Instrumenting Housing Need

The first commit is dated 2026-02-20. By 2026-09-15 the repository held 6,654 commits — sustained across seven months. The shape of that growth is not the smooth ramp of a planned build.

What the curve conceals is a change in kind. Early months added surface — pages, charts, fetchers. Later months increasingly added constraint: tests, guards, and documented rules about what the system may claim.

“Continuously rebuilt” is a claim worth making specific. 59 of the repository’s 71 workflows carry a live cron, on cadences from every thirty minutes to annually. They divide into three kinds, and the distinction matters: one kind brings in new facts, one computes over them, and one watches the other two.

Everything derived is generated, never hand-edited — the jurisdiction digests, the 484 place pages, the ranking index and its scenarios, the homepage snapshot, and the file manifests. A change to any of them is made by changing its generator and re-running it; CI fails a pull request whose generated files drift from what their sources would produce. That is the mechanism by which a one-person project stays internally consistent across 1,653 data files.

The third category — the watchers — exists because the first two run unattended. It is also, as §08 records, the least finished part of the system.

Housing policy is made at municipal scale — inclusionary ordinances, zoning, fee waivers, land contribution — while most published housing data is county-level. A county figure is almost never the figure a practitioner needs. Boulder County’s severe renter burden is one number; Louisville and Boulder city sit far apart inside it, and the spread is the product.

The system resolves 482 Colorado places by apportioning tract-level CHAS data through tract-to-place membership, then capping the result against the ACS occupied-household count. Every metric carries a geography_level, so a figure that is really the county’s own says so rather than masquerading as local.

The rule exists because the same defect recurred five times: a zero-household AMI band; $0 and −$320,000 equity figures; an absent Proposition 123 commitment rendered as non-participation; a flat chart line standing in for a dead data series; and $0 resale price and owner equity shown for 53 of 482 places.

The repository distinguishes green from verified: a passing check list answers “did the jobs report success,” not “did the work happen.” The documented tests for a real pass are whether the check actually ran, whether the passing test asserted anything, whether a cancelled outcome is being misread as neutral, and whether main is green at that moment.

Palisade, Mesa County (GEOID 0856970): 1,297 households, 433 renter and 864 owner. The town filed its Proposition 123 commitment on 2024-08-29, so it is eligible for that funding stream.

A strategy beginning with new construction reaches its first household in three to five years. A strategy sequenced by the data reaches 113 households within a year, 96 more within eighteen months, and reserves the slow instrument for the one band that requires it. Same total need, different order, materially different time-to-relief.

Of 482 places, 231 have at least 400 households and confident data. Within those, 63 meet a tractability test — acute burden concentration combined with an absolute target a single program could plausibly close. These are the jurisdictions where “eliminate the need” is a coherent sentence rather than a slogan.

The renter/owner column is the operationally important one, and it is the column county-level analysis cannot produce. Two towns can share a burden share to within a tenth of a point and have opposite tenure profiles. They need different programs. A regional analysis would assign them the same one.

These are choices about what the system is for. They are not derived from the data; they shape what the data is permitted to say.

The audience — municipal staff in towns of 500 to 5,000 households — is precisely the audience that cannot commission a consultant study or buy a data subscription. The marginal town in Table 3 does not have a housing department. A paywall would exclude exactly the jurisdictions where the tractability screen says the need is most closeable.

This is a policy choice before it is a technical one. Regional analysis is cheaper, more statistically stable, and matches how most housing data is published. It is also unusable for the instruments that actually get built: an inclusionary ordinance, a fee waiver, a land donation, and a deed restriction are all municipal acts. Choosing the place as the unit accepts wider error bars in exchange for matching the scale at which decisions happen.

A system may handle a missing value three ways: omit the row, substitute a default, or publish the absence. Omission makes a place look uncovered; substitution is the coerced-zero failure; publishing absence is noisier and less impressive and is the only one that leaves the reader correctly informed. Four sources are currently displayed as stale on the dashboard rather than having their age limits quietly raised to silence them.

Most assessments report a single “units needed” figure. This system reports cost burden and unit gap separately because they license different actions — burden licenses preservation and subsidy, gap licenses production. The Finding 2 result is only derivable because the two were not merged.

Where the data supports both, the sequencing in §03.1 places preservation first. The justification is time-to-relief rather than cost: a household severely burdened today is at risk this year, and the fastest unit to secure is one that already exists and is already occupied by the household that needs it. This ordering is a judgment, and a jurisdiction prioritising long-run stock growth over near-term stability could defensibly reverse it.

Commercial parcel data was evaluated and deferred: the cost could not be justified at current use. The consequence is recorded rather than hidden — parcel coverage is marked inactive, and the integration is left wired so it resumes if the decision changes. §07 revisits this as the clearest case where modest money would buy a capability the system cannot otherwise have.

Every decision below had a simpler alternative, rejected for a stated reason.

HUD CHAS — the only public source crossing cost burden with AMI band and tenure — publishes at tract level, and municipal boundaries do not follow tracts. The rejected alternative assigns each place its primary county’s rates; the repository documents why it fails: a parcel in Erie would see Weld County rates though half of Erie’s tracts sit in Boulder County. The error is systematic and largest exactly where towns straddle county lines.

The method: for each tract T overlapping place P the weight is max(area_share, population_share), and each CHAS count is multiplied by that weight and summed. (Area share alone collapsed small towns embedded in large rural tracts; the file header still documents the superseded rule.) The assumption is uniform household distribution within a tract — poor for slivers under roughly 10% of tract area, which are flagged rather than silently included.

Area weighting is geometrically sound and demographically optimistic: summed across overlapping tracts it overcounts households in roughly 31% of Colorado places, because population clusters in the built-up portion rather than spreading across tract area. The correction anchors each place’s total to the ACS occupied-unit count and scales the tiers to fit.

The property that matters: rates are preserved, only levels change. A cost-burden percentage — the number a policy is written against — survives the correction, while the household count becomes defensible against an independent source. The ordering is load-bearing: summary caches must regenerate before the CHAS build, or the cap silently no-ops.

The conventional price = income × 3 heuristic is rejected because it ignores interest rates — where rate movement alone shifts affordability 20–40% — ignores property taxes and insurance, worth roughly $500–$1,500 per month in Colorado, ignores mortgage insurance, and substitutes a single multiple for the debt-to-income threshold lenders underwrite to.

The model is not one model. The registry defines 7 of them — conservative screening, first-time buyer, conventional lender DTI, FHA, USDA, a Proposition 123 eligibility gate, and a fully custom one — differing in ratio convention, down payment, mortgage insurance and its loan-to-value gate. The default is conservative_screening, which screens at a 30% front-end ratio. A 43% back-end DTI is one of the others, and it answers a different question: what a lender will approve, rather than what a household can carry without being cost burdened. On the documented Denver illustration the heuristic and a PITI model agree on the affordable price and disagree entirely on the gap — roughly 117% over affordable against roughly 206% — so a town using the heuristic would size its subsidy at about half what a buyer needs. Which PITI model it uses changes the answer again. The full specification, with every parameter read out of the registry, is in the computational methods companion.

An earlier version of this section described a single model at “20% down, 6.5%, 0.65% tax, 0.85% insurance, 43% back-end DTI.” That is a recognisable description of conventional_dti — one of the seven, and not the default. Writing out the methods companion is what found it.

The numeric core of the epistemology is one line of language semantics: Number(null) === 0, and 0 is finite. A coercion two layers upstream therefore produces a value passing every downstream Number.isFinite() check. Three consequences are enforced: do not coerce a possibly-null value that feeds arithmetic or a renderer; guard <= 0 rather than < 0 wherever zero is not meaningful, since a price or rent of exactly zero means unknown rather than free; and carry the reason with the data. Fixing the renderer is insufficient — a renderer that formats null correctly never sees the null if the coercion happened upstream.

Rates were originally computed from unrounded accumulators while counts were rounded separately for display. Several shares land on exact .00005 ties, where a one-ulp difference in the denominator flips the fourth decimal — producing files that differed between machines and a freshness checker reporting five places as stale whose inputs had not changed in months. Deriving each share from the published counts makes the computation reproducible and, as a side effect, internally consistent: a reader dividing the published count by the published total now obtains the published share, which was previously not true.

The system is labour-rich and capital-poor. Its limits are not analytical — they are four specific missing inputs, each of which costs less than a single consultant housing needs assessment.

The National Housing Preservation Database extract has had no refresher since it was loaded — it is stale and displayed as such. This matters more than its size suggests: preservation is the fastest track in every sequencing above, and expiring affordability restrictions are time-bound in a way that need estimates are not. A unit lost to expiry costs far more to replace than to preserve. This is engineering time, not subscription cost.

Commercial parcel data was deferred on cost. The capability it would unlock is specific: moving from “this town needs deep-tier units” to “here are the parcels where they could go, with their zoning, ownership and size.” Need analysis identifies the target; parcel data is what converts a target into a site. This is a genuine annual subscription cost and the clearest case where modest money buys a capability that cannot be substituted with effort.

59 scheduled workflows now run with no map of themselves, alerting that covers failures but not silence, and at least one quality gate that runs only after merge. On 2026-09-14 the public site stopped deploying for thirty-two hours and three separate monitoring mechanisms each failed to report it. None of this is hard; it is unglamorous consolidation work that no volunteer evening prioritises over a new feature, and it is currently the largest tax on the maintainer’s attention.

A funder deciding whether to support an existing system is really asking what it would cost to commission one. The scope is measurable: 338,104 hand-written lines — excluding 226,824 generated lines of place pages — 62 pages of which 25 are substantial interactive tools, and 63 declared data sources feeding 59 scheduled pipelines.

Decomposed by workstream and summed, that comes to 6.1–11.1 person-years, or 12–21 months of calendar time for a team of 7 after allowing 12% for coordination and 25% for the fact that a team of seven is not seven times one person.

A paper describing a system built around admitting what it does not know owes the same treatment to itself.

This repository was built with substantial AI assistance over roughly nine months. That is a material fact about how the work was produced, it carries an energy cost, and a paper arguing that systems should publish what they would rather omit cannot omit it.

Two things are measurable. Continuous integration has run 40,000 workflow jobs, about 1,509 runner-hours of hosted compute. Model inference across the surviving session transcripts comes to 71,384 turns generating 68,149,998 output tokens.

Turning turns into kilowatt-hours is where the accounting stops being measurement. The published per-query energy figures span 140×, from 0.24 Wh for a short prompt on optimised serving to 33.6 Wh for a reasoning model on a long one. Agentic coding turns carry longer context than the high-end benchmark, so neither end is a bound so much as a signpost.

What is not claimed here is a net benefit. It would be easy and self-serving to set this against the travel and staff time of the consultant studies the system substitutes for, and declare the trade favourable. That comparison has not been measured, the counterfactual is unknown, and asserting it would be exactly the unsupported claim this paper spends nine sections arguing against. The cost is stated because it was incurred, not because it has been justified.

The narrower point stands on its own: scripts/paper/measure-compute-footprint.mjs computes these figures the same way every other figure here is computed, and they update with the rest. A project that discloses its compute cost once, in a launch post, has disclosed a historical curiosity. One that recomputes it has disclosed a number.

1. Read the tenure split before anything else. It determines which half of the toolkit applies, and it is the finding most often lost in county-level work.
2. Separate the shortage from the mismatch. Compare unit gap by band against burden counts by band. Where the gap is zero and burden is high, the units exist — acquisition and income targeting beat construction on both cost and time.
3. Find the severe-burden concentration. Households above 50% of income are at risk this year, and preserving an occupied unit is consistently faster than producing a new one.
4. Check the funding gate before designing the program. Proposition 123 eligibility turns on a filed commitment; 217 jurisdictions’ filings are tracked, so this is a lookup rather than a phone call.
5. Treat any “0” with suspicion and look for a stated reason. Where the system says “not published,” that is information about the source, not about the place.

Every figure in this paper is written by scripts/paper/build-paper-figures.mjs and injected by scripts/paper/inject-paper-figures.mjs. None is typed. Re-running them refreshes the paper against whatever the repository then says, and test/paper-figures-fresh.test.js fails continuous integration if this page and the data ever disagree — so the paper cannot silently go stale while the site beneath it moves.

A figure the generator cannot read is emitted as null with a stated reason and rendered here as that reason, never as 0 — the same rule the system applies to itself. That generator is what caught both corrections recorded in §01.

Authorship and standing. COHO Analytics is the sole product of pggLLC, built and maintained by one person. It has no institutional funding, no research grant, and no staff. That is relevant to reading this paper in two directions: it explains the limits catalogued in §08, and it is the reason the investment argument in §07 is framed around four specific inputs rather than a programme — the labour is already being supplied.

Figures reflect the repository at commit 0e3579672. The underlying data refreshes continuously; a reader reproducing these tables later should expect vintage drift and is encouraged to check each metric’s own as_of field rather than this page’s date.