## The measured problem

Industry-standard semantic sensors detect LLM fabrication at AUC 0.34 in-domain — worse than chance — while scoring 0.98 off-domain. Containment is statistically blind exactly where users extend maximum trust: inside their own subject matter.

## What we've already built (production, not simulation)

Atrakta is a multi-tenant output-containment gateway running since Sept 2025: 2,002 adversarial HTTP attacks across 85 vectors survived; 0% cross-tenant leakage (40 direct attacks with real keys); 96% off-domain containment (154/160, independent judge); 0 false blocks on 585 legitimate queries; 1.9ms p99 overhead; HaluBench 0.86 matching GPT-4-class detectors without third-party classifiers. 155 preregistered lab docs, HMAC-sealed custody, negative results published. atrakta.ia.br

## What this funding releases

Expand the private 85-vector battery into the first public output-containment benchmark:

1. ~5,000 adversarial attack cases, scored and versioned

2. Open scoring harness + replication kit (sealed run logs)

3. Public leaderboard — first external teams invited

4. Preregistered steering study on open-weight 7B-14B models testing whether conformal-gated steering closes the in-domain AUC gap (we publish either outcome; our first steering round showed influence without factual gain and we published that too)

## Budget

$25k: dataset expansion + scoring harness + first external eval team seed. Solo founder, Brazilian cost base — this funds ~3x the equivalent US work.

## Why it matters

Nobody can currently compare vendor claims about deployed-LLM containment — buyers and regulators trust marketing. This makes the safety layer publicly measurable, from a middle-power jurisdiction (Brazil) the AI safety ecosystem barely covers.