listings + housing: the OpenListing directory, and Redfin's open market data

[ralyodio]

Two adapters. One reads the spec published on logicsrc today; the other pulls the closest thing to open MLS data that exists.

openlisting — the reference directory

A listing is believed only when served from the origin its id claims — the rule OpenWebring established here. Otherwise anyone could serve a document claiming to be someone else's listing and this would carry it. A seller's index may only point at the seller's own listings, checked the same way.

Three decisions follow the spec rather than convenience:

• Prices stay decimal strings end to end. 2450.10 isn't representable in binary floating point, and a listing is a price. A test asserts the built item never turns it into a number.
• A street is dropped unless the seller said precision: exact or street — so a landlord publishing a postal code for an occupied home stays published at a postal code.
• bedrooms: 0 survives, because a studio is not an unstated bedroom count. The one documented place absent and zero differ.

Closed listings are ingested and tagged, not discarded. The spec asks sellers to keep a sold listing served 90 days with closed_at, because a directory learning from a 404 can't tell "sold" from "the server is down" — and the outcome is the part every listing site loses. There's a feed for exactly that.

It ships no default source. The spec published today and nobody serves a descriptor yet; a seeded source pointing at an invented origin would fail every run and read as the adapter being broken.

redfin-market — what the market actually did

The open S3 bucket Redfin publishes: median sale/list price, price per sq ft, homes sold, pending sales, new listings, inventory, months of supply, days on market, sale-to-list — monthly since 2012, by property type.

Six levels, measured 2026-09-25:

level	size	default
neighborhood	2.24 GB	opt-in
zip	1.48 GB	opt-in
city	955 MB	opt-in
county	230 MB	✅
state	8.6 MB	✅
national	0.5 MB	✅

Redfin is a brokerage with MLS access giving away the derived view — and that view is the half a person actually asks for. The listings themselves are not open and won't become open: 484 separate MLSs, each needing a licence, a broker and a signed agreement, none redistributable.

It is not a listing, and every row says so in data.note. OpenListing forbids publishing aggregate market data as a listing — no seller, no offer — so this is housing-market in its own collection, deliberately apart from /c/listings.

Two collisions caught while wiring up

• housing already exists, with six adapters (Land Registry sold prices, Eurostat HPI, permits, Freddie Mac rates). Redfin joins it and the existing description is extended, rather than creating a second collection with the same slug.
• The registry test's hardcoded collection list needed listings — the same check that caught parts earlier today.

Failing loudly

Both header maps throw rather than degrade: a Redfin file missing a column it reads fails instead of producing rows where every measure is absent. And an empty TSV cell is null, never 0 — a ZIP with no sales isn't a ZIP where the median price was zero, and charting it as zero puts a spike through every quiet market.

Checks

59 new tests (built on a real header and row from state_market_tracker), 2,101 across the suite, lint clean. The usual 5 Redis/Postgres integration failures this dev box can't run.

🤖 Generated with Claude Code