How we decided
Every feature was tested against five questions, in this order. A feature that failed any one of them left the plan or changed shape.
- Is it legal, and allowed by the platforms it depends on?
- Is it safe for members, including the ones who will misuse it?
- Is it current, or has the market already shown it does not work?
- Does it rely on people we would have to run: hosts, staff, moderators, venue partners?
- Is it worth its cost at the size we will actually be?
The fourth question arrived late. Plan v1.0 answered the first three well and still depended on paid hosts, venue deals and a moderation team. Plan v2.0 is the result of asking the fourth question of everything.
The self-serve rethink: v1.0 to v2.0
Research suggested paid hosts and platform-run tables. It was right that they create warmth and density. We cut them anyway, because the platform must not rely on people it runs, and because the more a platform controls a meetup, the more it owes for what happens there.
| Item | Before | Now | Why | Reason type |
|---|---|---|---|---|
| AI host composing tables of strangers | AI composes tables of 4–6 strangers every week | AI planner drafts; the member hosts; open seats on member plans | Composing groups and picking venues puts the platform in control of the meetup, and control brings a duty of care. It also needs weekly operations one founder cannot run. | Relies on people we would run |
| Paid table hosts, hub hosts and city ambassadors | Paid on show-up, about CAD 4–12k a month | None; organizers are members | Research suggested paying hosts on show-up; we cut it because the platform must not rely on people it runs. It was the biggest cost line and added contractor-law risk. Organizers are ordinary members, rewarded with reach and free Plus, never cash. | Relies on people we would run |
| Platform-run events and a paid-host launch pilot | Eight-week host-led pilot | Invite-only waitlist city-wide; hubs, chat and voice first | Research suggested an eight-week hosted pilot in two neighborhoods; we cut it because the platform must not rely on people it runs. Launch is invite-only and city-wide, with online surfaces first. | Relies on people we would run |
| Venue partnerships | 10+ venues signed before launch | None; public places from open map data | Signed venues are deals to manage and make the platform choose places for strangers. The AI suggests public places from open map data instead. | Relies on people we would run |
| Deposits and seat fees for meetups | Phase 2 revenue line | Dropped | Taking money for attendance makes the platform a party to the meetup, and BC refund rules on subscriptions and prepaid services made the model harder. Members pay for their own coffee. | Illegal or against terms |
| A staffed moderation team with a 24-hour review target | Staff implied | AI-first; vendor humans on demand | A team is people to hire and manage. AI-first moderation with graduated actions covers the routine; vendor reviewers on demand cover appeals and child-abuse hash matches. | Relies on people we would run |
| Safety check-ins sent to the platform | Countdown check-ins to the platform | Trusted contacts plus one-tap 911 | A platform that monitors check-ins takes on a duty to act and needs people on watch. Check-ins go to the member's own trusted contacts, with one-tap 911. | Relies on people we would run |
| Share Plan without the other person's consent | Shares the match's details | Needs the other person's consent | Sharing a match's details with a third party is personal information under BC and federal privacy law. The flow now asks the other person. | Illegal or against terms |
| Behavioral learning on by default | On by default | Explicit opt-in | What someone views and clicks can reveal sexual orientation, a sensitive inference. Behavioral learning became opt-in. | Unsafe |
| AI acting for members | A separate tier where AI could act | Drafts inside Plus; the member always sends | Nothing acts for a member. AI may draft a reply inside Plus; the member reads it and sends it. Only 4–13% of random pairs would accept AI-sent messages anyway. | Unsafe |
| Verification across Canada at launch | Canada-wide | BC and the rest of Canada first; Quebec after the filing | Quebec requires a biometric database to be declared 60 days before use. Verification is off in Quebec until the filing is done. | Illegal or against terms |
| A separate video vendor and a Next.js and Redis stack | Next.js, Redis and a video vendor | Cloudflare Workers, Durable Objects, R2, Queues, Images, Realtime and Postgres | One platform for API, rooms, media, queues and voice means fewer bills, generous free tiers and less to operate. | Cost |
| App-store billing | App stores at 15–30% | Web billing in Phase 2 | App stores take 15–30%. Web billing costs a few percent, and the MVP is free, so no merchant account is needed to launch. | Cost |
Cuts made in v1.0
These left the plan when the research was consolidated. Most carry legal exposure, safety risk or cost that a first-city MVP cannot absorb; a few are simply outdated.
| Item | Verdict | Why | Reason type |
|---|---|---|---|
| AI photo albums, share cards and short AI videos of members | Cut | Deepfake laws in the US, UK, EU and Canada, a public backlash that paused a similar feature at a major app in 2026, and cost. Only a labeled toon avatar survives, in Phase 2. | Illegal or against terms |
| AI-generated nudity or sexual content | Cut | Illegal in several markets without verified consent, against card-network rules, and open to abuse if verification is spoofed. Dropped permanently: it is one of the hard rules. | Illegal or against terms |
| Public video roulette, and the word "roulette" | Cut | Omegle shut down in 2023 after a lawsuit over matching a child with a predator. Replaced by voice speed-meets for verified adults and video by consent. | Unsafe |
| Live maps, exact location and "crossed paths" trails | Cut | Stalking risk, and precise location is sensitive data. Location is approximate for everyone and exact only for confirmed attendees. | Unsafe |
| AI agents messaging, creating accounts or uploading photos | Cut | Prompt injection, low acceptance by the people on the other end, and agents at another startup made profiles nobody asked for. | Unsafe |
| Boosts, "see who liked you" and priority likes | Cut | Paying to be seen is the most resented model in the category and fights the idea of rooms over profiles. | Outdated |
| Age- or profile-based pricing, outcome guarantees, lockouts over chargebacks | Cut | A USD 60.5M age-pricing settlement in California and a USD 14M FTC order against a market leader. One price for everyone. | Illegal or against terms |
| An infinite video or photo feed | Cut | Video-feed dating apps stalled or closed, and a feed competes with meetups for attention. | Outdated |
| Personality typing as matching science | Cut | Personality-type tests have weak validity for predicting relationships. Understanding comes from a short interview, Mirror and what people do. | Outdated |
| Matching people from their Discord servers | Cut | Discord's Developer Policy bans profiling users or their relationships. Communities can post events; no profiling. | Illegal or against terms |
| Scraping competitors' event pages | Cut | A terms-of-service breach. Organizers post their own events instead. | Illegal or against terms |
| Embedded music players on profiles | Deferred | The main streaming widget's terms restrict commercial use. Profile songs are a link-out, pressed by the viewer, until counsel clears embeds. | Illegal or against terms |
| Self-hosted GPUs at MVP | Deferred | Operations burden. Commercial model APIs are cheaper until volume justifies hardware. | Cost |
| Nine separate microservices | Changed | One modular application with the same boundaries ships faster and splits later. | Cost |
| Avatar rooms | Deferred | No evidence of demand; live audio rooms faded once the novelty went without recurring groups. Revisit in Phase 3 with data. | Outdated |
Earlier
| Item | Verdict | Why | Reason type |
|---|---|---|---|
| ReActor face-swap for avatars | Cut | The leading open-source face-swap tool and its models are not licensed for commercial use, and face swapping sits too close to deepfakes. Dropped early, in v0.2. | Illegal or against terms |
What did not change
Verification with the face optional, profiles and Mirror, hubs and personals, chat with free translation, the consent ledger, free safety tools and voice speed-meets survived every version. So did the hard rules: no AI-generated nudity or sexual content, AI never impersonates a member or acts as a romantic companion, consent is never delegated to AI, minors are banned everywhere.
Not legal advice. Laws here are described as what shaped the design, as of the date on this page. Check with counsel in your own jurisdiction.