Skip to content

CHANGELOG.md

Latest commit

 

History

History
111 lines (79 loc) · 5.06 KB

CHANGELOG.md

File metadata and controls

111 lines (79 loc) · 5.06 KB
Raw

Changelog

Unreleased

0.4.1 - 2026-08-12

Security

  • Replaced the blind judging packet's dynamic HTML assembly with DOM nodes and textContent, so task, response, alias, and rubric data are never reinterpreted as markup.

Added

  • Added a real Chromium workflow test that scores and advances a generated blind packet.

0.4.0 - 2026-08-12

Added

  • Real Playwright acceptance on desktop and mobile Chromium for keyboard validation, multi-pair review, long-answer scrolling, local history, responsive layout, and key WCAG A/AA checks.
  • CodeQL analysis for Python and JavaScript, npm dependency updates, and GitHub attestations for release distributions and checksums.

Changed

  • Pinned every third-party GitHub Action to an immutable commit while retaining version comments for automated maintenance.
  • Kept all three progress labels and the Study Mode link visible on narrow screens.

Fixed

  • Moved keyboard focus to the first invalid field and associated its error message programmatically.
  • Moved focus predictably between capture, each blind-review pair, results, and history; announced comparison and persistence changes through status regions.
  • Made long answer panes, horizontally scrolling result tables, JSON import, and all primary actions visibly keyboard reachable.
  • Added distinct accessible names for destructive actions, valid progress and summary semantics, and WCAG AA color contrast in the tested application states.

0.3.1 - 2026-08-05

Added

  • PyPI Trusted Publishing through a dedicated GitHub environment and short-lived OIDC credentials.
  • PyPI-safe README images and links, verified project metadata, and a social preview for shared links.
  • Structured bug and use-case issue forms so early adopters can report concrete workflows.

Changed

  • Made the primary installation path python -m pip install frontiertrials.
  • Split release construction, PyPI publication, and GitHub Release publication into least-privilege jobs that reuse the same verified distributions.

0.3.0 - 2026-08-05

Added

  • A zero-config Personal Lab for two-to-four-product blind comparisons in the browser.
  • Browser-local task history, aggregate preference scores, category coverage, price and latency context, JSON backup and restore, and self-contained HTML report export.
  • The frontiertrials open command, which serves the packaged Personal Lab on a loopback-only local server with a closed content-security policy.
  • Packaged web assets, server and CLI tests, and release-wheel smoke coverage for Personal Lab.

Changed

  • Repositioned the project around private AI subscription decisions for personal users, with the original rigorous workflow retained as Study Mode.
  • Rewrote the README and project website around Quick Compare, Personal Benchmark, and Study Mode.
  • Added an explicit Arena comparison and an honest masking-versus-memory limitation.
  • Replaced the hero with a plain-language personal decision workflow.

0.2.0 - 2026-08-05

Added

  • Blind-safe adjudication queues in JSON, CSV, and Markdown for disagreement, low confidence, ballot flags, abstentions, and tie votes.
  • Per-reviewer descriptive diagnostics and leave-one-rater-out ranking sensitivity.
  • Academic print-inspired website, report, reviewer packet, and graphical abstracts.
  • Versioned GitHub Release installation path, release automation, and repository audit.

Changed

  • Identity-bearing analysis now requires the revealed trial state.
  • Reveal now requires the complete assigned-ballot matrix unless an explicit override is recorded.
  • The fictional demo now includes adjudication artifacts and a deterministic seal timestamp.

0.1.0 - 2026-08-04

Added

  • Seven file-native artifact types for tasks, candidates, responses, rubrics, raters, pairings, and ballots.
  • Exact Markdown capture with SHA-256 verification and optional observed latency.
  • Complete-matrix checks, deterministic anonymous aliases, alternating pair orientation, and rater allocation.
  • Self-contained offline judging packets with downloadable JSON ballots.
  • Validated ballot-bundle import and completion tracking.
  • Bradley–Terry ranking with task-clustered bootstrap intervals.
  • Weighted pointwise rubrics, task-category sensitivity, reviewer agreement, position diagnostics, and verbosity association.
  • Structural and blinding audits plus a content-addressed trial seal.
  • A revealed portable report and a 191-artifact, fully fictional demonstration.
  • Eighty-five standard-library tests, cross-platform CI, package smoke tests, and GitHub Pages.