bitrep (Python)

Exact, order-invariant, bit-identical floating-point reductions and CRDTs — a thin Python binding over the Rust bitrep engine, so every result is the same across any machine, any merge order, any sharding.

pip install bitrep

import bitrep

# Exact sum — naive float gives 0.0; this gives 1.0, and it's order-invariant.
s = bitrep.SumF64()
for x in (1e16, 1.0, -1e16):
    s.add(x)
print(s.value())            # 1.0

# Exactly-rounded statistics
m = bitrep.MomentsF64()
for x in (2, 4, 4, 4, 5, 5, 7, 9):
    m.add(x)
print(m.mean(), m.variance(), m.count())   # 5.0 4.0 8

# Receipt: same multiset -> same 32-byte hash, regardless of order.
print(s.state_hash().hex())

# CRDT map that converges without a coordinator
a = bitrep.SumMap(); a.add("k", 1.5); a.add("k", 2.5)
b = bitrep.SumMap(); b.add("k", 2.5); b.add("k", 1.5)
a.merge(b)                   # idempotent, order-invariant

What's included

Exact accumulators (SumF64, SumF32, FastSumF64), exact dot product (DotF64, dot), convergent statistics (MomentsF64, Moments4F64 with skewness/kurtosis, CovF64 regression, WeightedMomentsF64, PnMomentsF64 retractable, CovMatrixF64 multiple regression — with the v0.5 exact tier: regression_exact() for correctly rounded coefficients and sub() for exact downdating/unlearning, plus SumF64.try_unmerge), HistogramF64 (exact counts + honest quantile bounds), ExtremaF64, signed receipts (state_hash), and the CRDT layer (SumMap, MomentsMap, ReplicatedSum, DeltasSum).

The convergence laws of the core are machine-checked in Lean 4. See the main repository for proofs, the Rust crate, and the paper.

License: MIT OR Apache-2.0.