generate real usability test plan."
3. System suggests target personas + sample sizes, screener + quotas, usability tasks, probes targeting the synthetic hypotheses, bias-check on each draft question, risk plan (what synthetic likely missed).
4. Researcher edits in the guide editor.
5. Researcher publishes guide; panel ops team uses screener to recruit.
6. After fielding, results are returned to the platform for calibration (Workflow 6).
### 14.4 Workflow 4 — Designer runs pre-test UX review
**Scenario:** A designer uploads a new moomoo membership page and asks active trader and HK high-AUM investor personas to review.
**Competitor reference:**
- **Uxia — Synthetic UX Review** ([uxia.ai](https://uxia.ai/)): per-screen synthetic walkthrough with persona reactions.
- **Maze — AI Insights on Prototype Tests** ([maze.co](https://maze.co/)): aggregated comprehension flags.
- **Simile AI — Persona Walkthrough** ([simile.ai](https://simile.ai/)): per-persona reaction text per screen.
- **MIPP deviations:** (1) risky financial wording auto-routes to a compliance review queue (no competitor has this); (2) every flag links back to real-research comprehension evidence rather than being LLM-imagined; (3) outputs always carry the "not real task completion" disclaimer.
```mermaid
flowchart TD
A["Designer opens
UX Feedback · 'New review'"] --> B["Upload screens
Figma links or PNG, ordered flow"]
B --> C["Select reviewer personas
(1–6)"]
C --> D["Run review"]
D --> E["For each screen × each persona:"]
E --> E1["Comprehension flags
(cited to comprehension evidence)"]
E --> E2["Risky financial wording"]
E --> E3["UX improvement suggestion"]
E2 --> F{"Flag severity"}
F -- high --> G["Auto-route to
compliance review queue"]
F -- low --> H["Stays in designer report"]
E1 --> H
E3 --> H
H --> I["Export report
(synthetic + 'not real task completion' disclaimer)"]
I --> J["Optional: 'Generate real usability tasks'
→ hand off to Workflow 3"]
classDef compliance fill:#fde2df,stroke:#b6261c
classDef ai fill:#fde2c4,stroke:#c2630e
class E,E1,E2,E3 ai
class G,I compliance
```
**Step detail:**
1. Designer opens UX Feedback module → "New review."
2. Uploads Figma frame links (or PNG export) for the membership page flow (3 screens).
3. Selects 2 reviewer personas: US active options trader, HK high-AUM investor.
4. Runs review. System generates per-screen feedback per persona (comprehension, risky wording, improvement suggestion). Unclear-terms / trust-barrier / drop-off are P1 — out of MVP.
5. Designer exports report. Compliance-flagged items auto-routed to compliance queue.
6. Designer clicks "Generate real usability tasks" to hand off to research.
### 14.5 Workflow 5 — Business stakeholder queries the knowledge base
**Scenario:** Business lead asks: "Why do Japan users hesitate to use margin trading?"
**Competitor reference:**
- **Dovetail — Magic / Ask** ([dovetail.com](https://dovetail.com/)): natural-language Q&A across research repositories.
- **Notion AI Q&A** ([notion.com/product/ai](https://www.notion.com/product/ai)): conversational answers over a workspace.
- **Glean** ([glean.com](https://glean.com/)): enterprise NL search with citation cards.
- **MIPP deviations:** (1) every retrieved answer must cite an evidence chunk — uncited LLM speculation is rejected; (2) market access scopes the corpus automatically per the requester's RBAC; (3) answers link to a persona when one exists, so insights become reusable.
```mermaid
flowchart TD
A["Lead opens Knowledge Base
natural-language search"] --> B["Question: 'Why do JP users
hesitate to use margin trading?'"]
B --> C["RBAC scopes corpus
to user's market access"]
C --> D["Retrieval: interviews + surveys
+ VOC + support · last 12 months"]
D --> E{"Each retrieved chunk
passes relevance threshold?"}
E -- no --> F["Banner: 'limited evidence'
(no fabrication)"]
E -- yes --> G["LLM synthesizes structured answer:"]
G --> G1["Top themes"]
G --> G2["Real verbatim quotes
(speaker context)"]
G --> G3["Linked persona(s)"]
G --> G4["Source assets · dates · freshness"]
G --> G5["Confidence on synthesis"]
G1 --> H["Render answer card"]
G2 --> H
G3 --> H
G4 --> H
G5 --> H
H --> I["Export for exec deck
(citations preserved)"]
classDef compliance fill:#fde2df,stroke:#b6261c
classDef ai fill:#fde2c4,stroke:#c2630e
class C,F compliance
class D,G,G1,G2,G3,G4,G5 ai
```
**Step detail:**
1. Lead opens Knowledge Base → search bar.
2. Asks the question in natural language. System scopes to JP market by default given user's market access.
3. System retrieves and synthesizes from JP interviews, surveys, support tickets, and VOC over the last 12 months.
4. Returns a structured answer with top themes, verbatim quotes, linked persona, source assets, and synthesis confidence.
5. Lead exports for an executive deck; export retains all citations.
### 14.6 Workflow 6 — Validate synthetic findings against real research
**Scenario:** After running the Event Contract concept test, researcher compares real findings with the prior synthetic panel.
**Competitor reference:**
- **No direct competitor analogue.** No current synthetic-research tool closes the loop with real research; their value prop is "skip real research." MIPP's calibration loop is the moat — it is the only mechanism that lets synthetic-panel accuracy improve over time, and it converts synthetic runs from one-shot guesses into a learning system.
- **Loose analogue:** weights & biases-style experiment tracking ([wandb.ai](https://wandb.ai/)), applied to research hypotheses instead of ML metrics.
```mermaid
flowchart TD
A["Researcher opens
prior synthetic-panel run"] --> B["Click 'Compare with real-user findings'"]
B --> C["Upload / link real research synthesis"]
C --> D["Align synthetic hypotheses
with real findings, hypothesis-by-hypothesis"]
D --> E1["Confirmed"]
D --> E2["Refuted"]
D --> E3["Partial"]
D --> E4["Unaddressed"]
E1 --> F["Update per-persona
predictive accuracy score"]
E2 --> F
E3 --> F
E4 --> F
F --> G["Adjust confidence on
affected persona attributes"]
G --> H["Bind new real-research
evidence chunks to persona"]
H --> I["Append to persona
calibration_history"]
I --> J["Audit log: calibration event"]
J --> K["Next synthetic-panel run
uses updated persona → loop closes"]
classDef loop fill:#d9efd6,stroke:#2c7d29
classDef gov fill:#ddd5f0,stroke:#5039a8
class F,G,H,K loop
class I,J gov
```
**Step detail:**
1. Researcher opens prior synthetic panel run.
2. Clicks "Compare with real-user findings."
3. Uploads or links the real-research synthesis.
4. System aligns synthetic hypotheses with real findings, marking each Confirmed / Refuted / Partial / Unaddressed.
5. Per-persona predictive accuracy updates the persona's calibration history.
6. Confidence scores for affected attributes adjust accordingly.
7. New evidence chunks from the real research bind to the persona.
8. Audit log records the calibration event.
---
## 15. MVP Scope
### 15.1 MVP markets
| Market | In MVP? |
|---|---|
| US | ✅ MVP |
| Hong Kong | ✅ MVP |
| Japan | ✅ MVP |
| Singapore | Post-MVP (Phase 2) |
| Malaysia | Post-MVP (Phase 2) |
| Australia | Post-MVP (Phase 2) |
**Rationale for top-3 prioritization:**
- **US** — Largest English-speaking market; deepest existing research corpus; sharpest competitive pressure (Robinhood, Webull, tastytrade, Schwab); event-contract whitespace driving 2026 product roadmap.
- **Hong Kong** — Highest-AUM cohort globally for moomoo; cross-border (HK + US + A-shares) complexity that no competitor persona platform can model; strategic gateway for greater-China expansion.
- **Japan** — Distinctive cultural & regulatory posture (JFSA, NISA, conservative risk framing); SBI/Rakuten competitor displacement campaign requires sharp local persona modeling; largest non-English market.
Singapore moves to Phase 2 — strong research signal already exists, but lower AUM density and smaller addressable cohort than HK in the same time window.
### 15.2 MVP modules
> P0 = ship-blocker for core value delivery or hard compliance rule. The scope below covers only P0 features; everything labeled P1/P2 in §10 is explicitly out of MVP.
| Module | MVP scope (P0 only) |
|---|---|
| Persona Library | Persona creation wizard from real research evidence; canonical persona profile page; market filter; evidence links; confidence score. 4 anchor personas per MVP market = 12 personas at launch. |
| Persona Chat | Single-persona threaded Q&A with citations, uncertainty warnings, out-of-scope refusal, and no-advice guardrail. (Multi-persona panel, follow-ups, compare view: P1) |
| Synthetic Panel | Persona/cohort selection; concept/PRD/image upload; open-ended simulation; hypothesis summary with confidence; synthetic-only labeling. (Structured questions, ranking, real-vs-synthetic compare, handoff to UX: P1) |
| Research Design Assistant | Interview guide generation; bias-in-questions detector. (Screeners, usability tasks, probes, translation: P1) |
| UX / Prototype Feedback | Screen / Figma upload; persona reviewer selection; comprehension feedback per screen; risky-financial-wording flag; persistent "not real task completion" disclaimer. (Unclear terms, trust barriers, drop-off, A/B: P1) |
| Knowledge Base | Document ingest; evidence extraction; NL Q&A across assets; clickable evidence citations; multilingual indexing (EN / JA / ZH / MS). (Auto-tagging, source-balance dashboard, freshness signals: P1) |
| Governance | Synthetic-output label; no-advice classifier; PII detect/redact; RBAC + market scoping; audit log; evidence traceability; per-output disclaimer; quote authenticity guarantee. (Confidence display, retention scheduler, red-team harness: P1) |
### 15.3 Out of MVP
- Multi-persona chat panel, follow-up questions, compare view (PC-02 / PC-03 / PC-04 — P1)
- Structured-question synthetic panel, ranking, real-vs-synthetic compare, UX handoff (SP-03 / SP-05 / SP-06 / SP-08 — P1)
- Screener / usability tasks / probe library / guide translation (RD-02 / RD-03 / RD-04 / RD-08 — P1)
- UX unclear-terms, trust-barrier, drop-off flags; A/B compare (UX-04 / UX-05 / UX-07 / UX-09 / UX-10 — P1/P2)
- Persona auto-tagging, source-balance & freshness signals (KB-04 / KB-06 / KB-07 / KB-08 / KB-11 — P1)
- Confidence display per output, retention scheduler, red-team harness (GV-06 / GV-08 / GV-10 / GV-12 / GV-13 — P1)
- Dovetail / Lark / data-lake integrations (Phase 2)
- VOC continuous ingest pipeline (Phase 2)
- Per-market compliance profiles (Phase 2)
- Persona similarity search, calibration history visualization (P2 / Phase 3)
### 15.4 MVP success bar (go/no-go)
| Criterion | Bar |
|---|---|
| ≥10 active researchers across 3 markets | Weekly active by week 8 of pilot |
| ≥3 product / design teams running synthetic panels | By week 10 |
| ≥5 documented decisions informed (not made) by the platform | By week 12 |
| Zero compliance incidents | Throughout pilot |
| Researcher rating ≥4/5 on guide generation | Post-pilot survey |
---
## 16. Future Roadmap
### Phase 2 (post-MVP, +3 months)
- Add HK, MY, AU markets.
- Panel templates; A/B compare in UX Feedback.
- Dovetail / Lark / data lake integrations.
- VOC continuous ingest.
- Per-market compliance profiles.
- Calibration history visualization.
- Red-team test harness.
### Phase 3 (+6 months)
- Persona similarity search.
- Behavior data binding (anonymized cohort signals).
- Multilingual response polish (native-language QA loops).
- Mobile-friendly read-only review mode.
- Inline Slack/Lark plugin for "ask a persona" from chat.
### Phase 4 (+9–12 months)
- Cross-platform persona portability (with strict access controls).
- Persona-driven prioritization scoring for product backlog.
- Real-time analytics-to-persona linkage (with anonymization safeguards).
- Optional: limited external partner access (separate compliance scope).
---
## 17. Success Metrics
### 17.1 Adoption metrics
| Metric | Target (post-MVP +90d) |
|---|---|
| Weekly active researchers | ≥80% of moomoo global research team |
| Weekly active PMs / designers | ≥50 across product orgs |
| Synthetic panel runs / week | ≥30 |
| Knowledge Base queries / week | ≥200 |
### 17.2 Quality metrics
| Metric | Target |
|---|---|
| Researcher rating: persona quality | ≥4/5 |
| Researcher rating: generated guides | ≥4/5 |
| Citation rate (% of responses with ≥1 citation) | ≥95% |
| Synthetic-hypothesis confirmation rate after real testing | Tracked, no target year 1 |
| Unsupported-claim flag rate | <5% of generated outputs |
### 17.3 Impact metrics
| Metric | Target |
|---|---|
| Time-to-hypothesis for new concepts | Median <1 hour |
| Time-to-first-research-guide draft | Median <30 min |
| % of new research projects citing ≥3 historical assets | ≥60% |
| Cross-market insight citations / quarter | ≥30 |
| Number of decisions documented as platform-informed (not platform-decided) | Tracked monthly |
### 17.4 Safety / governance metrics
| Metric | Target |
|---|---|
| Advice-classifier block rate (true blocks) | Tracked; investigated if 0 |
| Compliance-flagged outputs reaching reviewer SLA | 100% within 1 business day |
| Audit log completeness | 100% of outputs auditable |
| Synthetic-label preservation rate on export | 100% |
| PII detection precision/recall | ≥95% / ≥98% in evaluation set |
---
## 18. Risks and Mitigation
| # | Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|---|
| R1 | Teams over-trust synthetic outputs as evidence | High | High | Always-on "synthetic" labeling; mandatory disclaimers; calibration tracking; research-led enablement |
| R2 | Persona drift as base LLM updates | Medium | Medium | Persona regression eval suite; version-pinning of evaluator prompts; re-validation after each model upgrade |
| R3 | Ungrounded generation leaks despite guardrails | Medium | High | Post-generation classifier; red-team harness; human review queue for high-risk modules |
| R4 | Compliance incident (advice / regulated claim) | Low–Medium | Critical | No-advice classifier; per-market compliance profiles; pre-launch legal review; ongoing monitoring |
| R5 | Stale evidence misleading decisions | Medium | Medium | Freshness scoring; stale-evidence warnings; mandatory review cadence per persona |
| R6 | Underrepresentation of small markets | Medium | Medium | Diversity warnings on panel selection; sponsor research investment in under-covered markets |
| R7 | PII leak via uploaded transcripts | Low–Medium | Critical | PII scan on upload; redaction default; permission-aware retrieval |
| R8 | Cost overrun from large panel runs | Medium | Medium | Tiered model routing; prompt caching; per-team usage budgets |
| R9 | Researcher backlash ("AI replacing us") | Medium | Medium | Research-team-led design; framing as augmentation; researchers own persona quality |
| R10 | Synthetic outputs cited in regulatory submissions externally | Low | Critical | Internal-only enforcement; export watermarking; no external API at MVP |
| R11 | Bias amplification in personas (cultural / demographic) | Medium | High | Bias audit suite; per-market researcher review; diversity warnings |
| R12 | Loss of researcher tacit knowledge if reliance grows | Medium | Medium | Mandatory researcher review on persona creation/updates; calibration loop |
---
## 19. Compliance and Governance
### 19.1 Financial-services safeguards (hard rules)
1. **No personalized investment advice.** The system will never tell a user (synthetic or real) what to invest in. Post-generation classifier blocks any advice-shaped output. Reviewer audit weekly.
2. **No fabricated regulatory or product claims.** Any regulatory or product claim must be backed by an approved source. Approved sources are tagged in the Knowledge Base.
3. **No fabricated user quotes.** Real user quotes are traceable to source assets; synthetic quotes are visually labeled `[synthetic-quote]` and never appear in exports without the label.
4. **No simulation of personalized investment recommendations.** A persona may discuss preferences in general terms (e.g., "I tend to prefer broad-market ETFs"), but the system never returns "you should buy X."
5. **Real quote integrity.** Real user quotes always carry source citation and speaker context.
### 19.2 Privacy & PII
- PII scan on every upload; redact-by-default for PII fields.
- Researcher upload flow requires confirmation that PII has been handled per policy.
- PII never enters embedding indexes or persona outputs.
- Data retention configurable per asset type & per market regulation.
### 19.3 Access control
- RBAC by team and market.
- Source permission management cascades to derived outputs.
- Restricted assets cannot be cited by personas in unauthorized scopes.
- Admin can revoke user access; audit log preserves history.
### 19.4 Audit & traceability
- Every output has an audit ID.
- Every audit entry links to user, persona, retrieved chunks, source assets, model + version, prompt cache key.
- Audit retained per data-retention policy; exportable for compliance review.
### 19.5 Compliance review mode
- Dedicated reviewer dashboard.
- Flagged outputs: advice-shaped, regulated-claim-shaped, low-grounding, PII-suspect.
- Reviewer SLA: 1 business day.
- Reviewer actions: approve / reject / request edit / escalate.
### 19.6 Per-market compliance posture (initial)
| Market | Primary regulators | Notes |
|---|---|---|
| US | SEC, FINRA | Strict on advice & promotional content |
| Japan | JFSA | Localization of disclosures; conservative copy posture |
| Singapore | MAS | Risk warnings on margin / leveraged products |
| Hong Kong | SFC, HKMA | Cross-border investor considerations |
| Malaysia | SC Malaysia, Shariah considerations | Shariah-aware persona for relevant products |
| Australia | ASIC | Anti-greenwashing & advice rules |
### 19.7 Human-in-the-loop requirements
- Persona creation requires researcher sign-off.
- Persona updates require versioning + owner review.
- Major synthetic panel reports require researcher annotation before circulation.
- Compliance-flagged outputs require compliance reviewer sign-off before reuse.
---
## 20. Open Questions
1. **Persona ownership** — Does each persona have a single research owner per market, or a co-owner model across regional + global research?
2. **Calibration scoring methodology** — How do we score "synthetic predicted real" when real research findings are themselves qualitative and probabilistic? Do we adopt a simple confirmed/partial/refuted scheme, or build a more granular alignment score?
3. **Cross-market persona portability** — When is a US persona's evidence reusable for SG? Default to "never" or default to "with explicit cross-market evidence weight"?
4. **VOC inclusion thresholds** — How fresh / how voluminous must VOC be to influence a persona? Risk of single-event VOC distorting persona attitudes.
5. **Behavior data binding** — At what aggregation level can behavior data bind to a persona without creating a re-identification risk? Need DPO + DS alignment.
6. **Model provider strategy** — Single-provider (e.g., Anthropic Claude family) vs multi-provider with provider-portable prompts. Affects caching, eval, and contract negotiation.
7. **Inline integration depth** — Should the platform live as a web app, a Lark/Feishu plugin, a Figma plugin, or all of the above? MVP recommends web-app only.
8. **Synthetic panel size limits** — Is 8 personas the right cap? Diminishing returns vs cognitive load on the reader.
9. **External researcher access** — Are external research agencies (recruited per-project) ever granted access, or strictly internal? MVP recommends strictly internal.
10. **Persona retirement policy** — When markets shift (e.g., regulation change), what triggers automatic deprecation vs reviewer-driven deprecation?
11. **Right-to-be-forgotten flow** — When a research participant requests deletion, how does the platform purge embeddings, retrieval chunks, and any derived persona attributes that materially depended on that participant?
12. **Cost governance** — Per-team budgets vs central budget; how do we expose cost transparently without disincentivizing exploration?
---
## 21. Appendix
### A. Glossary
| Term | Definition |
|---|---|
| Persona | An evidence-grounded representation of a user segment, queryable as an AI agent |
| Persona agent | The runtime instantiation of a persona for a query (Module 2) or panel (Module 3) |
| Evidence chunk | A retrievable unit of research evidence (interview span, survey segment, VOC theme, support theme, analytics aggregate) |
| Grounded claim | Output supported by ≥1 retrieved evidence chunk above similarity threshold |
| Synthetic hypothesis | Output generated without real-user evidence; always labeled synthetic |
| Calibration | Comparing synthetic predictions to subsequent real-research findings |
| Confidence score | Persona-level (and per-attribute) measure of evidence coverage and quality |
| Freshness score | Time-decay-weighted measure of how recent supporting evidence is |
| RBAC | Role-based access control |
| Market scope | The set of markets a user is permitted to view |
| No-advice guardrail | Classifier + system-prompt contract preventing personalized investment advice |
| Compliance review mode | Reviewer workspace for flagged outputs |
### B. Example moomoo-specific scenarios
| Scenario | Modules involved |
|---|---|
| Pre-flight a new Event Contract concept page | Synthetic Panel, UX Feedback, Research Design Assistant |
| Localize a margin trading risk disclosure for JP | Knowledge Base, UX Feedback, Governance |
| Design a study on AI trading-tool adoption | Knowledge Base, Persona Chat, Research Design Assistant |
| Decide whether the new membership page resonates with HK high-AUM users | Synthetic Panel, UX Feedback |
| Compare ETF-product positioning across AU and SG | Knowledge Base, Persona Chat (panel) |
| Understand why options users churn to competitor X | Knowledge Base, Persona Chat |
### C. Open evaluation harness (proposed)
- **Grounding eval**: % of factual claims with retrievable evidence.
- **Persona regression eval**: held-out Q&A set per persona; compare answers across model versions.
- **No-advice red-team eval**: weekly adversarial set; track block rate.
- **Citation correctness eval**: human-reviewed sample of citations weekly.
- **Cross-market bias eval**: balanced question set across markets; check for US-default bias.
### D. Assumptions
1. moomoo has rights to ingest its own research, VOC, support, and analytics assets internally for AI processing under existing data-handling agreements (legal confirmation required pre-build).
2. A frontier-class LLM (Claude Opus / Sonnet) is available with prompt caching and reasonable rate limits.
3. Research team has bandwidth to seed 4 anchor personas per MVP market (12 personas total) over the first 8 weeks of build.
4. Compliance / legal can dedicate ≥1 reviewer per MVP market for the pilot.
5. An internal data infra team can provide retrieval / vector index infrastructure to security/compliance standard.
### E. Out of scope (explicit)
- Customer-facing chatbot or end-user-visible AI agent.
- Auto-generated marketing copy without human review.
- Trading signal generation.
- Market sizing / forecasting.
- Customer support automation.
- Recruitment automation (covered by panel-ops tools, not MIPP).
---
**End of PRD v1.0.** Comments, edits, and module-level deep-dives welcome via the platform Lark group and the PRD review thread.