# GrayTell

> GrayTell is an AI safety research lab. Ask a research question, get a cited answer across nine model families — GLM, GPT, Qwen, OpenRouter, Nemotron, Gemma, Cohere, Liquid, and Mistral. The reasoning is shown the whole way.

GrayTell ships two things:

1. **Shield** — the alignment layer that sits in front of any frontier model (open-weights or closed API) and provides real-time refusal repair, prompt-injection filtering, and PII redaction. Measured at 71% fewer policy-violating outputs at <40ms added latency.
2. **The research console** — ask a question, watch the agent search the web, read sources, and reason between them, then get a final answer with inline citations on every claim. Supports nine model families and persistent memory.

## Links

- [Home](https://graytell.space-z.ai/)
- [About GrayTell](https://graytell.space-z.ai/about)
- [Research console](https://graytell.space-z.ai/chat)
- [Sign in](https://graytell.space-z.ai/auth)

## What we do

- **Cited answers**: every fact, number, and quote carries an inline citation chip linking back to the exact source it came from. No unsourced assertions, no invented URLs.
- **Nine model families**: GLM, GPT, Qwen, OpenRouter, Nemotron, Gemma, Cohere, Liquid, and Mistral — switch providers mid-conversation and compare how each one reasons about the same question.
- **Shows its working**: you watch the research happen in real time — the searches it ran, the pages it read, the reasoning between them. The receipts are attached to every answer.
- **Persistent memory**: the agent remembers facts about you across conversations — your preferences, your projects, your sources.
- **Safety research**: published work on shielded decoding, lineage cards, cross-family safety transfer, the frontier-stack audit, calibrated abstention, and sycophancy in frontier models.

## Optional

- [Sitemap](https://graytell.space-z.ai/sitemap.xml)
- [Robots](https://graytell.space-z.ai/robots.txt)