Protocol & scope
United States; the cost, supply, and stability of housing — rental and ownership, market-rate and subsidized, with homelessness as the system's failure boundary. Imports the gubment method (M1–M8) in full: two-source rule, root-tracing, pre-registered hypotheses and adjudication criteria, anchor quarantine, and a deviations log that stays visible.
Executed across Phase 0 and full-execution passes on 2026-08-03, with two late workstreams added 2026-08-06: baseline with committed microdata pipelines (ACS, CPS-HVS, Building Permits Survey, AHS), supply and land use with pre-registered adjudication and its own admin-data instruments, construction economics with a committed labor-requirement model, finance and subsidy plumbing in two passes, federalism and legal mechanics, market actors, civil society, precedents (closing the Faircloth anchor at primary source), an architecture scorecard, and sequencing. A red-team pass corrected three scorecard cells and withdrew the draft headline before publication; an instruction-blinded second scorer then corrected thirteen more and replaced the headline again (2026-08-06). A Phase 1 verification pass against primary sources followed on 2026-08-10 — see the verification log — and a Phase 2 steelman the same day, in a separate session, built the strongest case against the filing’s own conclusion and partly won: see the steelman log. Its finding in one line: the diagnostics hold, the prescription does not travel as published, and the board’s missing rows — no tenant-protection instrument, no preservation instrument — matter more than any cell on it.
Anchor table — priors, stated before evidence, verified after
Every anchor was written down as an unverified guess before research began, so it could be broken. Three broke; one closed at primary source after two sessions of industry-restated figures.
| # | Anchor (unverified prior) | Verified value & delta |
|---|---|---|
| 1 ★ | Shortage estimates cluster ~3–7M units | Broken, and the bigger finding fired the kill condition: published estimates span ~1.0–8.2M and the number is a near-linear restatement of each publisher's assumed target vacancy — 1pp ≈ 1.59M units; Freddie Mac's 3.7M reproduces to within 3% from its stated target alone. |
| 2 ★ | ~Half of renter households are cost-burdened | Confirmed by direct ACS tabulation: 22.24M burdened renter households (48.2–51.8% depending on denominator); the circulating 50.3% reproduced exactly. The single-root caveat is closed: AHS 2023 microdata gives an independent instrument at 53.6% against ACS’s 51.8%. |
| 3 ★ | ~1 in 4 eligible households receives federal rental assistance | Directionally confirmed (~5.3M assisted; 3 in 4 eligible receive nothing), with a definitional correction: "eligible" means low-income and burdened/inadequately housed. |
| 4 | Ownership tax expenditures vs. rental assistance: multiple × | INDETERMINATE on strict two-source grounds; directionally supported on the narrow scope by ~35% (JCT FY2026 vintage) — too large for rounding, not clean enough to stamp. |
| 5 ★ | Tokyo out-built all of California | Confirmed and stronger per capita (~3.2×). The "disposable houses" caveat was retired on root-tracing: cohort turnover 65.0 vs 66.6 years — comparable stocks, larger advantage. |
| 6 | Auckland upzoning: measurable rent moderation | Downgraded from settled to contested: single author lineage, non-arm's-length rebuttal, three confounders inside the window. The institutional claim survives; the identification doesn't. |
| 7 | Construction productivity flat-to-declining for decades | Complicated by BLS's own revised deflators showing positive single-family productivity growth (secondary-relayed; primary PDF 403'd — logged). |
| 8 | Institutional investors own low-single-digit % of SF rentals | Supported as a surface, not a point: 1.9–5% depending entirely on ownership threshold; the "quarter of homes" figure is a purchase-flow statistic dominated by small owners. |
| 9 | Faircloth caps public housing at 1999 levels | Closed at primary source: not a cap, and not a freeze either — 42 U.S.C. §1437g(g)(3) conditions the use of Capital and Operating Fund money on not netting above the agency’s 1 October 1999 count, and subparagraph (B) permits building above that line without extra formula funding. HUD’s own list (04-03-25) sums to 278,345 units of headroom, computed twice independently, committed pipeline, and a third time on verification; one internal inconsistency in HUD’s file caught and disclosed. |
| 10 ★ | LIHTC per-unit cost approaches/exceeds market | Broken: the comparison has never been made nationally (Congress has assigned no federal agency to maintain and analyze the cost data) and four state like-for-like studies found no significant difference. |
Workstream findings
§2 · BaselineThe shortage is a sensitivity surface; burden and rationing are confirmed
Committed pipelines for ACS, CPS-HVS, and permits; the shortage back-test reproducing publishers' numbers from their vacancy assumptions; H2.1 (ownership vs. rental subsidy) indeterminate on strict grounds, narrow-scope supported by ~35%.
§3 · Supply & land useLegalizing a use is not legalizing a building
Envelope-adding prescriptive reforms produced units (CA ADUs, Oregon); unit-count reforms holding envelope produced almost none (SB 9, Austin HOME — decomposed from the city's own permit database). Five passes, pre-registered criteria, two disclosed test-design errors caught and corrected, one materially ambiguous criterion disclosed rather than buried.
§4 · Construction economicsHard costs aren't the constrained-metro lever; the labor exposure is near-term
Committed labor-requirement model: closing any defensible shortage band needs workforce growth 2.6–63× trend (3.8–31× on a ten-year ramp). Foreign-born share at record highs; undocumented share 15–23% (a band across three organizations measuring three different populations); no aggregate construction wage premium — the textbook shortage signature is absent.
§5 · Finance & subsidy plumbingVouchers leak where mobility points; insurance eats the operating side
Landlord denial 76–78% in the two study markets with no source-of-income protection (67% in a third with partial protection), worst in low-poverty areas; LIHTC equity ~83¢ with the par-gap undecomposable from public data; multifamily insurance premiums roughly doubled 2021→2024, structurally unabsorbable by rent-restricted owners.
§6 · Federalism & legal mechanicsCongress's direct reach into zoning is a paragraph; the state is the layer
The only federal statute on local land use this pass could locate requires describing barriers, not removing them — and the certification was defined down to "any action rationally related." Findings-clause drafting survives state-court challenge; preemption carries a procedural tax §3 first put at ~1 year and §6 revised to roughly twice that — Montana ran 27 months to a unanimous merits win.
§7 · Market actorsH7.1’s mechanism refuted at moderate confidence; the investor share is a threshold-dependent surface
Homeowners oppose upzoning that raises their own property values — asset-defense isn't the operating motive, which changes what instruments could work (venue/process, not compensation). Institutional SFR share resolves to 1.9–5% depending entirely on the ownership threshold.
§8 · Civil society & coalitionsThe swing vote is labor; the killer is home rule; four formulas, none replicated
California's 2018→2025 arc documents the Building Trades flipping from bill-killer to neutral once enforceable standards were attached — while the tenant camp stayed opposed and bills passed anyway. NY's and CO's preemption bills died to suburban home-rule revolts, not the tenant/supply split.
§9 · PrecedentsFaircloth closed at primary source; every marquee model's engine is non-portable
278,345 units of verified headroom (committed pipeline; HUD's own file has one internal inconsistency, caught). Houston's causal literature is one institution deep; the postwar "2M/year" is a peak-year artifact; Vienna's land bank and Singapore's acquisition powers don't travel; Sweden's queue is one distribution, not competing estimates.
§10 · ScorecardOne evidenced leader, and a lead one cell wide
Nine architectures, four objectives, six weightings including the neutral one the draft never ran. After the 2026-08-06 re-score, envelope-prescriptive preemption leads all six weightings; the Faircloth channel’s former stability lead rested on an unevidenced cell and is withdrawn; the land-value tax’s nominal second place is an ignorance artifact, not a finding. Housing-First carried unscored. Phase 2 steelman, 2026-08-10: 23 of the board’s 32 cells are neutral and the leader holds the only cell above neutral anywhere on it, so “leads all six” follows from the matrix’s shape and is worth w_supply — 0.25 of a point under the neutral weighting. That cell is the one the two adversarial checks split on — red team 3, blind scorer 5, reconciled to 4 — and read at the red team’s 3 it leads none of the 1,771 weightings swept, tying the land-value tax. Cells 4C and 4D are contested against two randomized trials absent from this record. Arithmetic now committed and reproducible.
§11 · SequencingThe seed order survives with amendments — except “demand-side last”
Prescriptive legal capacity first (four documented political formulas, matched to each state's veto players); the operating-side insurance constraint enters the finance tier as a first-class open question; the construction-workforce clock runs through everything; the homelessness track unexecuted at publication and picked up late in §12. “Demand-side last” downgraded 2026-08-10: §13 removed its inflation rationale and the ordering was left standing anyway; the Phase 2 steelman then found the two randomized evaluations of the voucher missing from the record entirely. The preconditions (source-of-income enforcement, receiving-market capacity, mobility services) stand; the queue position does not.
§12 · Homelessness (late pass)Housing First works on housing, and does not pay for itself
The parallel track the filing left unexecuted, run 2026-08-06 against RCT syntheses, the At Home/Chez Soi randomized economic evaluation, HUD’s 2024 AHAR and GAO-26-107517. Housing stability supported (~2.5× stably housed at 18–24 months) for a narrow high-needs adult population; self-financing refuted (net cost CAD $7,868 per participant per year after a 46% service offset); scaling constrained by units, subsidy and staffing, not by the model label. The scorecard row stays UNSCORED pending a fresh matrix pass.
§13 · Demand-side rent effectsThe inflation premise the scorecard inherited does not hold as a general finding
A late prospective check (2026-08-06) on H10.1’s imported claim that demand-side assistance without legal capacity inflates rents. Broad identified tests (Eriksen & Ross; HUD’s Housing Assistance Supply Experiment) find null-to-small overall effects; increases concentrate near voucher ceilings in supply-inelastic markets. Voucher and cash cells stay neutral; the sequencing rationale is access and localized risk, not general inflation.
Deviations log
Seventy-three entries — the fullest deviations record of any filing to date, spanning collision-renumbering across parallel PRs, two test-design errors caught before verdicts, a fabricated aggregator claim rejected by root-tracing, access failures logged rather than smoothed over, and the red-team corrections. The full table lives in the repository; highlights:
| # | Deviation | Effect |
|---|---|---|
| 15, 18 | Two §3 test-design errors (a no-ties ordinal test that structurally couldn't pass; a within-tier inversion count) | Both caught by hand-checking test logic before any verdict was recorded — both would have produced falsely negative results |
| 23 | §3's H3.1d criteria ambiguous between two readings that give opposite verdicts | Resolved toward the band reading on outcome-independent grounds, with the post-hoc nature disclosed — a reader applying the strict reading reaches REFUTED |
| 60 | An aggregator site claimed Katerra "remains a significant player" three years after its bankruptcy | Rejected via root-tracing — a positive control that the two-source discipline works |
| 63–64 | §8/§9 relaunched from scratch after the prior session's work didn't survive; Faircloth closed at primary source | Load-bearing facts independently re-verified by the primary session; anchor 9's gap retired |
| 65 | Red team corrected three scorecard cells and withdrew the draft headline | The same failure family earlier filings' checks caught recurred in new guise — evidence the check must stay structural, not remembered |
Red team
Seven attacks plus a master critique against the §10/§11 draft; six absorbed with corrections, one endorsed the draft's handling (Housing-First unscored), none dismissed without documented reasons.
Master critique — the "adversarial" weighting reweighted axes but never re-scored cells
Absorbed. Cells re-scored under a symmetric discipline; the neutral weighting added — it ties the top three under the strictest reading.
The top pair was scored as idealized composites; rivals as-implemented
Absorbed — the structural fix. The labor formula moved to the feasibility gate where its evidence lives; Houston relabeled as mechanism evidence; Vienna demoted from scored capacity to design guidance; one cell corrected for violating the scorecard's own evidence-floor rule.
The sequencing overreached twice
Absorbed. The insurance instrument's priority claim downgraded to an open question (no cost comparison exists in the record); California's labor formula demoted from "the" prescription to one of four documented state formulas.
Cash comparator — strawmanned?
No change needed. The evidence-floor rule held cash's score down correctly; the same rule now binds uniformly everywhere else too.