non-negotiables · bias & compliance
the audit that lets stage 3 ship (no black box)
We never collect protected attributes, so a demographic audit can only run on voluntarily-supplied data. What we monitor continuously is the proposal's real fairness worry (§17): how much does the score lean on school prestige? Verified education is a deliberate, job-related signal here — but proof-of-work dominates, and the engine selects talent far more evenly than a prestige-ranked system. Below is a live analysis run through the actual engine.
deterministic & reproducible
Same evidence in → same score out. The LLM never sets the score (§4 P1); every result carries a model_version.
fully explainable
Every pillar exposes a numeric subscore and a plain-English reason — no black box to defend (§8.2).
protected attributes excluded
Age, gender, ethnicity and name-derived inferences are never inputs (§5/§8.2).
proof-of-work dominates
An elite degree alone scores under 300/1000; shipped, verified work is the heaviest pillar (§4 P3).
education is job-related
Verified university & degrees are a weighted, domain-matched signal — a defensible validity argument, not a hidden proxy (§8.2).
human-in-the-loop
The employer view ranks evidence and never auto-rejects — a person makes every call (§8).
versioned & auditable
Scores are snapshotted (score_versions) and signed (§7.3) so any past decision can be replayed (§9).
adverse-impact analysis (4/5ths rule)
Synthetic cohort of 600 candidates where talent (proof-of-work) is independent of school prestige; selection = top half by score; threshold 555/1000. Reproducible (seeded). Model cs-heuristic-1.3.0.
careerscore engine
impact ratio 0.70Low-prestige candidates are still selected at 70% the rate of high-prestige ones — because proof-of-work is the dominant term, talented builders from any school surface. Verified education tilts the ratio below a strict 0.80, which is the honest cost of treating education as a job-related signal; we report it openly rather than hide it.
prestige-ranked baseline
impact ratio 0.00For contrast: a system that ranked purely by school prestige selects almost no low-prestige candidates (ratio ~0) — the adverse impact our proof-of-work weighting avoids.
methodology & limits (NYC Local Law 144-style)
- • Impact ratio = lowest group selection rate ÷ highest group selection rate; ≥ 0.80 passes the EEOC four-fifths rule.
- • Education (verified university & degrees) is an intentional, job-related signal, so the engine is not prestige-blind — but proof-of-work outweighs it, keeping the ratio far above a prestige-ranked baseline.
- • Protected-class impact ratios require demographic data we do not collect; before any regulated employer use, an independent audit runs on voluntarily-supplied data, refreshed at least annually.
- • Continuous monitoring re-runs this on real cohorts as volume accumulates (§4 P6); the tool augments a human decision and never auto-rejects (§8).