In one line: A tiny exact benchmark where freezing an oracle’s principal-line continuation can lose to blunders.
Canonical research site: https://psdg.pages.dev — motivation, definitions, empirical snapshot (blunder vs static rows), FAQ, technical report summary, related work, and worked examples. Game theory, ML evaluation, and alignment are different routes into the same benchmark under the same oracle on that site; this repo is the clone-and-run artifact.
This repo holds public, reproducible artifacts in Python only (no JavaScript solver in-tree): the reference solver, benchmark JSON, and scripts to generate, verify, and experiment—without a private monorepo.
It is intentionally different from the site’s long-form pages—for example Technical report (summary). Narrative and caveats live on psdg.pages.dev; numbers and protocols you reproduce start here.
psdg/
├── README.md
├── LICENSE
├── RULES.md # canonical rules v1.13 (Markdown); site mirrors with HTML nav
├── solvers/python/ # solver.py, oracle.py, helpers, small blunder JSON fixtures
└── benchmark/ # *.json suites, Python scripts, output/ logs
- Rules: RULES.md — same v1.13 text as psdg.pages.dev/rules.
- Benchmarks: benchmark/README.md for script roles and quick commands.
From the repository root:
python3 solvers/python/solver.py -r -s 42
python3 benchmark/verify_benchmark.py benchmark/benchmark_4d.jsonSee LICENSE (MIT).