両界Ryōkai
Evolved from Furiwake’s model router, Ryōkai routes tasks to agents and agent swarms.
- Agent routing
- Constrained selection
- Same specification
序A research lab that shipsThe space between
The space between
thinking and acting.
Blossom AI is a research lab and a product company. We study how models should reason, route, and collaborate with experts — and we put that work into systems that operate reliably inside operationally complex businesses.
思考と行動の間·The space between thinking and acting
製品の角度How we build products
Four angles. One platform.
01
両界Evidence before assignment
Agents or complementary teams are selected within budget, deadline and policy, and their work is verified against the same task specification.
02
校正Calibrated under load
We grade refusal, deferral, and uncertainty as first-class outcomes — not failures swept under the rug.
03
評価Pre-deployment evaluation
Evaluation and improvement before deployment, in CI/CD.
04
監査Auditable by design
Designed so decisions can be traced to a policy version, replayed and inspected by the team that has to defend them.
二層Two layers, one disciplineA lab and a product.
A lab and a product.
Each makes the other honest.
Research without deployment becomes performance art. Product without research becomes a wrapper. We hold the two together — what we learn in the lab ships into our products; what we see in production sets the lab’s next question.
研究室LabsResearch
How should AI systems learn, reason, and collaborate with human experts?
Blossom Labs studies the open questions underneath modern AI deployment — scalable knowledge discovery, calibrated reasoning, and reinforcement learning in simulated operations. We publish what we find.
- 推論Reasoning under distribution shiftHow models stay calibrated when the world moves.
- 協働Human–expert collaborationWhere to ask, where to defer, where to act.
- 模擬RL in simulated operationsPractice in environments before production.
製品ProductDeployments
Products and capabilities that put research-grade AI into operationally complex businesses.
Logistics, finance, manufacturing, healthcare ops — domains where a small share of bad decisions costs real money or worse. Operators get task routing to agents, pre-deployment evals, and long-horizon agents with human approval where it matters.
- 両界RyōkaiTasks routed to agents and agent swarms.
- 評価EvaluationPre-deployment evaluation and improvement in CI/CD.
- 行動AgentsLong-horizon work with human approvals where they matter.
製品The product
All product docs Products for agent labor.
評価Evaluation
Blossom Eval is designed for pre-deployment evaluation and improvement in CI/CD.
- Pre-deployment, in CI/CD
- Well-defined objectives
- Open-ended tasks
行動Agents
Long-horizon agents for operationally complex work. Human approvals where they matter, autonomy where they don't — with traces for audit, configured per project.
- Approval policies
- Tool sandboxing
- Replayable traces
流れHow it flowsFrom research question
From research question
to live operations.
We don’t separate “ideation” from “delivery.” Every loop closes back into the lab.
01問い
Question
A research question rooted in something we saw in real production.
02実験
Experiment
Simulations, evaluations, ablations. We publish what we find.
03橋渡し
Crossing
Findings become operating principles — codified, peer-reviewed, versioned.
04実装
Implementation
Principles ship into Ryōkai, Eval, or Agents — gated by an internal eval bar.
05運用
Operations
Production telemetry feeds back; the lab's next question begins.
研究Recent work
All research From the lab.
研究 · Research
Outcome Is Not Discipline
A strategic agent can hit the business KPI while failing to preserve the behavior that makes it deployable.
研究 · Research
When the Trace Decides
An agent can hit the outcome metric while learning the wrong behavior. Better evidence follows the trace.
研究 · Research
When the Benchmark Ruler Moves
A benchmark can have gold labels and still produce unstable evidence if the rubric, metric, or ranking rule moves.
研究室より
The questions worth answering only show up under load — and the products worth shipping demand the rigor of a lab.
From the lab — May 2026
研究室の現況
What the lab is working on now.
- Routing policies under multi-tenant drift
- RL agents in simulated logistics environments
- Expert-in-loop training for Operations
- Calibrated refusal for long-horizon work