weiss-sim 1.1.0

Deterministic Weiss Schwarz simulator with a Rust core and Python bindings.

Weiss Schwarz Simulator

Deterministic, batched Weiss Schwarz simulation for reinforcement learning and engine research.

What you get

• Rust engine (weiss_core) with deterministic advance-until-decision stepping
• PyO3 bindings (weiss_py) and Python API (python/weiss_sim) for batched training/eval loops
• Stable observation/action contracts (OBS_LEN=378, ACTION_SPACE_SIZE=527, SPEC_HASH=8590000130)
• RL-oriented fast paths for packed legal ids, no-metadata layouts, fused sampled log-prob stepping, and optional legal-action context tensors
• Replay and fingerprint surfaces for drift detection and reproducibility

Release status

1.1.0 is the stabilized research release line. Public compatibility boundaries are documented explicitly, and any future encoding, replay, WSDB, or action-space changes should update the corresponding version constants and docs in the same change.

5-minute start

Option A: install from PyPI

python -m pip install -U weiss-sim numpy

Option B: local build from source

python -m pip install -U maturin numpy
python -m maturin develop --release --manifest-path weiss_py/Cargo.toml

Run the commands in the Python environment where you want weiss_sim installed. If you are using a virtualenv, activate it first.

Option C: contributor setup (dev extras + local extension)

rustup component add rustfmt clippy
python -m venv .venv
# activate the virtualenv before the next commands
# PowerShell: .\.venv\Scripts\Activate.ps1
# Bash / zsh: source .venv/bin/activate
python -m pip install -U pip
python -m pip install -e ".[dev]"
python -m maturin develop --release --manifest-path weiss_py/Cargo.toml

This installs .[dev] extras (ruff, pytest, maturin, pip-audit) and builds the local extension in-place.

Virtualenv activation examples:

• PowerShell: .\.venv\Scripts\Activate.ps1
• Bash / zsh: source .venv/bin/activate

If you prefer the helper script, scripts/setup_dev_env.sh runs the same install flow in Bash-compatible shells.

Minimal high-level loop

import numpy as np
import weiss_sim

with weiss_sim.fast(num_envs=32, seed=0) as sim:
    reset = sim.reset()
    actions = reset.legal.sample_uniform(seed=123)
    step = sim.step(actions)

Minimal low-level loop

import numpy as np
import weiss_sim

legal_deck = (list(range(1, 14)) * 4)[:50]

pool, buf = weiss_sim.make_pool(
    mode="train",
    num_envs=32,
    deck_lists=[legal_deck, legal_deck],
    deck_ids=[1, 2],
    seed=0,
    layout="i16_legal_ids_nometa",
)
out = buf.reset()
actions = np.full(pool.envs_len, weiss_sim.PASS_ACTION_ID, dtype=np.uint32)
out = buf.step(actions)

Throughput-sensitive policy-gradient loops should prefer the fused sampled-logp helper when they need both sampled actions and behavior log-probabilities:

logits = np.zeros((pool.envs_len, weiss_sim.ACTION_SPACE_SIZE), dtype=np.float32)
seeds = np.arange(pool.envs_len, dtype=np.uint64)
step, actions, logp = buf.step_sample_from_logits_with_logp(logits, seeds)

Deck authoring flow (Python)

import weiss_sim

builder = weiss_sim.cards.builder(initial="starter_deck_ws02_v1")
report = builder.validate(rules_profile="approx", card_pool="all")
if report.ok:
    deck_ids = builder.build(rules_profile="approx", card_pool="all")

Bundled presets are the four release decklists: starter_deck_ws02_v1, control_deck_jj_s66_v1, main_deck_5hy_yotsuba_v1, and aggro_deck_5hy_nino_v1. Inspect the installed package with weiss_sim.cards.presets(). These presets require rules_profile="approx" because they include partially parsed card abilities.

Architecture at a glance

flowchart LR
  A["Python API\npython/weiss_sim"] --> B["PyO3 bindings\nweiss_py"]
  B --> C["Engine core\nweiss_core"]
  C --> D["Deterministic outputs\nobs/masks-or-ids/reward/status"]
  C --> E["Replay + fingerprint\nrepro & drift debugging"]

Documentation map

Start in docs/README.md. The docs are intentionally compact:

• docs/quickstart.md: install, first loop, decks, troubleshooting, and local checks
• docs/python_api.md: high-level API, low-level buffers, layouts, and RL helpers
• docs/rl_contract.md: step semantics, rewards, output schema, legal payloads, and compatibility constants
• docs/architecture.md: Rust/Python layers, rules/scraper boundaries, replay determinism, and invariants
• docs/performance_benchmarks.md: benchmark commands, current baselines, perf gates, and RL hot paths
• CONTRIBUTING.md: validation, release flow, docs rules, and PR checklist

Repository layout

• weiss_core/: Rust engine and deterministic rule runtime
• weiss_py/: PyO3 extension layer
• python/weiss_sim/: high-level and low-level Python interfaces
• python/tests/: Python API/contract tests
• scripts/: CI parity, coverage, perf, and docs checks
• docs/: user + contributor documentation hub
• .github/: CI, release, issue, and pull request templates

Local quality checks

Full local CI parity:

bash scripts/run_local_ci_parity.sh

On Windows, run the underlying commands directly from PowerShell:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace --all-features
python -m ruff format --check python scraper scripts
python -m ruff check python scraper scripts
python scripts/check_docs_links.py
python scripts/check_docs_constants.py
python scripts/gen_docs_snippets.py --check
python -m pytest -q python/tests

Skip benchmark gate during iteration:

SKIP_BENCHMARKS=1 bash scripts/run_local_ci_parity.sh

Docs-only checks:

python scripts/check_docs_links.py
python scripts/check_docs_constants.py
python scripts/gen_docs_snippets.py --check

Benchmark snapshot (main)

Last updated: 2026-05-12 12:11 UTC

Benchmark	Time
rust/advance_until_decision	21272 ns/iter
rust/step_batch_64	9581 ns/iter
rust/reset_batch_256	512004 ns/iter
rust/step_batch_fast_256_priority_off	39563 ns/iter
rust/step_batch_fast_256_priority_on	42891 ns/iter
rust/legal_actions	4 ns/iter
rust/legal_actions_forced	4 ns/iter
rust/on_reverse_decision_frequency_on	990 ns/iter
rust/on_reverse_decision_frequency_off	1006 ns/iter
rust/observation_encode	80 ns/iter
rust/observation_encode_forced	79 ns/iter
rust/mask_construction	180 ns/iter
python/reset_into	110.0 us/reset
python/step(mask)	1296605 env-steps/sec
python/step(ids)	1414805 env-steps/sec

Long-form benchmark docs: docs/performance_benchmarks.md

Compatibility policy

Contract constants are explicit compatibility boundaries:

• OBS_ENCODING_VERSION=2
• ACTION_ENCODING_VERSION=1
• POLICY_VERSION=2
• REPLAY_SCHEMA_VERSION=3
• WSDB_SCHEMA_VERSION=2

If encoding/layout semantics change, update code + docs in the same PR:

1. constants/encode implementation
2. docs/rl_contract.md checksum table
3. docs/architecture.md compatibility notes when replay/WSDB/runtime boundaries change

License

MIT OR Apache-2.0