Founder & CEO, Arobi Technology Alliance, Inc. · Superintelligence researcher. I build AI that can show its work: agents and machines that act only with current authority and leave evidence that anyone can check afterwards.
Ullumii · Download · HUE · Research · Benchmarks · Live status · Company
Intelligence is not authority. A system should be able to show who was allowed to act, under which policy, what actually happened, and what evidence supports the result.
- Three Pillars to Superintelligence Abundance (manifesto, 2026-10-31). Covers a universal standard of living, accountable superintelligence, and humanity united toward a cooperative future in the cosmos.
- Governing Autonomous Intelligence (white paper). Argues that safe AI autonomy is structural: intelligence proposes, an independent authority layer decides at the moment of action, and every action leaves evidence.
- The Arobi Familiar Memory Plane (whitepaper). Explains how Arobi keeps AI memory from becoming authority.
All of it is at aura-genesis.org/research.
| Product | What it is | Where |
|---|---|---|
| Ullumii Cloud | AI workspace in the browser. Free to start; paid tiers from $24/mo. | iorch.net/ullumii |
| Ullumii Desktop | Signed desktop app for Windows and Linux, with auto-updates and a built-in browser that agents can use when you approve it. | Download |
| HUE 2.2.6 | Terminal agent for Windows, Linux and macOS (Apple silicon). Signed builds with published SHA-256 checksums. | aura-genesis.org/hue |
| Q | The AI inside Ullumii that answers and does the work. Answers stream, with the first words in 1.0 to 1.6 s. | Status |
| Immaculate | Checks each consequential action against current authority and policy before it runs. If the check can't pass, the action doesn't run. | iorch.net/immaculate |
| LaaS + Verifier | Sealed evidence records for decisions and effects, plus a verifier that recomputes them on its own. | LaaS · Verifier |
| Arobi Network | A live production chain that anchors evidence. It has over 1.57 million blocks and 1,630 redacted receipts anyone can inspect. | Readiness benchmark |
| Whisper | End-to-end encrypted messaging. | aura-genesis.org/whisper |
The public status page lists 7 production services: Cloud, Q, Immaculate, the agent browser, Desktop updates, Arobi Network and evidence records. All 7 reported working on 2026-10-29.
Full tables, settings and limits are in BENCHMARKS.md. These are the highlights:
| Benchmark | Result |
|---|---|
| Rubik's cube, tool-assisted orchestration (2026-10-20) | Ullumii solved 8 of 8 scrambles from depth 3 to depth 12, with 0 invalid moves. It used about 2,000 tokens per game, 2 or 3 model calls, and a median of 3.9 s per game. |
| Rubik's cube, frontier models reasoning move by move (2026-10-18/19) | Codex solved 8 of 8 using 181k to 917k input tokens per game. Claude Code solved 3 of 4 before its preset plan-usage cap. |
| Visual reasoning, held out (2026-10-17) | 85% (51 of 60), 95% interval 74–92%. Charts, receipts, tables, grids, geometry and settings screens were all 6 of 6. Each answer cost about $0.0001 (median 5.2 s). |
| Q response latency (2026-10-18) | First visible word in 1.0–1.6 s with streaming. A buffered answer on the same day took 17.4 s. |
| Arobi readiness benchmark (live) | 100/100: 8 of 8 public probes pass and 0 fail. The probes cover the chain, receipts, seal, replay, policy gates, entitlement and checkout. |
Every benchmark was preregistered or run on a fixed game set, scored by fixed rules, and logged per run. Where a result is negative or doesn't support a claim, we say so. See What we don't claim in BENCHMARKS.md.
- Prereg first. We write down the hypotheses, thresholds and held-out seeds before the first model call. Deviations are appended, never edited away.
- Find what works by elimination. In the vision study, 6 popular prompting techniques and 2 tools were each tested against a preset bar. None cleared it, so none shipped.
- Honest comparisons. Every subject plays the same games under its own published plan. A tool-assisted result is labelled as tool-assisted and never shown as the same task as move-by-move reasoning.
- Replays are checked. Every benchmark clip is rebuilt from the engine log and checked state by state. If a single frame mismatches, the clip is deleted.
- Authority before effect; evidence after effect.
- Current authorization beats remembered permission.
- Material change requires revalidation; human hold where required.
- Exactly-once protection where duplicate physical or financial effects matter.
- Preserve failures, contradictions and inconclusive results.
- Independent recomputation instead of trusting a dashboard.
- Nothing mocked, placeholder or hardcoded is ever presented as production proof.
TypeScript / Node.js, Rust, Python and Go. Cloudflare Workers, Netlify, Postgres and Substrate. Cryptographic evidence, policy enforcement, agent harnesses, model routing, memory and context systems, signed desktop releases, and robotics assurance.
All Arobi source code is proprietary to Arobi Technology Alliance, Inc., and its repositories are private. This profile shows what is public: shipped products, live endpoints and measured results. Interested in a pilot, a partnership or an investment? Start at aura-genesis.org.
Arobi is not an insurer. No repository, benchmark, demo or public page is a safety certification or an independent third-party certification. Read each claim together with its date, method and stated limits.