LoginSign Up
← Back to all sparks
W

writeAlizer

INFRA · APIS
Velocity0.0

Generate Predicted Writing Quality Scores

github.com/shmercer/writealizer ↗

Six months of releases and not one of them touched the scoring models

writing-assessmentnlp-featuresmodel-artifactscran-complianceoffline-modereproducibility
◆Current state
writeAlizer generates predicted writing-quality scores from features produced by Coh-Metrix, ReaderBench and GAMET, downloading its trained scoring models on demand. Every release in this window — nine of them between September 2025 and February 2026 — is about that download path rather than the scoring: classed error conditions, checksum verification, an offline mode, a mockable artifact directory, and dependency reporting for the model families a user actually invokes.
◆Where it's heading
The package is being made safe to distribute. CRAN's policy on packages that reach the internet drove the first wave — graceful failure, tests that preflight their URLs and skip, examples seeded from a local mock model — and 1.7.0 turned the accumulated fixes into structure with named error classes for each failure mode. Only 1.7.2 adds anything a user would ask for: filename handling for Coh-Metrix and GAMET outputs that arrive as paths.
◆Prediction
With the artifact registry hardened and documented, the pressure that produced nine releases in six months should ease, and attention can return to the models themselves — the vignette on scoring-model development added in 1.7.2 hints at that. Nothing here promises new models.

◆Recent moves

  1. 6mo ago

    One example rewrapped to silence a CRAN check note

    A single example moved from a check-running wrapper to one that never runs, purely to keep CRAN's checks quiet. The smallest possible release.

    View source ↗
  2. 8mo ago

    Filename stems recovered from Coh-Metrix and GAMET paths

    The only release in the window that changes what a user can feed the package: keep_stem_before_txt() handles feature files whose identifiers arrive as full paths, wired into all three import functions. A vignette on how the scoring models were built ships alongside it.

    View source ↗
  3. 10mo ago

    Offline example guard, declared as no API change

    predict_quality()'s example seeds a local model and restores prior options so it runs deterministically offline. The note states plainly there are no user-facing API changes.

    View source ↗
  4. 10mo ago

    Named error classes for every model-download failure mode

    The release where six months of scattered fixes become structure: missing registry, unknown model key, failed download, submodel count mismatch each get their own condition class, and the loader gains hard checksum enforcement, file URL support and a mockable artifact directory.

    View source ↗
  5. 10mo ago

    Network failures degrade gracefully under CRAN policy

    Remote artifact fetches stop surfacing raw download errors, tests preflight their exact URLs and skip when unreachable, and examples seed mock artifacts. The start of the offline-determinism work that 1.7.0 completes.

    View source ↗
  6. 11mo ago

    wa_seed_example_models() exported and documented

    A previously internal helper is exported so examples stop reaching into the package namespace, per CRAN guidance. Documentation work rather than capability.

    View source ↗