Patina Multilingual Writing Editor

The workflows use read-only permissions by default, while dataset, npm, and GHCR publication require explicit Boolean inputs or constrained tag/branch conditions, protected environments, and source-commit checks. Browser tests impose a same-origin boundary, reject below-floor output, and exercise cancellation. Private text, keys, and tokens are explicitly within the security policy, and the README says BYOK keys are neither stored nor logged; however, the supplied files omit the relevant implementation and a complete privacy or retention policy, so data-flow transparency and sensitive-data handling do not earn full marks. Runtime dependencies are minimal and some release actions are commit-pinned, but several actions use moving major-version tags and no lockfile, vulnerability scan, or dependency-update policy is shown. Rejected CLI candidates are not written and source files remain unchanged, offering partial recovery, but there is no complete user-facing rollback procedure. MIT copyright, NOTICE references, inspirations, and repository ownership are clearly attributed.

The README, package metadata, workflows, and tests are broadly aligned on version 8.10.0, Node requirements, commands, and verification floors. Release automation checks metadata, tests, benchmark drift, and artifact identity. A visible inconsistency remains: the version badge's alt text says 8.8.0 while its displayed value is 8.10.0. The registry is also stated to remain at 8.9.0; this is explained, but still complicates consistency. Dependencies and backend choices are documented and offline scoring is available, though online rewriting, the hosted API, external CLIs, and browser use rely on services whose availability guarantees are not established here. Browser and CLI tests cover unavailable service responses, semantic drift, cancellation, timeout, and rejected writes with visible, specific failure states, supporting full credit for failure messaging.

The source addresses Korean, English, Chinese, and Japanese and distinguishes browser, Agent skill, CLI, CI, batch, blog, and document-type scenarios. Document Type, Persona, and Register are explicitly independent. Boundaries are unusually clear: this is an editing aid rather than authorship proof or detector evasion, scores have false positives and negatives, and omitted axes preserve source characteristics. Commands and modes are precise, but the suggested Agent installation instruction fetches from the mutable main branch rather than a pinned revision, reducing trigger and installation precision. Node requirements, configuration examples, offline operation, multiple backends, and narrow-viewport browser tests provide strong environment-fit evidence.

The README has coherent navigation to installation, the CLI contract, architecture, authentication, cookbook, examples, pattern catalog, research, FAQ, contribution guidance, and changelog. Quick Start covers the no-install browser, Agent skill, and Node CLI, while commands, configuration, and before/after examples are concrete and the example is explicitly labeled synthetic. Known limitations, retired clients, and the non-authorship purpose are disclosed. LICENSE and package metadata consistently identify MIT. Version and release-channel handling are described and a changelog is referenced, but its contents are absent and the badge alt text is stale, so version governance is not fully demonstrated. SECURITY.md identifies main as the maintained branch and gives private reporting and issue-routing paths, but the unverified publisher is not a deduction by itself; the absence of a named responsible maintainer or response commitments keeps maintenance responsibility below full marks.

Outputs support diff, audit, score, verify, structured receipts, and gate states. Tests show that approved results are copyable while drifted, interrupted, and cancelled results are not, and that rejected CLI candidates are never written, providing strong output-usability evidence. Four-language pattern packs, independent style axes, and meaning safeguards plausibly add value beyond generic rewriting, but the complete pattern catalog, calibration artifacts, and cited study are not supplied, so the claimed 189 patterns and measured efficacy cannot be fully evaluated. Browser, offline, BYOK, Pro, and local CLI routes offer useful choices, but pricing, latency, resource use, and backend cost comparisons are absent, limiting the cost-benefit score.

The README maps claims about pattern counts, calibration intervals, research, benchmarks, and architecture to named repository documents. Workflows check benchmark drift, source commits, artifact identity, and release versions, while the supplied browser and CLI tests corroborate gating, cancellation, credential transport form, and rejected-write behavior. However, most referenced research, catalogs, authentication documentation, implementation files, and generated reports are not included, so central effectiveness statistics are traceable only to filenames and cannot be checked from this evidence set. The source carefully distinguishes synthetic examples, historical recordings, live runs, and measured scores, and states that scores are editing signals rather than authorship proof, earning full credit for fact/inference separation.