batter 0.10.2

Deterministic Python research engine for prediction-market evidence-backed backtests

batter

Built for Pancake — strategy results at https://usepancake.com/<handle>/<strategy_slug>/v/<version_n> use this engine to verify strategy math.

batter is a deterministic Python research engine for prediction-market evidence-backed backtests. Given a backtest spec and an EvidenceDataset, it produces a canonical result_hash — identical bytes across ubuntu / macos / windows on Python 3.12+ — enabling reproducible research and auditability of strategy claims. Engine 0.4 adds Monte Carlo bootstrap confidence intervals and a sign-permutation Sharpe test so credibility signals travel with every result.

The PyPI package is batter; the Python module is pancake_engine (sklearn-style rename: pip install batter then import pancake_engine).

What is this for?

Pancake is a prediction-market research platform. batter is the math layer, extracted as a standalone package so the formulas can be verified independently of the platform.

When a strategy backtest runs on Pancake, it runs through batter. The platform stores the result_hash; anyone can reproduce that hash locally by running the same spec and dataset through pip install batter. This is the platform connection: batter has origin in Pancake but is genuinely independently usable by anyone doing prediction-market research.

Install

pip install batter

Quickstart

import json
from pancake_engine import run_backtest, BacktestSpec, EvidenceDataset, BacktestConfig

spec    = BacktestSpec(**json.load(open("spec.json")))
dataset = EvidenceDataset(**json.load(open("dataset.json")))
config  = BacktestConfig()

result = run_backtest(spec, dataset, config)
print(result.result_hash)   # deterministic SHA-256 over canonical JSON
print(result.metrics.sharpe)
print(result.bootstrap_ci)  # 95% CI on cagr / sharpe / sortino (0.4+)

Determinism

The same (spec, dataset, config) produces the same result_hash across ubuntu / macos / windows on Python 3.12+. Verification method, fixture set, and numeric bounds are documented in docs/math-audit-0.4.md §"Verification verdict".

Supported Python versions: 3.12 and 3.13. Python 3.11 is permanently out of scope — sum() semantics changed in 3.12 (compensated float accumulation), causing the bootstrap CI values to differ by 1 ULP and producing a different result_hash. No code change can reconcile this without reverse-engineering 3.12's exact internal accumulation path. See docs/py311-investigation-2026-05-27.md for the full root-cause analysis and docs/math-audit-0.4.md §"Known scope qualifier — Python 3.11" for the audit entry.

Cross-platform

CI enforces a 6-cell matrix (ubuntu-latest + macos-latest + windows-latest) × (Python 3.12 + 3.13). See the badge above.

Cite batter

If you use batter in academic or independent research, please cite:

@software{mustopo2026batter,
  author       = {Mustopo, Michael},
  title        = {batter: Deterministic Python research engine for
                  prediction-market evidence-backed backtests},
  year         = {2026},
  version      = {0.10.1},
  url          = {https://usepancake.com/engine},
  repository   = {https://github.com/usepancake/batter},
  license      = {Apache-2.0},
  note         = {The math layer of usepancake.com. Produces canonical
                  SHA-256 result hashes reproducible across Ubuntu,
                  macOS, and Windows on Python 3.12+.}
}

Verify any result

No trusted party required. Given a self-contained bundle (spec + inline dataset + expected hash), batter verify re-runs the engine locally and checks that the computed result_hash matches the declared one. Anyone can audit a Pancake result independently.

pip install batter

# local bundle file
batter verify --bundle result-bundle.json

# or fetch the replay bundle for a result directly
batter verify --url https://www.usepancake.com/api/replay-bundle/<short_id>

Exit codes: 0 verified · 1 hash or dataset integrity mismatch · 2 input/validation error · 3 unverifiable (pointer dataset — rows not inline, license-gated)

JSON output (stdout):

{
  "verified": true,
  "expected": "<sha256>",
  "computed": "<sha256>",
  "engine_version": "0.8.1",
  "num_trades": 42
}

Bundle shapes accepted:

• regen-style (what examples/*/regen.py produces): {spec, dataset, config?, expected_result_hash: "<sha256>"}
• fixture-style: {spec, dataset, config?, expected: {result_hash: "<sha256>", ...}}

The dataset must carry storage_mode: "inline" with rows present. Pointer datasets (rows held under license) print a clear message and exit 3.

batter verify also checks dataset integrity first: it recomputes rows_sha256 and schema_sha256 over the bundle's actual bytes and compares them to the declared values. A tampered bundle is caught before the engine even runs.

If the bundle declares an engine_version that differs from the installed version, a warning is printed — result_hash values are only comparable under the same ENGINE_VERSION.

License

Apache-2.0 — Copyright 2026 Michael Mustopo

See also

• usepancake.com — the prediction-market research platform
• usepancake.com/engine — batter's platform page (methodology, JSON-LD, citation)
• usepancake.com/methodology — platform methodology