Independent Review

Independent review

The repository includes a small review kit that makes the first reproduction step one command:

git clone https://github.com/ryanpavlicek/pyaegean.git
cd pyaegean
python scripts/reproduce_review.py

It runs on CPython 3.10 through 3.14 with the standard library and the checked-out zero-dependency package source. The command verifies the SHA-256 of every record in the canonical review manifest using its declared byte-exact or canonical-LF mode, runs the project-authored offline benchmark fixture and one ordinary baseline-pipeline sentence, and compares the exact result with the reviewed expectation. When Git is available, it also identifies the commit and rejects a dirty checkout by default.

The verifier does not use the network, download or run the neural model, write bytecode, or create a pyaegean cache. The offline fixture is a regression test, not the published neural benchmark: reproducing the neural measurements needs the separate pinned protocol in Benchmarks.

What a pass establishes

flowchart LR
    A["Clean checkout"] --> B["Evidence hashes"]
    A --> C["Deterministic fixture"]
    B --> D["Exact receipt"]
    C --> D
    D --> E["Human review of claims and limits"]

Loading

A pass establishes that the public records have their reviewed bytes and that the small deterministic result matches. It does not establish that an interpretation is correct, rerun every benchmark, or amount to external peer review. Use --json for the complete receipt. --allow-dirty is for diagnosis only.

Review map

Start here	What it answers
Review kit	One-command instructions, exact manifest, model card, data card, and receipt map
Benchmarks	Measured values, sample sizes, protocols, and comparisons
Methodology	Data, architecture, evaluation, leakage, and review methods
Data and Provenance	Dataset origins, licenses, hashes, and cache boundaries
Limitations	Known evidence, licensing, accuracy, and design boundaries
Validation and Review	What has and has not received outside review

Report a discrepancy

Open the independent-review discrepancy form and include the exact command, version, Git commit or source-archive hash, manifest digest, deterministic-result digest, environment, observed output, and every local modification. Record a mismatch exactly; it is useful evidence.

People interested in maintenance can choose an ownership slice in CONTRIBUTING.md before taking on a larger change.