Where it stands

One command starts the application on a fresh checkout; a setup page creates the first organization and admin; members are invited with roles.

One gateway with OpenAI, Anthropic, and OpenAI-compatible adapters; keys encrypted per organization; every call recorded with its cost or as unpriced; a restricted call to a third-party provider refused before anything leaves. Probing a local model server was not exercised on the checking machine.

Scheduled jobs with progress, plugins discovered and installed per organization, the audit log, and the isolation tests against a real database.

Documents are extracted, split into passages, and queued for review.

Approve, reject, annotate, or correct; bulk approve that still records one decision per document.

Search returns approved content only, each hit with a complete citation; a labeled reviewer mode includes unapproved content and cannot be cited.

The product's own tokens, navigation, and shared components on every page.

Confluence Cloud spaces and public websites sync on a schedule; spreadsheets keep their rows and columns when cited.

A changed source waits for re-review while the approved version stays in use, flagged; a title-only change creates no new version.

Standard search-quality and speed measures per run, with a gated comparison before a change to chunking or embeddings is confirmed.

Small additions to the interface between knowledge and workbench, a record of whether a sync was live or replayed, and the website passages in the evaluation suite, added after the check.

Search naming an approver by identifier, and the other small defects the check listed, are fixed in 0f.

Findings with citations, a conflict recorded as an open question with both passages, gaps with coverage counts, and requirements each with a citation or an assumption.

Hand edits as new versions; a changed document makes the requirement set stale within one check; the merge view shows conflicts element by element; the interface between the knowledge side and the workbench fixed at its first stable version.

The check found no screen that records a ready mark. The stage page now has one for the reviewer and one for the owner's self-check.

The check found regenerate failing for discovery findings, no agent help in the merge view, and no screen that starts a regeneration chain. Regenerating now keeps each element's identifier, merge has the agent's help, and the overview and the stale badge start a chain of regenerations.

The check found the stage workflows drafting only at the end, so a cancelled run had nothing to salvage. Drafts now stream as they form.

Imported from a token file, a Figma variables export, or by hand, with the theme map confirmed and contrast checked.

Each stage producing its artifact, with screens rendered from tokens only and viewers for each.

A sandboxed, clickable mockup that exports as one offline HTML file, with an annotated walkthrough.

Backlog with Given, When, Then scenarios, a bundle validated against its schema, citations resolved at export, and a read-only seat for engineering.

The reference initiatives with metrics and a judge, the comparison of two runs, and spend per initiative.

Selecting an element on a rendered screen to edit it, and notes anchored to an element that feed the next draft.

Every criterion checked, with the full scripted run from sources to the bundle.

An optional container with a small open embedding model, connected from the command line or from Models and keys, so search matches meaning with no hosted key.

The marketing site, its captures from the seeded application, its journeys, and this page.

Short rounds of questions before each stage drafts, answered in the run panel and recorded on the version. The pictures of them on this site are design renders, not pictures of the application.

Each stage page opening on its current output, and the overview's stage cards grouping outputs and counting gaps.

The bundle written in the folder layouts of spec-driven methods, with a prompts folder for engineering's planning.

User, administrator, and technical guides, the reader's guide for engineering, and the presenter's script for the fictional energy retailer's demo.

The mortgage finance company's corpus, its scripted run and demo script, and a comparison of the two clients' scores.

The defects the first scripted run and the 0c check logged, and the browser tests split across workers.

The product on one UK or EU virtual machine behind TLS, with backups and a restore drill, exercised once. The last item of the phase.