Measured, not marketed

The figures below are previously reported project results. This page does not include the raw runs, version pins, or enough methodology to reproduce them; request those artifacts before using the numbers for a deployment decision. They are not capacity or accuracy guarantees for your application.

	Result	What it means
GAIA — public agentic benchmark	+18.9 pts	Reported absolute score difference; inspect the task set and model/version controls
Grounded scale	256 concurrent agent workloads per 8×H100 node, 98.0% grounded success (single run)	We measure grounded answers, not HTTP 200s
Token economy	23× fewer tokens per task (~27K vs ~615K)	Under a tenth of the cost per answer
AgentDojo — prompt-injection defense	attack success 26.8% → 17.3%, benign utility held at 90.7%	Hardened without making the system useless

Full methodology and per-run data available on request — contact@meganova.ai.