57.6% of Our Development Pipeline Scores Are the Number 33. That Number Means the Permit Feed Is Empty.

A correction, and the rent data we should have been checking against all along.

The Setup

Locus scores 5,002 hexagonal cells across 22 US metros. As of today, 2,881 of them — 57.6% — carry a development pipeline score of exactly 33.

Not "about 33." Exactly 33, in every one of them.

Eleven metros — Atlanta, Charlotte, Dallas, DC, Houston, Miami, New York, Phoenix, Portland, San Diego, and Las Vegas — score 33 in every single cell, with a standard deviation of 0.00. A score that does not vary across 1,546 cells spanning eleven different housing markets is not measuring those markets.

It is measuring whether we have a permit feed.

The Chain

The development pipeline score combines seven signals: permit activity, construction detection from Sentinel-2, NLCD land-cover change, federal Opportunity Zone status, permit velocity, permit volatility, and an LLM-extracted permit scope score.

Here is what the score object actually looks like for a Washington DC cell, straight out of production:

score: 33 confidence: 0.0375 subScores: [ { "Permit Velocity": 33, weight: 0.04 } ] sourcesUsed: [ "Building Permits Trend" ] staleSources: [ "building_permits" ] sourcesMissing: [ "Building Permits", "Sentinel-2", "NLCD", "Federal OZ", "Building Permits Volatility", "AXL-5", "AXL-108" ]

One signal out of seven. Six missing. One stale. Confidence: 3.75%.

And the 33 itself is not a market level — it is an artifact of arithmetic. Permit velocity is normalised onto a scale running from −10% to +20%. A metro with no permits has a velocity of zero. Zero on a −10-to-20 scale lands at 10/30, which is 33.33.

Zero permits → zero velocity → 33. That is the whole mechanism.

The score has been telling us it was 3.75% confident this entire time. We published the 33 and never published the 0.0375.

Why the feed is empty

Across 22 metros there are 872,878 building permits carrying an issue date. Boston holds 654,008 of them — 74.9%, going back to 2006.

The rest of the picture:

This invalidates a specific kind of claim we made repeatedly: cross-metro pipeline rankings.

"Boston's development pipeline score is 92.7% higher than any other metro's" is a true statement about Boston's twenty-year permit archive and everyone else's few months. It is not a statement about construction. Boston has 225× as many dated permits as Los Angeles because of what we ingested, not because of what was built.

Two more corrections in the same family:

• Austin's "10,749 permits in 90 days, +89.5% month-over-month" was a ramp-up artifact. The Austin feed opened on 2025-01-02. The surge was measured against a baseline that did not exist.
• Dallas has appeared in comparison tables since 2026-05. Its permit feed stopped on 2020-05-29. Every Dallas pipeline figure we published was computed against six-year-old silence.

The regime break nobody announced

On 2026-07-12 the scored corpus went from 2,000 cells to 5,002. On the same day the median development pipeline score moved from 6.0 to 20.2 to 31.9, business vitality from 26.6 to 35.6, and amenity demand fell from 51.8 to 20.7, where it has sat unchanged for seventeen consecutive days.

Nothing was built on 12 July. The scoring regime changed — including the no-data sentinel, which moved from 4 to 33. That is why our older posts describe a "pipeline floor of 4" and the newer ones describe a floor of 33. Both were describing the same emptiness under different arithmetic.

There is a second trap in that history. Before 2026-07-12, the daily snapshot sampled only 2,000 of 5,002 cells, and the number of metros in each snapshot oscillated between 22, 18, 11, and 2 from day to day. Daily averages in that period move with which metros got sampled. Any trend line drawn across that window — including across the 12 July break — is measuring our sampler.

What the market actually did

Here is the part we should have run first. Zillow's rent index has been sitting in our database the whole time: 421,167 rows, monthly, by ZIP, back to January 2015. Until this week it had never been joined to a single cell score.

It is joined now. Across the ZIPs containing our scored cells:

The eleven metros whose pipeline score reports them as identical moved between +0.3% and +13.8% on rent from 2023 to 2026. New York gained 13.8%. Miami gained 6.3%. Phoenix gained 0.3%. Those are three completely different markets sharing one score of 33.

And the longer series reorders the story we have been telling. Ranking metros by rent growth in the 2019–2023 boom against their growth since:

The rank correlation between boom-era growth and growth since is ρ = −0.387 across 22 metros (Pearson r = −0.356). That is moderate, not deterministic — Miami is high in both periods, and the relationship is carried substantially by the extremes. But the direction is consistent and it runs against the Sun Belt narrative that our permit-composition posts implied.

San Francisco is the sharpest reversal in the set: worst in the boom, best since, and currently running +10.9% year over year — the strongest in our footprint. We characterised SF as a stalled repair economy on the strength of a renovation-to-new-construction permit ratio. Its rent series disagrees.

The Implication

A score built from one signal at 3.75% confidence should not render as a number that looks like every other number. It should render as absent. A reader cannot distinguish "this market is mid-tier" from "we have no data here" when both print as 33, and the burden of that distinction should never have been on the reader.

Three changes follow from this:

1. 1.Coverage-gated publishing. Any signal with zero cross-cell variance in a metro is now blocked from being written as a market finding. It can be written as a coverage story, which is honest and useful, or not at all.
2. 2.Confidence travels with the score. Any group score quoted in an analysis carries its confidence and its populated-signal count in the same paragraph.
3. 3.Outcome-anchored claims. Every forward-looking claim gets tested against the rent and price panels before publication, not asserted from the signal alone.

What to Watch

• Permit feed coverage for the ten zero-permit metros. The pipeline score for half our corpus is uninformative until those feeds land. Coverage is the metric to watch, not the score.
• Whether amenity demand unfreezes from 20.7. Seventeen days of a constant across 5,002 cells is a stalled input, not a stable market.
• Whether SF's reversal holds through 2026. A single metro flipping from worst to best in one cycle is the observation most likely to be noise.

Limitations

The rent panel maps each cell to its nearest ZIP centroid; all 5,002 cells resolved within the 15 km tolerance, but a centroid is not a boundary and dense urban cells can sit near a neighbouring ZIP. These figures are rent across the ZIPs containing our scored cells, which is a narrower footprint than the metro-wide numbers Zillow publishes — they will not match a headline CBSA figure, and should not be quoted as one.

The lead-lag result is a metro-level correlation across 22 observations. It is not a causal estimate, it does not control for unit mix, interest rates, or migration, and n = 22 is small. A negative rank correlation between consecutive-period growth is also partly what mean reversion looks like in any price series. The honest claim is directional, not predictive.

Permit counts reflect what we have ingested, not what municipalities issued. Every coverage statement above is a statement about our pipeline.

Data current as of 2026-07-29. Sources: municipal building permit open-data portals; Zillow Observed Rent Index (ZORI), January 2015 – February 2026; FHFA House Price Index, 2016 – 2026Q1. This post supersedes and corrects fourteen earlier posts that reported uniform cell scores as market findings.