About this role
Summary: Designs evaluations of conceptual reasoning in AI safety by turning expert questions into well-posed tasks with opinionated rubrics. Runs these tasks against frontier models and iterates until scores distinguish strong answers from weak ones.
- Engagement: Hourly contract, 20–40+ hours/week
- Start: ASAP
- Contract length: ~3 months, with potential to renew
- Reports to: [Ryan Carey](https://www.trajectorylabs.com/team) (MTS, Head of Conceptual Reasoning)
- Location: Remote, or on-site at Constellation in Berkeley
- Compensation: $100-150/hr
Overview
Trajectory Labs is hiring researchers to design evaluations of conceptual reasoning in AI safety. We build evals to measure how well models reason about hard, open-ended problems in AI governance, alignment, macrostrategy, and economics. You'll work 20–40+ hours a week, remotely or alongside us at Constellation in Berkeley.
We're an early-stage startup, so you should expect and enjoy that your responsibilities will grow and priorities change quickly.
About us
Our mission is to [automate AI safety](https://www.lesswrong.com/posts/F3j4xqpxjxgQD3xXh/ai-for-ai-safety), to pave the way for a future where the vast majority of AI safety work is done by AI models.
To that end, we build safety and alignment evals, red-teaming programs, and RL environments for frontier AI labs. Our areas of focus include red teaming, control, anti-distillation, AI safety reasoning, reward-hacking propensity evaluations, and embedded assessments.
Although [Trajectory](https://www.trajectorylabs.com) is only a year old, we have already been featured in Anthropic's [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf#page=59) (the only commercial red team to jailbreak its cyber safeguards) and Google DeepMind's [Gemini 4 Argon Safety & Security Report](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf), led Anthropic's [prompt injection evaluation of Auto mode](https://claude.com/blog/auto-mode-default-in-claude-code), and been featured in Meta's [Muse Spark Safety & Preparedness Report](https://ai.meta.com/static-resource/muse-spark-safety-and-preparedness-report/).
About the role
Over time, we want AI labour to be used effectively towards AI safety work, and this requires AI models to get better at some relevant [hard-to-verify tasks](https://arxiv.org/abs/2605.06390), like evaluating safety cases or reasoning about risks from future powerful systems. We're partnering with frontier AI labs to close that gap.
It's important to us that our work is robustly positive, so we focus our evals on the capabilities that differentially accelerate AI safety, not capabilities in general. While we take seriously the risk of spillover into other capabilities, we think it is crucial that AI models are able to help with safety and alignment, in time for when this work needs to be done.
What you'll do
- Find hard conceptual questions in AI safety — alignment, macrostrategy, economics, governance — where frontier models still reason badly.
- Turn each into an evaluation: a well-posed problem and an opinionated rubric that rewards excellent reasoning and penalizes slop.
- Interview experts in the field to turn their insights into conceptual reasoning tasks.
- Run your tasks against frontier models and iterate until the scores separate strong answers from plausible-sounding weak ones
What we're looking for
We're looking for researchers in any of the four areas below, or adjacent fields, who have strong, defensible opinions about what good reasoning looks like in their specific area of expertise.
- AI Governance: compute governance and verification, international agreements and security, AI law and institutions
- AI safety & alignment (conceptual): alignment theory, AI control and oversight, evaluations, AI futurism
- Macrostrategy: power concentration, takeoff dynamics, long-run trajectories, post-AGI ethics, digital minds and AI welfare
- Economics & decision theory: growth theory, AI’s effects on labor and capital, game theory and mechanism design, decision theory
### Essential
- Domain depth: you have produced original analysis on AI risk, alignment, macrostrategy, economics, or governance that you stand behind.
- Articulable taste: you can say why one argument in your field is better than another, precisely enough to encode it in a rubric.
- Calibration under pushback: you revise your judgment when the argument against it is good, and defend it with reasons when it isn't.
- Agent-native research: you work fluently with frontier models and agentic tools — you get real leverage from Claude Code or similar in your own work, and can tell the difference between output that is correct and output that merely sounds right.
### Desirable
- PhD-level research experience. All of our CR staff have studied and/or worked at top universities, but a PhD is not a hard requirement.
- Experience writing grading rubrics, exam questions, or evaluation criteria
- Peer-reviewed or widely-cited work in one of the domains above
- Experience using frontier models to answer messy questions in your own research
Role highlights
The evals you produce become part of the benchmarks that frontier labs use to measure — and improve — how their models reason about making TAI go well. You’d be one of the few people having a very large impact in this space.
Work with our top research team including PhDs or research fellows from Oxford University (x4) and LSE (x2) and the leading figures in your field.
Other details:
- Remote-friendly, async by default; desk space available at Constellation in Berkeley
- 20–40+ hours/week, flexible as your availability changes
- Starts immediately, rolling review