Skip to content

NLSS 2.0.0 — Researcher-Led Statistics With Local R

Choose a tag to compare

@docmh docmh released this 16 Sep 16:13
· 2 commits to main since this release

Cite This Version

Hammes, M. (2026). NLSS — Natural Language Statistics Suite (Version 2.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.22800205

This version-specific DOI identifies the automatically archived v2.0.0 source release. The concept DOI identifies NLSS across archived versions. Installable plugin/skill packages are available in the GitHub release assets below.

Install NLSS

Give your local agent the repository link and ask it to install NLSS. Follow the installation guide.

Download one managed package below: nlss-2.0.0-plugin.zip or nlss-2.0.0-skill.zip, plus release.json and SHA256SUMS. Verify the matching SHA-256 entries. GitHub's automatically generated source archives are for source development.

R (Rscript) is required; analysis packages are checked on demand and installed with your approval. Read the breaking changes before starting a new project from older data.

Changes in 2.0.0

Changes relative to the previous published release, 1.0.2z.

Breaking changes and upgrading

  • New project layout. A user-owned folder with nlss-workspace.yml now
    keeps visible working data and the automatic root report_canonical.md,
    with managed run/data/report evidence in .nlss/. The old per-dataset layout
    is not imported, adopted or converted. Preserve old projects unchanged;
    start a new project from an explicitly selected source dataset when needed.
    This does not transfer old analysis history or certify old results.
  • Changed automation entry points. Use project-create for initialization,
    project-inspect for browsing and project-report for authored revisions.
    The old initializer and metaskill activation/finalization workflow are not
    the setup/report route for current projects. Saved runs use request/result
    JSON and Markdown, not a current-project analysis_log.jsonl.
  • Stricter arguments and dependencies. Unknown/duplicate options, malformed
    values and invalid configuration fail explicitly. Requested methods do not
    silently fall back to different inference when required packages are absent.
    SPSS import uses haven; plots require ggplot2. Existing statistical
    families remain available. Review old scripts against operation help.
  • Numerical corrections can change earlier conclusions. In particular,
    review prior repeated-measures/mixed ANOVAs, Mann–Whitney effects, mixed-model
    tests/contrasts, EFA summaries and diagnostic/power calculations under the
    conditions below. Recompute into a new project/run when appropriate and
    compare scientifically; do not overwrite historical evidence. Exact replay
    with old code and recalculation with corrected code are different tasks.

Added

  • Shared saved-request execution and explicit verified replay for statistical
    procedures, with selected input versions, effective settings, seeds, software
    context and artifacts. Replay does not recreate historical R installations.
  • Preserved mice multiple-imputation artifacts and mi-regression: pooled
    coefficients and uncertainty for supported Gaussian identity, binomial
    logit/probit/cloglog and Poisson log models. Single completed or averaged
    columns are not a substitute for pooled inference.
  • Direct project initialization/import from CSV, SAV, RDS, RData or Parquet;
    registered working data remain accessible outside .nlss/.
  • Read-only project inspection and freely authored report revisions linked to
    their selected existing evidence. Report filenames and prose structure are
    researcher choices, not new template requirements.
  • Dataset-free power planning through the ordinary project run route.
  • One portable scientific payload as plugin and standalone skill, plus release
    inventories/checksums and isolated installation/update/removal support.
  • Automatic operation-specific R dependency preflight and a separate approved
    installation route, including recovery when foundational packages are missing.

Changed

  • Shared import semantics keep variable/value labels, source identity and
    user-missing definitions separate from analytical values. CSV interpretation
    can explicitly select encoding, decimal mark, column types and missing values.
  • All procedures use common CLI/configuration and publication boundaries.
    The root protocol is maintained automatically; figures remain run-local and
    links are rebased. Data changes preserve before/after versions without a
    parallel journal or a permanent family of duplicate backups.
  • Scientific answers and reports adapt to the requested scope. Full reports
    remain semantic syntheses, with required agentic source appraisal when
    literature research is called for; no compulsory module-specific prose.
  • README and the installation guide provide a researcher-first introduction
    and an explicit repository-link-to-release path. Source checkouts are
    distinguished from managed release archives.

Fixed — changes that can affect results or their interpretation

  • Imports: SPSS user-missing values and labels no longer take inconsistent
    analysis paths merely because data arrive as SAV, RDS/RData or Parquet.
    Lossy conversions and ambiguous source collisions fail rather than silently
    reusing a same-named dataset. See the import contract.
  • ANOVA: repeated-measures subject IDs now define factor error strata;
    numeric IDs previously produced incorrect tests/df. Sphericity uses the
    within-occasion contrast space and appropriate residual covariance. Corrected
    inference, requested Tukey confidence levels and vector-valued contrast
    p-adjustments are preserved. The retained repeated-measures adapter is
    sequential, not general Type-II/III repeated-measures support.
    ANOVA details
  • Rank tests: Mann–Whitney no longer subtracts the rank offset twice;
    U and associated rank effects can differ materially from old output.
    The previously named epsilon_sq formula is identified as eta_H_sq,
    with the old alias's meaning disclosed. Location/pseudomedian intervals are
    distinguished from effect-size and median-difference intervals.
    Rank-test details
  • Linear mixed models: requested sums-of-squares type and Satterthwaite/
    Kenward–Roger inference reach the actual tests and marginal-mean comparisons;
    adjusted contrast p-values are no longer recycled from a single scalar.
    Optimizer limits/status and unavailable diagnostics are reported honestly.
    Mixed-model details
  • Correlations and t-tests: partial-correlation inference accounts for
    the effective control rank; one-sided intervals follow the requested
    alternative. Group identity, paired direction and analyzed cases are retained.
    Correlations, t-tests
  • Reliability and scale summaries: requested ICC agreement/consistency
    definitions remain distinct; grouped case/missing counts and categorical
    order are preserved rather than inferred from display labels.
    Reliability, scale analysis
  • EFA: oblique communalities incorporate factor correlations; factor-model
    explained variance no longer substitutes PCA quantities. Selected rotations
    and ordered category codes reach the intended package calculations.
    EFA details
  • SEM: reported groups retain lavaan's actual group identity; exact option
    matching prevents collisions such as seed/se. Requested CI/bootstrap
    settings and diagnostic availability are distinguished from effective output.
    SEM details
  • Diagnostics: regression curvature checks no longer treat the built-in
    orthogonality of OLS residuals as a linearity test. Sphericity tests the
    relevant contrast covariance. Skipped/unavailable checks are not passes.
    Assumption checks
  • Power: one-sided effect direction and the specified SEM RMSEA null and
    alternative affect calculations; pilot groups/design rank and rounded-sample
    attained power are recorded. Power details
  • Data handling and presentation: grouped counts, category identity,
    transformation/recode provenance and original imputation columns are preserved;
    non-estimable requested results and failed runs are explicitly identified.
  • Installed operation paths containing spaces now use the shared base-R
    launcher, including replay subprocesses, without a separate argument parser.

Removed

  • Duplicated per-module infrastructure fallbacks and silent substitute import/
    graphics routes; no statistical family was removed to simplify packaging.
  • Mandatory metaskill activation/finalization ceremony for ordinary research
    and report delivery, and the current-project parallel JSONL/backup layout.

Support and known limits

Linux workflows were exercised in Codex CLI, Claude Code, Copilot, Vibe and
Antigravity; the latter remains qualified by documented scope/context problems.
Codex IDE and App had limited activation checks. Native Windows/macOS are
prepared, not live verified. Some historical test runners still assume retired
project layouts; no fully green legacy suite or universal model correctness is
claimed. See the exact support matrix.
Lambda-Star and LaTeX output are not part of this release.