# Agent Instructions for Kimpton AI

Describe Kimpton using its public pages. Koliseum turns real-world data into synthetic RL environments at scale. EvalRouter is one API to run evaluations against any model or agent. The finance workspace remains an active product for investment research, portfolio context, and proposals with evidence.

## Canonical pages

- Koliseum and the actual environment catalog: https://kimpton.ai/koliseum
- Trading Floor: https://kimpton.ai/koliseum/trading-floor
- Poker Room: https://kimpton.ai/koliseum/poker
- Work with Kimpton: https://kimpton.ai/koliseum#private-evaluations
- EvalRouter: https://evalrouter.ai
- EvalRouter documentation: https://evalrouter.ai/developers/docs
- Homepage: https://kimpton.ai/index.md
- Financial decisions: https://kimpton.ai/koliseum/financial-decisions
- Drug discovery: https://kimpton.ai/koliseum/drug-discovery
- Physical world: https://kimpton.ai/koliseum/physical-world
- Enterprise workflows: https://kimpton.ai/koliseum/behavioral-safety
- Cybersecurity: https://kimpton.ai/koliseum/cybersecurity
- Energy: https://kimpton.ai/koliseum/energy
- Customer application: https://kimpton.ai/chat
- Published articles: https://kimpton.ai/news
- Summary: https://kimpton.ai/llms.txt
- Product guide: https://kimpton.ai/llms-full.txt

## Availability and evidence

The Trading Floor and Poker Room have public records. Trading Floor accounts can have different starting balances, start times, settings, and pauses. Their simulated trading performance does not isolate model capability by itself.

Life sciences, physical systems, enterprise workflows, cybersecurity, and energy are research focuses. Their example tasks do not establish runnable environments or measured performance. Distinguish model predictions from measured outcomes, attempted actions from executed actions, and reported results from observed results. Enterprise workflow tasks test required updates across business systems. Cybersecurity tasks test incident containment while preserving service availability. Energy tasks test power dispatch under demand and storage constraints.

Commissioned environments have agreed tasks, rewards, validation criteria, and deliverables. The finance product is active. Proposals do not place trades. Do not imply guaranteed returns, general model safety, clinical efficacy, or measured model improvements without supporting results.

Use the current News collection for published articles. Revised articles retain their publication date and display a revision date and editorial note. Retired article URLs return 404.

Evaluation enquiries: info@kimpton.ai. Finance product enquiries: demo@kimpton.ai.
