# ๐๏ธ Two Doorways
**Civic forecasting infrastructure for neighborhoods. Prediction markets are the mechanism โ not the point.** Resolves from the city's own data, and knows the difference between an expert and an insider.
[](https://two-doorways.vercel.app)
[](https://two-doorways.vercel.app/agent)
[](https://two-doorways.vercel.app)
Built for Frontier Forecast 2026 ยท SoMa, San Francisco.
---
## ๐ก The one-liner
One market question has two doorways. To a trader tracking regional CPI, it's a mispriced probability to correct. To a tenant, it's an early read on next year's rent โ before the Rent Board makes it official. Same price, two motivations โ accuracy from one side, foresight from the other. Neither is gambling.
## Why this, why here
Cities make million-dollar decisions on surveys that take months and dashboards that are stale the day they ship. Cities already know their own future โ that knowledge is real, distributed, and unpriced. There's no instrument for it, only rumor. Two Doorways continuously fuses public data, local expertise, and accountable forecasting into a living signal โ and settles it from public infrastructure instead of a bookmaker's say-so. AI agents participate too, through the same MCP surface humans use, bound by the exact same governance โ no separate, looser rules for machines.
## The domain edge (the part nobody else in the room has)
San Francisco rent moves in three regulatory tiers, and knowing which is which separates a live market from a dead one:
| Tier | Behavior | Predictable? |
|------|----------|--------------|
| Rent Ordinance (pre-1979) | Increase = published CPI formula; landlords always max it in a hot market | **Deterministic โ no market** |
| AB 1482 (state cap) | Capped at 5% + CPI; also maxed | **Deterministic โ no market** |
| New builds (uncapped) | Increase is a real decision | **Genuine uncertainty** |
So the uncertainty doesn't live in "will rent rise." It lives **upstream** in the CPI input that sets every capped increase, and **downstream** in the uncapped tier. We build markets on those two, not the dead middle.
## The oracle recipe (how trust is specified, not assumed)
Every market ships with a declared resolution method, visible before you trade:
1. **API-resolved** *(live on stage)* โ permits settle from DataSF `p4e4-a5a7`; CPI markets from the BLS release. One field, zero human judgment.
2. **Public-data-computed** โ rent buckets from the Rent Board Housing Inventory (`gdc7-dmcn`): authoritative, penalty-of-perjury, but annual and coarse. Method published up front.
3. **Attestation** โ storefront closures via staked local observers. Subjective, small-scale โ the frontier problem, named honestly rather than faked.
The tier *is* the honesty. We price the trustworthy-but-slow vs. fast-but-proprietary tradeoff instead of hiding it.
## ๐ The innovation: insiders attest, outsiders trade
Most prediction markets struggle because they assume every informed participant should trade. We separate incentives instead: experts trade, insiders attest, everyone else just reads the number. Some knowledge is expertise; some is influence over the outcome. A market has to tell them apart.
- **Fair (expertise):** knowing the CPI formula sets capped increases. Public rules, privately understood. Belongs in trading.
- **Foul (influence):** knowing you're about to re-list a specific unit at market. Private action you control. Belongs in attestation, never trading.
Participants who can move a resolution are flagged โ barred from trading it, routed to attest instead. Being an insider becomes a **credential**, not a weapon. This is the conflict-of-interest problem a property operator feels from the inside, turned into a mechanism. Not a slogan โ `POST /trade` really returns a 403 for a flagged insider, and `POST /attest` really records their staked claim instead. Try it on [`/agent`](https://two-doorways.vercel.app/agent), one tap runs both.
## ๐ด What we ship today
- **Create** โ a market factory: pose a neighborhood question, attach a declared oracle recipe.
- **Trade** โ a play-money automated market maker (LMSR) quotes a live, *shared* price โ every trade is an immutable record in a durable log (Vercel Blob), replayed to compute the price, so any two callers (human or agent) always see the same market. No counterparty needed. Accuracy scored in reputation, not dollars โ no regulatory exposure.
- **Resolve** โ the permit market settles live from DataSF, end to end, on stage.
- **Try it yourself** โ [`/agent`](https://two-doorways.vercel.app/agent), linked via a QR code on the homepage: one curl (or one tap, no terminal needed) places a real trade, another triggers the insider gate's real 403 โ attestation, and a live feed shows everyone's activity landing in the same market in real time.
Two live SoMa markets. One resolution you can watch happen. **Play money on purpose** โ the real-money regulatory picture (Kalshi/CFTC) is unsettled, and our contribution is the mechanism, not the wagering.
## Market potential โ two doorways, two buyers, one flywheel
**Doorway 1 โ the independent operator (signal, sooner).** Institutional owners already underwrite everything that matters to them โ CoStar, Yardi Matrix, in-house analysts, broker relationships, and increasingly AI-driven monitoring of public filings. That's roughly 3% of San Francisco's rental stock, and we don't compete there โ that edge only gets thinner as AI keeps commoditizing public-record tracking. The other ~97% is small, independent owner-operators with none of that infrastructure, running on gut feel and informal neighborhood word-of-mouth โ the exact position this project's own founder knows firsthand. AI can automate watching a database; it can't automate a contractor who knows a job is delayed or a neighbor who noticed scaffolding go up. That dispersed, undigitized knowledge is what a prediction market aggregates โ the "attestation" tier of the recipe, structurally out of reach of scraping. We're not selling faster access to the public record; we're selling the neighborhood's collective private judgment, to the owners who currently have no way to access it at all. And because this only requires participants to value being *right* โ reputation, not cash โ it already works on play money, the same reason reputation-scored forecasting platforms sell aggregated forecasts today without real-money stakes.
**Doorway 2 โ the renter (signal, viral, free).** Not a hedge โ a number. ~71% of San Francisco's rental units are under rent control, and most of the rest fall under AB 1482's state cap, so most SF renters face the same annual formula: allowable increase = a fixed share of regional CPI. We surface the CPI market's live price directly as a free, shareable stat โ "here's the crowd's current odds your rent jumps more than 2% next year" โ the same mechanic that makes a Zestimate or a Spotify Wrapped spread on its own. No payout, no money movement, no insurance question โ it's an informational estimate, not a promise. (We looked hard at a literal rent hedge here โ buy the side that pays out if your capped increase comes in high โ and killed it: paying out specifically because *your* rent went up reads as personal-loss indemnification, i.e. insurance, regulated state-by-state with actuarial reserves, a much harder bar than the event-contract lane Doorway 1 lives in. Not worth the exposure for a feature that doesn't need it.)
**Why this is one flywheel, not two businesses:** every renter who checks their number is a potential participant adding an opinion to that same CPI market, and every neighborhood that gets attention this way is a candidate for the next permit market. Consumer virality on the renter side feeds the participant density that makes the operator signal on the other side worth paying for.
## The horizon
A neighborhood is a tribe with concentrated knowledge. Give it an instrument and its wisdom becomes legible โ to itself first, and to capital on the tribe's terms. Later: autonomous resolvers with onchain identity, staked reputation, and provenance for every settlement. Instruments, not oracles-by-decree.
## Beyond housing
The recipe โ public data + local expertise + accountable attestation โ isn't specific to rent. Same mechanism, new domains:
- Wildfire and evacuation forecasts
- Supply chain disruption
- Public health signals
- Permitting and construction timelines
- Homelessness services capacity
- Infrastructure maintenance backlogs
Housing is the wedge. The platform is the point.
Prediction markets have existed for decades. We believe the missing layer isn't another market โ it's trust infrastructure that connects public data, human testimony, and AI into one accountable forecasting system.
*Places hold knowledge. This makes it tradable without making it extractable.*
---
## Repo map
```
two-doorways/
โโโ README.md โ you are here (also the pitch)
โโโ vercel.json โ headers, rewrites, outputDirectory
โโโ package.json
โโโ public/ โ deployed web root
โ โโโ index.html โ the scrollytelling explainer โ also a live tool + WebMCP surface
โ โโโ agent.html โ /agent โ QR-linked, one-tap starter trade + insider-gate demo
โ โโโ slides.html โ /slides โ the 5-slide presentation
โ โโโ screenshots/ โ same images as docs/screenshots, servable on the live site
โ โโโ vendor/qrcode*.js โ vendored qrcode-generator (Kazuhiko Arase, MIT), no CDN
โ โโโ robots.txt โ AI bot allow rules + Content Signals
โ โโโ sitemap.xml
โ โโโ .well-known/mcp.json โ MCP Server Card (SEP-1649 discovery)
โโโ api/ โ Vercel serverless functions โ the agent-curlable surface
โ โโโ markets/index.js โ GET list all markets
โ โโโ markets/[id]/index.js โ GET one market's price + recipe
โ โโโ markets/[id]/resolve.jsโ GET live resolution
โ โโโ markets/[id]/trade.js โ POST a trade (insider gate enforced, durable log)
โ โโโ activity.js โ GET the shared live trade feed (also lazily seeds empty markets)
โ โโโ starter-pack.js โ GET a one-call onboarding trade for a fresh pseudonymous trader
โ โโโ mcp.js โ real MCP server (streamable-http, stateless)
โ โโโ docs/[name].js โ markdown content negotiation for the docs below
โ โโโ _lib/
โ โโโ registry.js โ shared market registry (REST + MCP both use this)
โ โโโ tradelog.js โ event-sourced trade log on Vercel Blob โ the durable shared AMM state
โโโ src/
โ โโโ markets/ โ market factory + market definitions
โ โ โโโ factory.js โ create a market from (question, oracle recipe)
โ โ โโโ examples.js โ the two live SoMa markets (CPI + permit)
โ โโโ amm/
โ โ โโโ lmsr.js โ play-money pricing (cost fn, price, buy/sell)
โ โโโ oracles/
โ โโโ recipe.js โ recipe interface + the 3 recipe types
โ โโโ datasf-permit.js โ LIVE resolver โ reads DataSF p4e4-a5a7
โ โโโ bls-cpi.js โ LIVE resolver โ reads BLS CPI-U SF-Oakland-Hayward
โโโ scripts/
โ โโโ refresh-cache.js โ re-fetch live data into /data before stage
โ โโโ dev-server.js โ local stand-in for `vercel dev`, no login needed
โ โโโ test-mcp.js โ real MCP client test against api/mcp.js
โโโ data/ โ cached API responses for offline demo safety
โโโ docs/
โโโ ONE-PAGER.md โ the required one-page summary
โโโ DEMO-SCRIPT.md โ 3-minute live demo runbook
โโโ RESOLUTION-SPEC.md โ exact settlement rules per market
```
## Run
```bash
npm install
# Local site + API, no Vercel login needed
node scripts/dev-server.js # โ http://localhost:3000
# Or with the real Vercel dev server (needs `vercel login` first)
vercel dev
# Test the live permit resolver against DataSF (no API key needed)
node src/oracles/datasf-permit.js
# Test the live CPI resolver against BLS (no API key needed)
node src/oracles/bls-cpi.js
# Test the real MCP server end-to-end with an actual MCP client
node scripts/test-mcp.js
# Refresh the /data offline-fallback cache (run during venue setup)
node scripts/refresh-cache.js
```
Deployed, the site is dual-legible: humans get the scrollytelling page, agents
get `curl /api/markets`, a real MCP server at `/api/mcp`, and markdown
versions of every doc (`/one-pager`, `/readme`, etc. โ send `Accept:
text/markdown` or just curl them). See section 08 on the live page for the why.
`/agent` is the same idea aimed at a human holding a phone: scan the QR on
the homepage (or just visit the URL) and tap through the exact same calls
an agent would curl โ one starter trade, one insider-gate 403 โ no
terminal required.
Both resolvers fail over to the cached snapshots in `/data` automatically if
the live call errors, so a dead venue wifi doesn't kill the demo.
## Deliverables checklist (per hackathon guide)
- [x] GitHub repository
- [x] 3-minute demo (see `docs/DEMO-SCRIPT.md`)
- [x] 5-slide presentation โ [`/slides`](https://two-doorways.vercel.app/slides)
- [x] One-page project summary (`docs/ONE-PAGER.md`)
- [x] Live demo โ permit market resolves from DataSF on stage
## Scoring map (how each piece earns points)
| Criterion | Weight | Where we win it |
|-----------|--------|-----------------|
| Real-world Impact | 25% | Three-tier domain edge; free renter-signal use case; public-data settlement |
| Technical Execution | 25% | Working LMSR + live DataSF resolution end-to-end |
| Innovation | 20% | Insiders-attest-outsiders-trade; declared oracle recipes |
| Market Potential | 20% | Two-buyer flywheel: signal to the 97%-of-market independent operators CoStar ignores (sooner, reputation-only), fed by a free viral renter-facing signal (no hedge, no regulatory exposure) |
| Presentation | 10% | Scrollytelling explainer + tight deck |