DriftLab

A browser-based battery of cognitive psychophysics tasks (Stroop, Flanker, N-back, and more) paired with a Python analysis pipeline that uses drift-diffusion modeling (DDM) to separate processing speed from decision threshold.

🚧 Work in progress — Fase 1 (MVP). Stroop and a simple RT task are playable; DDM fitting comes in a later phase.

What is this, and why?

Tasks like the ones below ask you to respond as fast and accurately as possible to a simple stimulus — a color word, a row of arrows, a sequence of letters to remember. Cognitive science has used tasks like these for decades because they're cheap to run and surprisingly revealing about how the mind processes information.

Most analyses stop at averages: mean reaction time and percent correct per condition. But an average hides the thing that's actually interesting. Two people — or the same person on two different days — can post the same average reaction time for opposite reasons: one is fast because they're genuinely quick at gathering evidence for a decision, the other is fast because they're willing to commit to a shakier answer. A plain average can't tell those apart; it only sees the outcome, not the process behind it.

That's what a drift-diffusion model (DDM) is for. Instead of treating each trial as a single RT number, it models a decision as noisy evidence building up over time toward one of two thresholds, and fits two separate quantities from the full shape of your reaction times and errors: how fast evidence accumulates (drift rate, roughly "processing speed") and how much evidence you require before committing to an answer (decision threshold, roughly "caution"). Two conditions that look identical on RT alone can turn out to differ cleanly on one of these and not the other.

DriftLab exists to run that idea end to end, in the open: play a battery of classic cognitive tasks in the browser (no install, no lab equipment, no EEG), export your own data, and feed it into a reproducible Python pipeline that fits a DDM to it — checking whether well-known effects, starting with the classic Stroop interference effect, show up as a difference in drift rate, threshold, both, or neither, and whether that matches what's been published.

You don't need to run any code to get the idea — the rest of this README is for anyone who wants to try the tasks or run the analysis themselves.

Play the demo

Browsers block fetch() on file:// for CORS reasons, so serve the repo over a local static server instead of opening index.html directly:

python -m http.server 8000

Then open http://localhost:8000/ and pick a task.

Run the analysis pipeline

python -m venv venv
venv\Scripts\activate          # Windows
# source venv/bin/activate     # macOS/Linux
pip install -r analysis/requirements.txt
pip install -e analysis        # installs driftlab_analysis in editable mode

Run the tests:

pytest analysis/tests/

Run the notebook:

jupyter notebook analysis/notebooks/01_stroop_basic_analysis.ipynb

It uses real data from analysis/data/raw/ if any is present (exported from the browser tasks), and falls back to a versioned synthetic sample otherwise — so it runs out of the box without playing the task first.

Architecture

Each task lives in its own self-contained folder under experiments/, using a shared data export module (shared/js/data-export.js) so any task's CSV export can be loaded by the analysis pipeline without task-specific code. See each task's own README.md for what it measures and its parameters.

License

MIT — see LICENSE.