← back to the tool · all reviewer documents

This is a working document from the project, published verbatim for reviewers. File paths in it refer to the project workspace.

RESIDUALS: what the replication could not derive, and why

Written 2026-08-11 when the recovery gate went green (zero exception cells; every reproduced line within 0.5%, most within 0.01%). Companion to SPEC.md and METHOD.md. Plain language.

The score

All 8 prototype tables (4 housing types x 2 scenarios x 21 years x 13 lines) reproduce from: published rates and values (derived lines), plus four recovered internal series per table (units built, average size, students, population proxy), plus a handful of constants. Each recovered series is validated by at least one independent published line it was NOT taken from:

units series: taken from the property-tax line; confirmed by fire levy, income tax (0.14%), surcharge and excise.
size series: taken from the surcharge line; confirms road excise to 0.0003%.
students series: taken from school debt service; confirms operating schools to 0.0003%.
population series: taken from all-other revenues; confirms all-other expenditures (0.0003%) and other debt service (0.0006%).

Residual 1: the resale schedule

Transfer tax includes resales of the accumulated stock (recordation is exactly transfer/2 everywhere, so both lines share one sales base). No constant turnover rate fits the recovered resale series: the study used an internal year-by-year schedule it never published. We carry the recovered schedule as data. For OUR tool the fix is better than the study's approach: a visible turnover-rate slider instead of a hidden schedule.

Residual 2: Table 2's own mix

Table 2 (the famous per-unit summary) is a standalone prototype whose per-type income and school mix inputs are unpublished; they do not match the year-0 column of the C tables (gaps up to ~$990/unit for condos) and no constant income-tax share explains all four types. Its property/fire/ recordation/capital components derive exactly; the remainder is carried as two labeled recovered constants per type. The C tables, which we reproduce in full, are the real validation of the architecture. Practical consequence: our future tool should present per-unit results built transparently from its own stated inputs, not chase Table 2's unpublished blend.

Residual 3: income-tax wobble (0.14%)

The income-tax line fits stock x one constant per type to at worst 0.14% (most cells far better). The wobble is consistent with the model internally carrying slightly time-varying income mixes. Within tolerance; noted for honesty.

Residual 4: the enrollment reconciliation, now quantified (2026-08-12)

Actual HCPSS enrollment in 2017-18 was 56,799 (official count, externally CONFIRMED; see VERIFICATION.md). The county appropriation per student was therefore 572,871,655 / 56,799 = 10,086 dollars. The model's recovered per-student cost is ~9,213-9,267 dollars, i.e. 91.3-91.9 percent of the full appropriation. This matches the methodology's stated fixed-vs-variable split (pp. 40-41): the model charges new students only the share of school spending it classified as variable, about 91-92 percent. The exact split percentage remains unpublished, but the gap is no longer mysterious; it is the split, measured.

Honest accounting of what "reproduced" means per line

Of the 13 lines in each table, five are the SOURCES the internal series were recovered from (real estate tax, school surcharge, all-other revenues, transfer tax, recordation tax). For those, reproduction is partly by construction. The genuine cross-predictions are the other eight lines (fire, income tax, road excise, both school lines, both other per-person lines, and net). The validation strength lives there.

Scope not attempted (published output we have NOT replicated)

The countywide scenario tables (Appendix B), the planning-area tables (Appendix D, which also carry the PAYGO adjustment we have never exercised), the per-planning-area prototype findings in the report body, and the nonresidential prototypes (C-13 onward). The study's HEADLINE numbers (the 63.4M six-year APFO cost, 151.7M over twenty years) come from those tables and are therefore still unverified by us. Replicating Appendix D is the natural next audit if we ever want to cite the headline figures.

What this means for v0 acceptance

SPEC.md's Table 2 targets are met via the derived components plus the two recovered Table 2 constants per type, with the caveat in Residual 2 stated openly. The deeper claim v0 actually earned: the model's ARCHITECTURE and constants are recovered and proven against ~2,000 published numbers at basis-point precision.