CT69.A1.Reparations: Methodology, sources, assumptions

Status: first full pass, 2026-11-12. All numbers come from python scripts/run_all.py; see data/processed/.

Scope. This values the labor of enslaved people in what became the United States, 1619–1865, as a clinical economic quantity. It does not value the harm of enslavement, lost freedom, violence, family separation, or post-1865 discrimination. Every figure is a model output under stated assumptions, not a measurement. The interest-rate assumption dominates everything else (§5).

1. Questions and headline answers

2. Data sources (all in data/raw/, SHA-256 in data/raw/MANIFEST.json)

Literature used for checks and context: Craemer, T. (2015), "Estimating Slavery Reparations: Present Value Comparisons of Historical Multigenerational Reparations Policies," Social Science Quarterly 96(2): 639–655. Hacker, J. D. (2020), "From '20. and odd' to 10 million: the growth of the slave population in the United States," Slavery & Abolition 41(4): 840–855. Evans, R. (1962), "The Economics of American Negro Slavery," in Aspects of Labor Economics, NBER, pp. 185–256. O'Malley, G. (2014), Final Passages: The Intercolonial Slave Trade of British America, 1619–1807.

3. Method

Labor value. For state s and year y:

V(s,y) = Enslaved(s,y) × participation × hours/yr × w_real(y) × wage_multiplier × (1 − maintenance_share)
PV(r)  = Σ_s Σ_y V(s,y) × (1 + r)^(2025 − y)

• Enslaved(s,y) combines colonial estimates for 1630–1780 and census counts for 1790–1860, linearly interpolated year by year. Virginia is set to 20 for 1619–1620. 1861–1865 holds the 1860 count.
• w_real(y) is the hourly wage of unskilled free labor, in 2025 dollars. MeasuringWorth's unskilled-labor index is converted to dollars per hour using the average ratio of production-worker pay to the index over 1850–1870 (k = $0.000777 per index point, which gives $0.078/hr nominal in 1860). It is then deflated with MeasuringWorth's CPI. Before 1774 there is no series, so the 1774 real wage is held constant.
• The result: 1860 real wage ≈ $3.11/hr (2025 $), or about $9,300 a year at 3,000 hours.
• r is a real (inflation-adjusted) rate, because the values are already in constant dollars.

Validation against the literature. Rebuilding Craemer's construction with these inputs (1776–1865, nominal wages compounded at 3% to 2009) gives $3.9T, against his published $5.9–14.2T. Same order of magnitude, a little lower, consistent with this model's lower wage and participation assumptions. Note that a 3% nominal rate on nominal wages is roughly a 1.5% real rate. That is why the "real 3%" result here ($332T) is so much larger than Craemer's.

Land purchase (per state, value V to spend):

1. Purchasable land = non-federal acres (CRS) split into four tiers, bought in this order:
2. Non-family corporate farmland
3. Partnerships and family-held corporations
4. Family or individual farms
5. Other non-federal land: non-federal acres minus land in farms. This includes state, municipal, urban and forest land.
6. Excluded: federal land, and farmland under "Other: estate or trust, prison farm, grazing association, American Indian Reservation", so tribal land is never "bought".
7. Price: USDA 2025 farm real estate $/acre for every tier (§6). The cost is acres bought × price. Value beyond the state's purchasable land is reported as unspent.
8. Two allocation modes:
9. State where the labor was done: each state's share of the present value buys land in that state.
10. All states, same share: the national present value buys the same fraction of every contiguous state's purchasable land.

Lives lost (item 5). Deaths = embarked − disembarked, by decade. Two separate methods:

• (a) Deaths × VSL. The VSL is already a present-day value, so it is not compounded.
• (b) Forgone labor: each person who died is credited with 30 years of the same hourly labor value from the decade's midpoint, capped at 1865, and compounded at r.

Neither is added to the labor total, and descendants who were never born are not counted.

4. Assumptions (all live in scripts/params.json)

5. Sensitivity to the interest rate (the dominant assumption)

Present value of the labor, in 2025 dollars. The band varies A1 and A2 together.

Each extra point of real interest multiplies the result by about 7× (at low rates) up to about 20× (at high rates). That is because labor from 1700 compounds for 325 years.

Which rate is defensible?

• Long-run real returns on safe assets (government debt) have historically been around 0–3%.
• Real equity returns have been around 5–7%. Applying them over 200–400 years produces sums far larger than all the world's wealth. That is a property of compounding, not a finding.
• This project does not pick a single rate. It shows the whole range. The visualization defaults to 3% only because Craemer's widely cited work uses 3% (nominal).

6. Known limitations, stated plainly

1. Land prices. USDA prices farm real estate. Urban and suburban land costs far more per acre, so "% of state land" is overstated where land is urban, and the cost is understated. There is no consistent statewide market price for all land types. The FHFA/AEI county land-value data cover residential land only.
2. Corporate vs family farms. The Census of Agriculture classifies operations, not owners. About 39% of farmland is rented, often from non-operator landlords, including investors and trusts. "Non-family corporations" hold only 16.6M of 880M farm acres (1.9%) nationally. So "take from corporations first" runs out almost immediately, and most of the purchase falls on family-run operations and other land. USDA's only owner-level survey (TOTAL, 2014) is outdated.
3. Sectors. The crop split is a national 1850 snapshot of people of all ages, with no stated basis. It is not a state-by-state or year-by-year allocation of labor value. A better version would use 1860 Census of Agriculture crop output by county.
4. Colonial counts are HSUS estimates (decennial, rounded), not censuses.
5. Middle Passage figures cover deaths on board only. Deaths during capture, overland marches and coastal holding in Africa are much larger and not counted.
6. Free Black people are excluded from the census years (enslaved counts only) but included in the colonial years (A7).
7. Not modeled: stolen wages after 1865 (convict leasing, sharecropping), Jim Crow-era losses, and the value of the people as "capital" (slave prices). Each would be a separate study.

7. Verification

• Every cited number, and where and how it was checked, is listed in data/manual/source_verification_log.csv.
• scripts/03_qa.py re-runs the checks every time and writes QA_REPORT.md. It covers:
• source checksums
• transcription totals
• cross-source agreement, including official 1860-census decade totals, CRS state sums, the NASS US average, and SlaveVoyages estimates vs documented voyages
• literature figures re-found in the saved sources
• model unit and totals invariants, and an independent recomputation of the present value
• land-purchase invariants
• one-at-a-time sensitivity runs
• The run exits non-zero if any check fails.
• Two cited figures are flagged as partial or unverified (the ~1M interstate sales; the ~50k reshipped from the Caribbean). They are used for context only, never in a calculation.
• The US DOT VSL page blocks scripts, so it was verified by hand in a browser (logged as MANUAL).

8. Reproduce

python3 scripts/run_all.py            # fetch (skips existing) → parse → model → QA → .md→.html
python3 scripts/capture_images.py     # optional: refresh images/ with headless Edge/Chromium

Then open index.html (double-click; three.js r147 is bundled in lib/three/). Requires Python 3.10+. Optional pypdf enables QA check L3.