← Lab Handbook

Lab Expectations: Operating as a DRI

This page describes what successful project ownership looks like in the lab. The framing is adapted from Ethan Perez’s tips for empirical alignment research — a rubric built for a high-throughput empirical ML organization — modified for a methodological research environment, where correctness and understanding carry more weight than experiment volume. Companion pages: Starting a Project and Communication Rules.

By roughly the one-year mark, the goal is for you to operate as the Directly Responsible Individual (DRI) for your research projects: the person who moves the work forward, tracks progress, and identifies roadblocks. No one arrives as a DRI, and you are not expected to. In your first year, most tactical direction will come from me; the handoff is gradual, and we will discuss where you are in it explicitly during 1:1s.

The percentages below are a mental model of relative importance, not an evaluation formula.

Executing & understanding (70%)

Thoughtful implementation (45%). In methodological research, speed and understanding are in constant tension. We do not prize a massive volume of hasty experiments — a high volume of poorly designed simulations is worse than no work at all, because someone eventually has to figure out which results to trust. The expectation is a medium volume of high-quality work: implementations careful enough that you deeply understand the data, can refine the methodological approach, speed up your algorithms, and ask progressively better questions each cycle.

Troubleshooting & autonomy (25%). When code breaks or a simulation misbehaves, you take the first pass at diagnosing why. Between meetings you can independently explore logical next steps and resolve standard R and Python issues, including problems with your own local development environment. Getting unstuck is a research skill, not overhead.

Driving project direction (20%)

Tactical direction (10%). You make sound day-to-day decisions about which analyses and evaluation metrics to prioritize. You operate well with standard check-ins, and you reach out when genuinely blocked — not before you’ve made a serious attempt, and not after you’ve burned three days on it.

Strategic vision (10%). You are beginning to spot new methodological gaps and project ideas in your research area. For some of you that area sits squarely within the lab’s core focus — adaptive interventions, causal inference for mobile health, and sequential decision-making — and for others it sits further afield. Either way, the expectation is the same: a vision that builds outward from the lab’s methodological foundation rather than floating free of it.

Clear communication (5%)

Your central working document — for us, the Project Landing Page, the Google Doc described in Starting a Project — stays current. Whether you share an RMarkdown file, a Jupyter notebook, or present live, your updates and plots are clean enough to digest quickly — so our 1:1 time goes to the implications of your results rather than untangling raw output or messy code. Weekly updates follow the template in Communication Rules.

Lab citizenship (5%)

You help your peers, contribute to shared lab resources, and take constructive feedback seriously. Five percent of the rubric, but disproportionate in determining whether this is a good place to do research.


Default norms for projects with me

Research is a stochastic decision process. We are exploring unknowns, so expect variance in timelines and be prepared to pivot when evidence points somewhere better. Jacob Steinhardt’s Research as a Stochastic Decision Process is the canonical statement of how we sequence work: attack the highest-uncertainty steps first, because they are the cheapest place to learn you are wrong.

OODA loops over strict deadlines. We prioritize rapid iteration — Observe, Orient, Decide, Act — over rigid schedules, following Ben Kuhn’s playbook for running major projects. Update your priors quickly when models fail to converge or data behaves unexpectedly; the iteration cycle, not the calendar, is the unit of progress.