O — X

symbiotic intelligence

What we build

01

Agent Governance

Multi-agent work loses context when sessions end. We make claims, handoffs, and decisions durable, reviewable, and tracked in version control.

dot_swarm

02

Agent Security & Privacy

The safety layer agentic AI usually ships without: protected credentials, human approval on sensitive actions, complete activity records, and supply-chain review on the tools agents use.

oasis-claw

03

Alignment & Evaluation

Human values don't reduce to a single score. We map the geometry of competing objectives — helpfulness, safety, honesty, privacy, agency — so AI systems can navigate them without flattening trade-offs.

alignment_research

Principles

Human agency first Local-first privacy Auditability by default Risk-tiered deployment Ecological, adaptive design

Global AI Challenges

AI systems have to do more than perform tasks. They have to protect privacy, respect agency, and stay governable across their lifecycle.

Challenge Oasis-X response
Human Well-Being Workflows that preserve context across sessions and surface conflicts in human preferences instead of averaging them away.
Responsible AI Multi-objective value modeling, geometric safety barriers, and supply-chain audit on every installed skill.
Privacy Plaintext credentials never enter the LLM context. Local-first architecture. Signed logs for prompt-injection incidents.
Human-Centered Design A public risk ladder. Approval gates in the gateway. Explicit lifecycle primitives — agents can sleep, compact, and hand off.
Governance & Oversight Git-native traces, signed logs, and lifecycle artifacts. We run our own development on the protocol we ship.
Human-AI Interaction Inspectable, interruptible, and accountable. Operator approval runs through an explicit gate, and transcripts are append-only.

AI Risk Ladder

  1. 01 Assistive AI drafts, summarizes, suggests, or explains. Human decides.
  2. 02 Delegated AI performs bounded tasks with reviewable outputs and logs.
  3. 03 Operational AI can affect systems, customers, money, or data. Requires approval gates, monitoring, and rollback.
  4. 04 High-Impact AI affects health, legal, finance, employment, or essential services. Requires formal governance.
  5. 05 Prohibited Deceptive, coercive, privacy-invasive, or designed to remove meaningful human control.