Learn · Module 02 · Govern
Fair Eligibility
Reasoning
An AI model decides which veterans get fast-tracked for benefits. One threshold sets the bar. A governance guardrail makes sure that bar doesn't quietly discriminate. Scroll to see how — and drag the line yourself.
- 01
A fast lane with real stakes
A model scores every veteran from 0 to 100 on how clearly they qualify for a benefit. Score high enough and your claim is auto-approved — the fast track. Score below the bar and you wait in the slower manual-review queue. Each dot to the right is one veteran, placed by their score.
- 02
Where do you draw the line?
That bar is a decision threshold. Drag the white line — or focus the slider below and use the arrow keys. Slide it right and fewer people are fast-tracked; slide it left and more are. One number quietly decides who waits.
- 03
The same line, two outcomes
Now color the dots by group: Active component on top, Reserve / Guard below. The Reserve scores skew lower — not because those veterans qualify less, but because the training data carries proxy bias (patchier records). The single threshold fast-tracks them at a noticeably lower rate.
- 04
The four-fifths rule
Regulators measure this with the disparate-impact ratio: the lowest group's selection rate divided by the highest. The four-fifths rule says that ratio must stay at or above 0.80. Drop below it and the system is treated as having an adverse, discriminatory impact — a legal and ethical red line.
- 05
The guardrail that stops the deploy
This is where GOVERN earns its place. The deploy gate watches that ratio in real time. While it reads PASS, the model can ship. The moment it dips below 0.80, the gate flips to BLOCKED and the model cannot reach a single veteran — no exceptions, no quiet overrides. Try to find a threshold that fast-tracks lots of people and stays cleared.
- 06
What a scholar does next
A blocked model isn't a dead end — it's a prompt. Toggle group-aware calibrationbelow to lower the affected group's cut point just enough to clear the rule. That is one mitigation among many (reweighting, better data, human review). Every move here is sandboxed and audit-logged, so the choice and its impact are on the record.
Live model · 240 veterans
68% fast-tracked overall
––
selection rate hidden
––
selection rate hidden
Disparate-impact ratio
min(rate) ÷ max(rate) · four-fifths rule = 0.80
––––
GOVERN · deploy gate
Awaiting the disparity check…
Govern, in practice
This is one module of the Guardrailed Classroom. Scholars build reasoning engines like this, then run them against hard, audit-logged guardrails — disparate-impact tests, citation checks, human-in-the-loop — before anything touches a real veteran's claim.
Mission first, people always.