How Current Conditions becomes Latent Value.
No black boxes.
1 · THE TWO-SCORE MODEL
Every place gets two scores, never one: CC (Current Conditions) and LV (Latent Value, a band). Gap = LV − CC is the investment thesis — the bigger the gap, the bigger the correction return.
2 · THE SUPPRESSION WEIGHT RECIPE
v1 placeholder calibration: ρ = 0.80, Ceiling = 100. Per-factor weights wᵢ are proprietary; the prototype ships the v1 placeholders that live in each place record.
3 · SOURCED SUPPRESSION FACTORS
Each factor carries its receipt. No unsourced suppression counts.
| Domain | Distortion factor | Source (receipt) |
|---|---|---|
| Housing & wealth | Redlining / disinvestment (1930s, still predicts today) | HOLC maps / Mapping Inequality (Univ. Richmond) |
| Housing & wealth | Appraisal bias (homes devalued for Black presence) | Brookings (Andre Perry); Freddie Mac |
| Credit & finance | Biased credit instruments ("credit invisible") | CFPB Credit Invisibles |
| Credit & finance | Predatory lending (higher rates at equal risk) | DOJ settlements; HMDA |
| Health | Medical-racism distrust (documented harm, earned skepticism) | Hoffman et al., PNAS 2016 |
| Health | Biased clinical algorithms (care under-allocated) | Obermeyer, Science 2019; eGFR; NEJM pulse-ox |
| Environment | Environmental burden (pollution / siting / particulates) | EPA EJScreen; CDC |
| Environment | Food apartheid (low-access designation) | USDA Food Access Atlas |
| Reach / capital | Under-captured reach (cultural value generated, capital not retained) | Royalty / talent-pay-gap / ownership data |
| Data | Majority-baseline modeling (one model applied uniformly) | Buolamwini & Gebru, Gender Shades 2018 |
| Data | Data deserts / undercount (under-measured → "low value") | Census undercount |
4 · THE KILL RULE
“If a factor's residual disappears once you add a legitimate control, it is dropped. Suppression that cannot survive controls does not count.”
5 · PUBLIC VS. PROPRIETARY (THE FICO MODEL)
| Layer | Disclosure |
|---|---|
| Suppression factors + sources + direction | Public — the credibility and the defense; the most transparent component. |
| Factor-level breakdown of a given score | Customer / auditor — they see what drove their own score. |
| Exact weights wᵢ, ρ, Ceiling, formula | Dharma only — the IP and the moat (factors public, formula private). |
“The thumb was already on the scale. Dharma is the one taking it off.”
6 · SCORE VALIDITY & RELIABILITY
Two properties, never conflated. Consistency / reproducibility is ours by design, today: a deterministic pipeline (fixed formula, no randomness), everything versioned as one pinned bundle (weights + factor table + citations), a factor-level audit trail, and confidence expressed as band width — missing data widens the band, it never silently moves the point.
Reliability / validity is partly proven now, fully at pilot. Does Latent Value actually predict recoverable value? It is defensible now via construct validity (every factor independently documented) and the ρ cap; proven once pilots close the loop.
The backtest plan is the centerpiece: find natural experiments where suppression was already partially lifted (appraisal-bias corrections, credit models adding rent/utility data, environmental remediation, desegregation investment), score the “before” state, compare predicted uplift to observed change, publish the error, and calibrate wᵢ, ρ, Ceiling against the error — not intuition. Consistent by design. Reliable by evidence. Validated by pilots.