Documentation / glossary
The terms,
without the noise.
Definitions for reading the CLI output, benchmark reports and paper.
Lossless compression
Restoration reproduces the original bytes, including whitespace and formatting.
Structural projection
A reversible representation that separates recurring record layout from changing values.
Typed column
A stream of related values encoded using a selected lossless model, such as a dictionary, a delta or a constant.
Pattern dictionary
A store of reusable byte strings. A shared-pattern archive can depend on matching dictionary content.
Standalone archive
A self-contained archive that includes the patterns needed for restoration.
Native engine
PIZD’s Rust matching and entropy-coding path.
Hybrid engine
A PIZD path that uses Zstandard as a backend.
LZ77 match
A reference to previously occurring bytes. Match length and backward distance identify the copied bytes.
Huffman coding
A variable-length entropy code that gives shorter bit strings to more probable symbols.
rANS
A member of the asymmetric numeral systems family of entropy coders. A better coder cannot compensate for a poor data model.
Shannon H0 entropy
The empirical byte-frequency entropy, in bits per byte. It describes an independent-byte model; it is not a lower bound for a compressor that models order and structure.
Context model
A model that conditions a symbol or value on surrounding information, such as prior values or a field identity.
BLAKE3 / SHA-256
Cryptographic hash functions. Their use for data integrity is separate from a release’s build-provenance attestation.
Attestation
A verifiable statement linking a software artifact’s digest to the workflow and source identity that produced it.
Compression ratio
Compressed bytes divided by original bytes. A lower ratio means fewer stored bytes; the direction should always be stated.
Pareto frontier
Measured choices for which no other measured choice is better on every selected objective. A frontier depends on the workload and objectives.
Radin
An explicitly lossy summarization mode. It does not preserve the original data and is excluded from lossless codec comparisons.