Crypto-style market making in 60 seconds: the flow is anonymous, unsegmented, and some of it knows exactly what the asset is worth. No payment for order flow, no flow labeling, no way to buy protection — you quote a two-sided market against a hidden jumping value, and toxicity is something you infer from the tape (and from the silences) or pay for. Glosten–Milgrom under the hood; your PnL decomposes exactly into spread captured + adverse selection + inventory cost, and the post-round replay shows you every trader who picked you off — and every one who looked at your spread and walked.
(Stylized render of the replay screen from a real L4 round — regenerate with python docs/render_replay_gif.py, or replace with a screen recording.)
- Pick a level (α = share of informed flow, 10%→50%); drag your bid/ask lines — or nudge with W/S (ask) and P/L (bid), Shift = ±5.
- Tight spreads attract fills; stale quotes attract informed fills. Quiet markets mean no customers — or customers who refused your price. You can't tell which.
- At the close your book is marked at the true value; watch the replay to see who knew what. First visit gets a 30-second tutorial (replayable via "?" on the level screen).
Levels were calibrated by a playability gate: a Bayesian dealer (reads fills and silences) must beat a naive fixed-spread dealer by >30% mean PnL over 30 paired seeded rounds. 8 of 72 regimes passed (grid); the shipped one (mean ticks/round):
| Level | α | naive bot | Bayesian bot | edge |
|---|---|---|---|---|
| 1 | 0.10 | 144.8 | 114.5 | 0.79× — naive play genuinely wins |
| 2 | 0.20 | 85.9 | 86.8 | 1.01× — parity |
| 3 | 0.30 | 69.4 | 93.7 | 1.35× |
| 4 | 0.40 | 35.1 | 61.1 | 1.74× |
| 5 | 0.50 | 18.2 | 55.6 | 3.05× |
That 0.79× is the design: when flow is harmless, zero-edge "optimal" quoting loses to a wide dumb spread. By level 5, spread income alone cannot survive the toxicity. The arc is the game.
- writeup/writeup.md — the 3-minute version: why crypto MM is a GM problem, calibration results, mechanism, and an n=3 human-vs-bot self-experiment (the human beat the bot twice — and got annihilated once, −872 vs +60, by never repricing after jumps).
- DESIGN.md — the full spec (source of truth): model, exact accounting identity, censoring rationale, golden-vector schema.
- notebooks/human_vs_bot.ipynb — drop your own session exports into
sessions/and compare yourself to the Bayesian dealer on the identical stream.
| Path | What |
|---|---|
| sim/ | Python simulator — canonical engine, Bayesian bots, playability gate (236 tests) |
| web/ | React + TS + Vite game; TS engine verified against the same frozen vectors (25 tests) |
| vectors/ | 12 frozen golden test vectors consumed by both engines, exact integer equality |
| sessions/ | Session exports analyzed by the notebook (add yours) |
| writeup/ · notebooks/ | Research writeup · gate grid + human-vs-bot analysis |
cd sim && python -m pytest # Python engine: 236 tests
cd web && npm install && npm test # TS engine: conformance vs ../vectors
npm run dev # play locallyMIT — see LICENSE. Topics: market-making · glosten-milgrom · adverse-selection · market-microstructure · trading-game