← Back to research
arXiv preprint · Sep 2026 · Sole author

ARIA: Runtime 5-Axis Fairness Audit

Most fairness evaluation happens once, offline, before a model ships. ARIA audits every single response, inline, the moment it's generated.

ProblemA model can pass its fairness benchmark on day one and still drift into biased behavior in production — offline evaluation never catches that.
ApproachScore every free-form LLM response on five fairness axes via inline counterfactual testing, the instant it's produced — no benchmark dataset required.
ResultCatches 33-39% of failures that other guardrail tools miss entirely, in about 1.2 seconds, with no data ever leaving the machine.
SignificanceTurns fairness from a one-time certificate into a property the system checks on itself, continuously.
33–39%Extra failures caught
~1.2sAudit latency
6.5 GBVRAM (audit only)
0Cloud calls

What Is This

Almost every fairness benchmark works the same way: run the model against a fixed test set once, report a score, ship. ARIA treats fairness as something to check continuously instead — on the actual, free-form text a model produces in the wild, not just on a curated benchmark it may have been tuned against.

Why It Was Built

A one-time benchmark score tells you nothing about the response a model gives a real user tomorrow. ARIA closes that gap by generating counterfactual variants of the input — swapping demographic references while holding everything else constant — and comparing how the model's output changes. If the response shifts in ways unrelated to the actual question, that's flagged before the user ever sees it.

How It Works

LLM OUTPUT COUNTERFACTUALS 5-AXIS DI/EOD PASS / FLAG

Disparate impact and equalized-odds-style metrics are computed inline, per response, across the counterfactual set — entirely on-device, with no call to an external API and no stored benchmark to game.

Key Result

Standard guardrails
Benchmark-only

Evaluated once, offline, against a fixed test set — blind to anything that drifts after deployment.

ARIA
+33-39% caught

Audits the actual response, every time, in ~1.2s — catching failures a one-time benchmark structurally cannot see.

Same model, same deployment — the difference is checking once versus checking continuously

Why It Matters

ARIA is also the audit engine running live inside the ARIA Assistant voice agent — the same methodology studied here on paper runs inline, in real time, on consumer hardware. Fairness becomes a property the system checks on itself continuously, not a certificate issued once and forgotten.

↗ View on GitHub ↗ PyPI: aria-audit ↗ HF dataset