All projects

MetaMatching

Used in a published dissertation

A web platform I built for a friend’s PhD dissertation in behavioral science. It runs a real-time, three-player decision game, follows the study’s experimental design to the round, and exports every round for analysis.

What it does

Three participants play together in a live session. Each round, every player picks a row on a grid; the payouts reward individual choices or a shared group outcome, which is what the study measures. A session runs 9 blocks of 21 rounds.

A moderator creates a room, shares an invite code, and starts the run, and can fill empty seats with bots for testing. When a session ends, the moderator downloads a CSV with one row per round. No off-the-shelf survey tool could run a synchronous multiplayer design like this, which is why I built it from scratch, working with my friend on the experiment.

How it’s built and run

Stack
Node.js with Express 4 and Socket.IO 4, EJS templates, and MongoDB. Sessions are shared between HTTP and sockets.
Hosting
A systemd service running as its own user, with filesystem and privilege hardening and automatic restarts, behind nginx and Cloudflare. Installed and updated by script.
Data
Round data stays out of the database: it lives in memory during a session and is written to CSV and JSON at the end. Downloads are limited to the room’s moderator and admins, and participant files are kept out of version control.
Access
Invite codes for participants. Accounts use salted scrypt password hashes.
Tests
A schedule test that proves the counterbalancing, and an end-to-end test that drives a live server with bot players and checks the payouts. Both run manually; there is no CI.

Engineering notes

  • Counterbalanced schedules

    Each block contains every combination of player, condition, and incentive exactly once, plus rounds with no incentive, shuffled into a fresh order. The schedule actually used is written into the export, and a test checks the counts across all 189 rounds.

  • Surviving a refresh mid-session

    A study session can’t restart because one participant’s browser reloaded. Players keep their seat and game state across a refresh or dropped connection, and a short grace window keeps brief disconnects from disturbing the other two players.

  • Paying out each round exactly once

    When the last person and a bot lock in at almost the same moment, a guard stops the round being scored twice. Payouts go out in turn order and are capped at what is left in the shared token pool.