Research Engineer, Model Evaluations

<div class="content-intro"><h2><strong>About Anthropic</strong></h2> <p>Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.</p></div><h2 class="text-text-100 mt-3 -mb-1 text-[1.125rem] font-bold" data-sourcepos="3:1-3:18;40-57">About the role</h2> <p class="font-claude-response-body break-words whitespace-normal leading-[1.7]" data-sourcepos="5:1-5:267;59-325">We're looking for Research Engineers to build the evaluations that tell us — and the world — what Claude can actually do. Your work will turn ambiguous notions of "intelligence" into clear, defensible metrics that researchers, leadership, and the public can rely on.</p> <p class="font-claude-response-body break-words whitespace-normal leading-[1.7]" data-sourcepos="7:1-7:548;327-874">You'll design and implement evaluations across the full spectrum of Claude's capabilities and personality, and build the infrastructure that runs them reliably at scale. You'll partner closely with researchers throughout the lifecycle of a new capability — from defining what to measure, to running the eval against live training checkpoints, to interpreting the results. The goal is to make Anthropic the leader in extremely well-characterized AI systems, with performance that is exhaustively measured and validated across the tasks that matter.</p> <h2 class="text-text-100 mt-3 -mb-1 text-[1.125rem] font-bold" data-sourcepos="9:1-9:24;876-899">Key responsibilities</h2> <ul class="[li_&amp;]:mb-0 [li_&amp;]:mt-1 [li_&amp;]:gap-1 [&amp;:not(:last-child)_ul]:pb-1 [&amp;:not(:last-child)_ol]:pb-1 list-disc flex flex-col gap-1 pl-8 mb-3" data-sourcepos="11:1-18:112;901-2127"> <li class="whitespace-normal break-words pl-2" data-sourcepos="11:1-11:212;901-1112">Design and run new evaluations of Claude's capabilities — reasoning, agentic behavior, knowledge, safety properties — and produce visualizations that make the results legible to researchers and decision-makers</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="12:1-12:152;1113-1264">Build and harden the distributed eval execution platform so hundreds of evals run reliably against checkpoints throughout production RL training runs</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="13:1-13:180;1265-1444">Own the dashboards researchers and leadership use to monitor model health during training, improving signal-to-noise, reducing latency, and making regressions impossible to miss</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="14:1-14:178;1445-1622">Debug anomalous eval results mid-training-run, determine whether the cause is a model change or an infrastructure issue, and communicate the answer clearly under time pressure</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="15:1-15:104;1623-1726">Improve the tooling, libraries, and workflows researchers use to implement and iterate on evaluations</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="16:1-16:155;1727-1881">Partner with research teams across the full lifecycle of a new capability — from defining what to measure to interpreting results as training progresses</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="17:1-17:134;1882-2015">Run experiments to characterize how prompting, sampling, and scaffolding choices affect results on internal and industry benchmarks</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="18:1-18:112;2016-2127">Communicate evaluations and their results to internal stakeholders and, where appropriate, external audiences</li> </ul> <h2 class="text-text-100 mt-3 -mb-1 text-[1.125rem] font-bold" data-sourcepos="20:1-20:26;2129-2154">Minimum qualifications</h2> <ul class="[li_&amp;]:mb-0 [li_&amp;]:mt-1 [li_&amp;]:gap-1 [&amp;:not(:last-child)_ul]:pb-1 [&amp;:not(:last-child)_ol]:pb-1 list-disc flex flex-col gap-1 pl-8 mb-3" data-sourcepos="22:1-26:113;2156-2682"> <li class="whitespace-normal break-words pl-2" data-sourcepos="22:1-22:84;2156-2239">Strong Python programming skills, including production or research infrastructure</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="23:1-23:131;2240-2370">Experience building or operating distributed systems, data pipelines, or other infrastructure that needs to be reliable at scale</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="24:1-24:106;2371-2476">Clear written and verbal communication, especially when explaining technical results to non-specialists</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="25:1-25:93;2477-2569">Comfort operating in an on-call or production-support capacity when training runs are live</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="26:1-26:113;2570-2682">Care about the societal impacts of your work and an interest in steering powerful AI to be safe and beneficial</li> </ul> <h2 class="text-text-100 mt-3 -mb-1 text-[1.125rem] font-bold" data-sourcepos="28:1-28:28;2684-2711">Preferred qualifications</h2> <ul class="[li_&amp;]:mb-0 [li_&amp;]:mt-1 [li_&amp;]:gap-1 [&amp;:not(:last-child)_ul]:pb-1 [&amp;:not(:last-child)_ol]:pb-1 list-disc flex flex-col gap-1 pl-8 mb-3" data-sourcepos="30:1-38:43;2713-3386"> <li class="whitespace-normal break-words pl-2" data-sourcepos="30:1-30:113;2713-2825">Hands-on experience using large language models such as Claude, including prompting, sampling, and scaffolding</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="31:1-31:107;2826-2932">Background in data visualization and a track record of building dashboards people actually trust and use</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="32:1-32:70;2933-3002">Experience developing robust evaluation metrics for language models</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="33:1-33:76;3003-3078">Experience with observability, monitoring, or experiment-tracking systems</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="34:1-34:51;3079-3129">Background in statistics and experimental design</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="35:1-35:73;3130-3202">Experience with large-scale dataset sourcing, curation, and processing</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="36:1-36:62;3203-3264">Experience running or supporting ML training infrastructure</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="37:1-37:79;3265-3343">A bias toward picking up slack and operating flexibly across team boundaries</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="38:1-38:43;3344-3386">Enjoy pair programming — we love to pair</li> </ul> <h2 class="text-text-100 mt-3 -mb-1 text-[1.125rem] font-bold" data-sourcepos="40:1-40:27;3388-3414">Representative projects</h2> <ul class="[li_&amp;]:mb-0 [li_&amp;]:mt-1 [li_&amp;]:gap-1 [&amp;:not(:last-child)_ul]:pb-1 [&amp;:not(:last-child)_ol]:pb-1 list-disc flex flex-col gap-1 pl-8 mb-3" data-sourcepos="42:1-45:151;3416-4106"> <li class="whitespace-normal break-words pl-2" data-sourcepos="42:1-42:222;3416-3637">Stand up a new eval that tests a specific reasoning capability from scratch — define the task, build the dataset, implement the scoring, validate against known signals, and ship a dashboard that makes the result legible</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="43:1-43:187;3638-3824">Diagnose a mid-training regression: an eval suite returns anomalous numbers, and you need to determine within hours whether it's the model, the harness, the data, or the infrastructure</li> <li class="whitespace-normal break-words pl-2" data-sourcepos="44:1-44:131;3825-3955">Take a flaky distribut