TastePolicy
AI that helps without replacing you.
TastePolicy is the permission layer for taste-aware AI.
Qloo tells an agent what culturally fits a person. TastePolicy decides whether the agent may actually go that far — given the person’s real-world requirements, the facts we can verify, and the authority they explicitly granted.
Margaret’s mobility changed. Her taste didn’t.
Margaret is 72. She still loves intimate live jazz, classic cinema and Italian dining. But today she also needs a step-free entrance, a short walk, a quieter setting and an evening within budget.
A generic AI can produce a plausible recommendation. Qloo can identify what genuinely fits Margaret’s culture and taste. TastePolicy answers the question that becomes critical once AI starts acting for people:
May the agent actually go this far with that choice, for this person, right now?
Start here
- Watch the 1:56 demo: https://youtu.be/T0oudn4f-L8
- Judge walkthrough: https://tastepolicy.vercel.app/judge
- Public Margaret demo: https://tastepolicy.vercel.app/demo/margaret
- Use TastePolicy as yourself: https://tastepolicy.vercel.app/experience
- Product: https://tastepolicy.vercel.app
- Public source: https://github.com/VictorCyberDev/TastePolicy
The judge walkthrough and Margaret demo require no login. The personal experience is account-backed so every person uses only their own taste, requirements, rules and history.
The 30-second judge story
- Margaret’s five favourites are resolved to real Qloo entities.
- Qloo discovers places that genuinely fit her taste.
- A required fact is missing, so TastePolicy returns ASK instead of guessing.
- Margaret supplies the missing fact. The same place is re-evaluated and can become ALLOW — but only up to her saved action ceiling.
- If the agent tries to go further and present the place as the final choice, TastePolicy returns ASK again.
- If a different venue violates a hard requirement, TastePolicy returns REJECT, even when Qloo says it is a strong cultural match.
That is the product in one sentence:
Taste intelligence is not permission.
Why this is not another recommendation app
Most recommendation systems stop once they know what you might like.
TastePolicy separates five things that are usually collapsed together:
cultural fit → factual evidence → requirements → authority ceiling → requested action
The result is one of four honest outcomes:
- ALLOW — requirements are satisfied and the requested action stays within the person’s authority ceiling.
- ASK — a critical fact is unknown, or the agent is trying to go further than the person allowed.
- REJECT — a known hard requirement or explicit rule is violated.
- UNAVAILABLE — evidence required for the decision cannot currently be obtained.
Missing evidence never silently becomes permission. Provider failure never becomes “the user dislikes this.” Cultural affinity never overrides an explicit boundary.
Qloo is structurally indispensable
TastePolicy does not recreate or imitate Qloo’s taste graph.
The live product uses the official Qloo API to:
- resolve a person’s cultural anchors to real Qloo entities;
- verify those identities;
- perform cross-domain discovery from those anchors;
- evaluate specific candidates with affinity and explainability.
Qloo is the cultural-evidence layer. TastePolicy is the authority layer that governs what an agent may do with that evidence.
Without Qloo, TastePolicy intentionally does not substitute an LLM guess and pretend it is equivalent cultural intelligence.
One candidate, multiple outcomes
TastePolicy’s clearest proof is that the same culturally strong candidate can move through different states without changing Margaret’s taste:
Unknown walking time → ASK
Margaret answers “8 minutes.”
Same place + new evidence → ALLOW SHORTLIST
Agent attempts “present as the final choice.”
Same place + action beyond authority ceiling → ASK
Different venue has no step-free entrance.
Strong Qloo fit + hard requirement violation → REJECT
This is why TastePolicy is not a scoring wrapper around Qloo. It is a deterministic authority system around Qloo-powered cultural intelligence.
Cultural Receipts: why this happened
Every decision can expose a receipt that keeps the layers separate:
- Qloo cultural evidence
- place facts
- the person’s requirements
- the person’s authority setting
- the requested action
- the final ALLOW / ASK / REJECT / UNAVAILABLE decision
Presentation media never masquerades as factual proof. A venue image can help someone understand a place, but it cannot silently satisfy an accessibility requirement.
A real person can use TastePolicy as themselves
TastePolicy is not only a Margaret judge demo.
A new user can:
- create an account;
- enter five favourites in ordinary language;
- resolve them with live Qloo and confirm the intended identities;
- save what matters to them;
- choose how far AI may go without asking;
- receive Qloo-grounded recommendations;
- see ASK / ALLOW / REJECT outcomes;
- inspect why each decision happened;
- return later to persisted personal state.
Real-user pages never silently inherit Margaret or another demo persona.
Real place experience
TastePolicy’s place layer uses open-data components including MapLibre/OpenFreeMap, OpenStreetMap, KartaView and Panoramax where coverage exists.
Media is explicitly labelled by what it actually represents — exact place, nearby street, area context or illustrative context — and is never used as hidden requirement evidence.
The model generalizes beyond Margaret
Margaret is the clearest story, not the limit of the system.
The same authority contract supports additional functional scenarios such as:
- sensory-aware outings with lower-noise requirements;
- low-energy days with walking-time limits.
TastePolicy represents functional requirements rather than inferring medical conditions from cultural taste.
Engineering depth
The human product sits on one deterministic authority contract exposed through multiple surfaces:
- Next.js / React / TypeScript
- Qloo API with credentials kept server-side
- PostgreSQL persistence
- versioned rules and action ceilings
- Cultural Receipts
- bounded authority leases and drift checks
- portable signed proofs
- REST
- TypeScript SDK
- remote MCP
- A2A
- OAuth 2.1 / PKCE
- automated unit, browser, accessibility, conformance and hosted acceptance testing
The final M26 source passed 953 unit tests and the full Chromium E2E suite at 190/190. The real-user production flow was also exercised with live Qloo on desktop and mobile.
Safety boundary
TastePolicy grants preference-choice authority only.
An ALLOW decision never grants payment, booking, messaging, account modification, private-data disclosure or legal authority. Those require separate execution permission.
Memory is not permission. Affinity is not authorization. A verified receipt is not authorization.
Why this matters
AI agents are moving from answering questions toward choosing on behalf of people: what to watch, where to go, what to buy, what to shortlist and what to put into plans.
As those agents become better at understanding people, they also become more capable of overstepping.
Qloo answers: what fits this person?
TastePolicy answers: may this agent go this far with that fit?
That is the category TastePolicy is building: personal agency infrastructure for culturally intelligent AI.
Built With
- a2a
- better-auth
- kartaview
- maplibre
- mcp
- next.js
- oauth-2.1
- openstreetmap
- panoramax
- postgresql
- qloo
- react
- typescript
- vercel