How a deal becomes a pick.

We cover Bay Area regions on a weekly cadence, one board per region. What enters the pipeline is decided by rules that run the same way every week:

Growth catalysts still matter here: a funded transit line is worth telling you about. It changes zero numbers in the model. A catalyst is a flag on the page, never a rate in the math.

The wealth comparison lives in one deterministic module of our pipeline, a file named engine.cjs, with every default pinned in defaults.json. The public calculator on this site runs the same math. The horizon is ten years: buy the building, live in one unit for two years while renting the rest, then move out and run it as a full rental. Against it, the alternative: never buy, keep renting, and put the same starting cash in an index fund.

The winner is decided on hold net worth: portfolio plus property value minus mortgage debt, with reserves counted back in. A liquidation figure that subtracts selling costs prints beside it as the secondary number. Margins are signed: when the index wins, the page says so in negative dollars, and no prose argues past that verdict.

The headline number is the S&P breakeven: the annual index return at which never buying would catch the deal over the decade. We assume 10%. A deal whose breakeven sits above 10% wins at our defaults, and the higher it sits, the stronger the deal. One function turns that number into the letter grade on every record:

Grades D and F mark deals whose cautious case loses to the index. They render as losses, they can't be picks, and their pages explain why they lose. A second solved number prints on every record too: the crash floor, the appreciation rate at which the deal and the index tie. It's derived from the model, and the answer doesn't move when our appreciation convention does. That makes it the honest headline for appreciation risk: this deal needs X%/yr, whatever anyone forecasts.

Published verdicts run with the income-tax model off. Depreciation and interest deductions are opt-in, in the calculator and in your signed-in overlay; turning them on only helps the buy side, so the published record stays the cautious version.

Deliberate simplifications, on the record:

• PMI terminates at 78% of the original purchase price (the Homeowners Protection Act rule), so optimistic appreciation can't cheat PMI away early.
• Emergency reserves earn 0% and return to final wealth: real cash, never spent, never grown.
• Wages grow with inflation in both arms. No real raises assumed.
• One maintenance-and-vacancy rate applies in both phases.
• Named tax events we don't compute, listed in the calculator's disclosure block: Schedule E income tax on rental profit, passive-loss limits, depreciation recapture at sale, capital-gains tax on the index side.

So every property runs three times:

In San Francisco a stricter rule sits underneath the cautious run: rent-controlled buildings anchor to Rent Board registry figures for what comparable buildings actually collect, and the run refuses to start when a cautious rent basis doesn't conform. A deal that only wins at optimistic rents never publishes.

Each run produces its own S&P breakeven and a ten-year wealth margin in dollars. Percentages on seven-figure assets hide six-figure differences; we want the money visible as money.

Three inputs used to be judgment calls. They're now conventions: one value, set in one file, enforced mechanically. A scenario that disagrees with a convention doesn't run.

Appreciation is a threshold, never a forecast

One set of appreciation rates covers every property in every ZIP: 2.5% cautious, 4.0% base, 5.5% optimistic, defined in a single module named appa-convention.cjs. Read the cautious rate as a publishing standard: we publish deals that beat the index while needing under 2.5%/yr appreciation, in a city whose realized ten-year median has run near 4.4%.

We used to write a per-neighborhood appreciation band. We measured it and retired it. Re-derived under an unchanged rubric, the same place kept its rating about a third of the time. Nothing we built could forecast appreciation out of sample either; a flat 2.5% beat every model we tried. A number that can't survive its own re-derivation has no business pricing your decade.

Rent growth: a cap is a ceiling

Where rent control applies, modeled rent growth is the smaller of the market baseline and the jurisdiction's formula cap, derived in one module (rga-convention.cjs) from a maintained table of ordinances. San Francisco's cap computes to 1.8% and binds. A statewide cap of 8% doesn't bind against a 2% baseline, because a ceiling on rent growth is never a forecast of it.

The same regime prices the other arm. The renter you'd otherwise be lives under the same ordinance, so the cautious case grows their rent under the same cap. We fixed this in August 2026: a generic 3% on the renter's side against a capped 1.8% property side flattered buying, and flattering buying is the one direction the cautious case must never lean.

Financing: the FHA floor, and a dated rate

3.5% down everywhere. It's the FHA floor, and for 2–4-unit buildings no 3%-down program exists. The modeled mortgage rate is pinned to the Freddie Mac PMMS 30-year weekly average and stamped with its as-of date; a test in our suite goes red 35 days after the pin, so the model can't quietly keep running on last quarter's money.

Pick status is conferred by criteria, and there are three:

• It wins the cautious case at our published defaults.
• It has room for a rent shortfall: 15%.
• Its facts are verified line by line against the listing itself.

The rent cushion is the across-the-board rent haircut at which the cautious verdict flips. Read it as: rents can fall this far and the deal still beats never buying. The metric and the floor are defined in exactly one file, rent-cushion.cjs, and the gate is enforced at three separate layers of the pipeline.

Ranking is deterministic, top to bottom: highest cautious S&P breakeven first, ties broken by the signed cautious margin, then by lower price. The picks are the top of the set that passed, at most five per region. There's no override lever: rank #11 with beautiful evidence is still rank #11, and nobody can force it in.

If the cautious case loses to the index, it isn't a pick at any rank. We'd rather ship a short list than talk anyone into a $1.5M mistake.

Every input into the engine carries a recorded origin: what produced it, the exact basis text, and whether anybody actually looked. A convention is recorded as a convention. An input nobody checked can never pose as a verified one, because "we did not look" is stored as exactly that, never as "there is nothing."

Physical facts (units, beds, baths, square footage, year built) carry one of three states: corroborated, divergent, or unchecked. A divergent fact keeps both values and says which one we used. When sources disagree, independent records outrank a listing site's data fields: an assessor's roll, a rent-board filing, or the listing's own prose beats the number in a portal's summary chip, because we've caught those chips wrong on multi-unit buildings too many times.

When an analysis would rest on an unverified fact, that fact is pulled out of the lead argument and rendered in a visible needs-verification block on the page. The page may state the conditional ("if the lower flat is deliverable vacant, then X"); it never leads with it. Paid members can request verification of a specific fact, and the answer lands on the record.

Record-keeping is append-only:

• Unit-count and commercial rulings live in an append-only verdicts ledger (listing-verdicts.json), hash-guarded and never hand-edited. When an extraction rule changes, verdicts are re-derived under the new rule's version; nobody edits an old row.
• Permanent bans are a separate store, and a ban requires the listing's own text stating the disqualifier verbatim: a timeshare interest, a deed restriction barring tenants. Suspicion routes to a verification request, never to a ban.
• Listing identity (which MLS numbers and address spellings are the same building) lives in a registry that only grows: a wrong entry is corrected by adding a new row, never by rewriting one.
• Predictions: the pipeline appends a row for each scenario it runs, inputs included, to an append-only archive. Our past claims sit on disk in the form we made them.

The public face of that discipline is the track record: every pick logged when it publishes, then graded against the numbers we claimed that week. Wrong calls stay visible.

Modeling isn't the safety layer. The safety layer is a stack of deterministic checks that fail closed, and then people.

The weekly product is the regional board plus a short notification email when it's live. The weekly send itself is retired; past issues stay readable in the archive. The board shows the entire run: picks on top, then every other property we underwrote that week, winners below the cut, losers, and rejections, each with its signed margin and a one-line reason.

Every evaluated property, whatever its verdict, gets a permanent property page carrying the full underwriting at our published defaults, with the index comparison as the reference line. Publication is a record; endorsement is a separate, smaller thing. The picks are the commitment layer, the accountable calls the track record grades. The rest is the database layer, published so you can check us. A rejection's page explains why it lost, with the number that decided it. A surface that showed only winners would be a rendering bug here, and we'd treat it as one.

The board prints its own screening funnel at the top of each region's page: how many listings were screened, how many beat the index at our defaults, how many were rejected. Those are that run's real numbers, computed by the pipeline. The rejection rate is the point.

The rejection list is the product. Rejecting a winner we can't verify buys more trust than shipping it, and a thin week published honestly beats a fat week padded quietly.

What's free and what's paid

The top four picks on each live board are members-only while they hold their rank; a page goes public once its property drops out of the top four. Pick five is free in full every week. The rest of the board, every rejection and its reason, the full history and every property page below the top four: public. Locked rows still show rank, ZIP and the signed margin, the same shape the track record uses.

Run it at your numbers

Every record takes your rent: type what you pay and the verdict re-runs at that number, badged as yours. Members set their tax bracket, horizon and down payment once, and every record they open carries a toggle between the published defaults and their numbers. The published defaults line never moves, because that line is the record. One discipline holds for everyone: cautious rents compute every verdict class, including your personalized one.

For an address we haven't covered, the lookup gives you the parcel's own facts free, and runs the same engine and the same gates on demand for $79 — one San Francisco address, delivered within one business day.