What is new in Rubrkit.

Drift watch re-audits one pinned version on a schedule (daily, weekly, or monthly) and reports when its score leaves the range measured at calibration. The artifact never changes, so a move outside that range means the model or the rubric shifted, not your prompt. Create one from the Drift panel or POST /artifact-bundles/{id}/drift-monitors.

A Check now control and POST /drift-monitors/{id}/check to run a check immediately instead of waiting for the schedule. The first check on a new monitor measures the version's own noise band before it can report anything as drift.

Send real production results back with POST /artifact-bundles/{id}/outcomes (or the rubrkit_record_outcome MCP tool) to unlock two signals the offline rubric cannot give: input drift, which compares recent traffic against your reference window, and validity drift, which measures whether the rubric score still tracks the real outcome. Both stay blank until there is enough data to answer them honestly.

The background job drainer now runs every ten minutes instead of every minute, so a queued job may take a few minutes longer to start but the database is no longer held awake around the clock.

An advanced comparison mode that audits two versions of a file several times each and reports whether the difference is larger than the evaluator's own run-to-run noise — better, not just different. Every rubric dimension gets a 95% confidence interval, corrected across dimensions, so a bar that crosses zero is not a real change. Available in the Compare panel or by POSTing an eval run with mode advanced, repeats, baselineVersionId, and candidateVersionId.

Golden sets: pin the real inputs a bundle must keep handling correctly, so the last thing that broke can never break silently again. Manage them from the Golden set panel, over REST at /artifact-bundles/{id}/golden-cases, or with the MCP tools rubrkit_add_golden_case, rubrkit_list_golden_cases, and rubrkit_retire_golden_case.

Proof reports exported from an advanced comparison now carry the statistical verdict — the effect size, confidence interval, and which dimensions moved — instead of a bare pair of scores.

Advanced comparisons measure rubric-judged quality across repeated audits, not real-user outcomes, and the interface says so wherever a verdict appears. A single offline run is still a first gate, never a claim of statistical significance about your users.

Permanently delete an archived artifact bundle from the dashboard or the API (DELETE /artifact-bundles/{id}/permanent). A permanent delete removes the bundle and all of its files, version history, and run history, and only works once a bundle has been archived.

A Load more control on the Artifact Bundles dashboard, backed by a new offset parameter and nextOffset response field on the list API, so long bundle lists page in instead of silently capping.

New MCP tools rubrkit_delete_artifact_bundle (archive) and rubrkit_hard_delete_artifact_bundle (permanent), plus a force option on rubrkit_start_audit that bypasses the audit cache — the same as the CLI's --no-cache flag.

Artifact type auto-detection now takes precedence over a format-level type sent by a client, so a SKILL.md uploaded by the CLI is detected as a skill instead of plain markdown. Re-upload an existing bundle to pick up the corrected type.

A curated Skill Library at /skills — open, install-ready skills for frontend and backend development that you can browse, copy, or pull from GitHub at github.com/Rubrkit/development-skills with npx skills add.

The published skills are graded by Rubrkit in CI on every push: a skill that scores below 75 fails the build, so the open collection is held to the same bar as everything else.

The CLI test --remote now syncs into an existing artifact bundle — updating changed files, adding new ones, and removing stale ones — so a pinned CI bundle no longer conflicts on repeat runs (rubrkit 0.4.0).

A new Account section in the dashboard where you can export all of your data as a single JSON file — profile, bundles and file metadata, audit/eval/proof history, usage, credits, and billing identifiers.

Permanent account deletion from the Account page: a type-to-confirm step removes your profile, bundles, files, API keys, and run history, and cancels any active subscription. Financial records are anonymized and retained as required by law.

Account export and deletion require a signed-in session (login token); API keys cannot trigger either action.

Audit results are cached per account by a content hash of their inputs, so re-running an audit on unchanged files returns the identical result instantly and charges no credits — repeated CI runs are reproducible.

A --no-cache CLI flag (and force API option) to bypass the cache and force a fresh audit run.

The audit API and CLI now report whether a result was served from cache.

Public API reference at /docs/api, generated from the canonical OpenAPI contract so it can never drift.

CLI reference at /docs/cli for the rubrkit npm package: pull, validate, test, audit, eval, and report.

MCP server docs at /docs/mcp with the full tool catalog and connection config.

Collapsed the separate doc links into a single Docs menu in the header.

Audit jobs no longer stall when a rubric dimension returns zero findings.

Proof report exports retry transient blob-storage failures instead of failing the job.

Tightened API-key scope checks on file-write endpoints so a read-only key can never mutate a bundle.

Eval runs that score rubric-backed test cases against a rewrite.

Exportable proof reports that show an instruction held up to its eval.

The credits ledger now previews the cost of a run before you start it.

Restoring a file version now writes a new version instead of overwriting history.

Corrected an off-by-one in bundle version-history pagination.

Full version history for artifact bundles and individual files.

Restore any prior bundle or file version from the dashboard or API.

File uploads are content-addressed, so identical content no longer creates duplicate versions.

The rubr_flow validator now accepts the optional ON and STATE blocks.

Conversion job progress events now stream in order.

Convert a loose instruction into a bounded rubr_flow procedure through the API.

A dedicated rubr_flow page covering the anatomy, how to write one, and worked specimens.

A before-and-after examples gallery showing weak instructions graded and rewritten.

The dashboard bundle list now shows each bundle’s last-audited score.

Contact and newsletter forms now confirm submission with a snackbar instead of a silent reset.

Login tokens are verified before an identity is mapped to a Neon-backed user.

Expired or revoked API keys now return a clear unauthorized error instead of a generic failure.

Scoped API keys with explicit read and write scopes per resource.

Artifact-bundle, file, and audit endpoints under /api/v1.

Async job polling for long-running audits and conversions.