Feature status

Honest inventory against the v2.4 spec. Working = runs end to end with no keys. Wired = built but needs a credential or external service. Not built = absent.

Part 3 — Creator Studio

Part 4 — Film Studio

Part 5 — The Theater

Part 6 — Publishing Hub

Part 7 — Supporting systems

Part 9 — Legal & safety

/admin/readiness shows all of this live. Five blocking checks currently fail, so public launch is correctly blocked.

Verification

See the v2.4 section below — 214 assertions across three suites, run with npm run verify.

v2.4 additions

Verification

214 assertions across three suites — npm run verify.

The adversarial suite (verify-security.ts) is the important one: every case asserts that something is refused. It caught two real bugs during development — an OAuth unique-constraint crash and a foreign-key failure in account deletion — both documented in the README.

Final sweep — completed after the v2.4 audit

Everything below was found missing in a line-by-line re-read of the spec and has since been built.

Still not built, and why

Verification

476 assertions across five suites, plus a design-system check — npm run verify.

Two further scripts run against a live server:

• npm run smoke — all 59 routes render (scripts/smoke-routes.ts)
• npm run smoke:content — 58 assertions that pages contain what they claim (scripts/check-content.ts)

The style check caught four real bugs: bg-cine-900 and bg-cine-950 did not exist in the palette, so those backgrounds silently rendered as nothing.

Second full-spec sweep

A second line-by-line pass over the outline found nine more items with no implementation. All nine are now built.

Bugs this sweep found in existing code

• Head-of-line blocking in the generation queue. runQueue(n) drained the global FIFO oldest-first, so "enqueue N then run N" silently processed someone else's backlog while the caller's work stayed queued. runQueue now accepts the caller's own ids.
• Abandoned generations were never reaped. A crash mid-render left credits reserved against a job that would never finish — charged for nothing, and unspendable. reapStaleGenerations() now runs on every queue pass, fails them past 30 minutes and refunds the hold.
• The first compute cap broke batching. A 12-view model sheet is one user action; the initial concurrency limit of 2 made a documented feature impossible. The cap now ignores a just-submitted burst (60s grace) and the floor is set by the largest legitimate action.

Verification

583 assertions across seven suites, plus the design-system check — npm run verify.

Against a live server: npm run smoke (61 routes) and npm run smoke:content (70 assertions that pages contain what they claim).

Community forum & help desk

Three connected surfaces sharing one knowledge base.

Design decisions worth knowing

Articles live in code, not the database. They document the behaviour of specific functions and ship in the same commit as the features they describe. Documentation stored separately drifts, and a help centre that lies is worse than none.

The assistant can only cite articles that exist. That constraint is the whole design — it structurally cannot describe a button that is not there, which is the failure mode that makes most product chatbots untrustworthy. When the catalog has no answer it says so and offers the forum.

It works with no API key. In grounded mode the answer is assembled from the retrieved articles directly, which is genuinely useful because the articles are written as answers. The mode is labelled on every response.

Author labels are denormalized. A forum where every old answer says "deleted account" is a forum nobody trusts, and refusing to delete accounts is worse.

Bugs this work found

• Two broken cross-references in the help catalog pointed at slugs that did not exist. Now asserted — every related link must resolve.
• Short keywords matched gibberish. Plain substring matching meant a two-character keyword like x matched inside zzzqqxx, so nonsense scored against real articles. Search now matches on word boundaries.
• A single keyword counted as a full phrase match. "The render farm" inside a seventy-character nonsense question scored as a phrase hit on the export article. A phrase match now requires the keyword to cover half the query.

Verification

693 assertions across eight suites, plus the design-system check — npm run verify.

Against a live server: npm run smoke (71 routes) and npm run smoke:content (86 assertions).

Graphics Model Test

Compare engines side by side before committing a whole film to one.

Models added this pass

grok-imagine, grok-imagine-image, nano-banana-pro, ideogram-3, recraft-v3, firefly-image-4, pika-2.5, mochi-1, flux-schnell, sd-3.5-large.

35 models registered, of which 8 are self-hostable. Adobe Firefly is a second licensed engine, so Cleared Content mode is no longer one model deep — previously a single point of failure for the whole premium SKU.

Bugs this work found

• The UI and the service disagreed on the run threshold. The service allowed

≥2 selected / ≥1 runnable; the button required 2 runnable. With no provider keys that meant a permanently disabled button and no explanation.

• A preferred model from a different modality produced no explanation. Asking

for video on a project whose chosen stills engine was set silently routed elsewhere. It now says the preferred engine is not a model of that kind.

Verification

774 assertions across nine suites — npm run verify. Live: 73 routes, 86 content assertions.

Activation inventory, reachability and mobile

Admin → Activation

A live inventory of every feature and whether it is switched on, at /admin/activation. Three states:

• active — working now.
• needs-key — built and tested, waiting on a credential. Configuration, not development.
• not-built — genuinely absent, with what building it involves.

Computed from the environment, the credential vault and the model registry on every load — never a hand-kept list. A verification assertion flips FFMPEG_PATH and asserts the state follows, which is what stops it silently rotting into a page that says "active" for something broken a month ago.

Every pending item states what unlocks it and what happens today instead. Nothing fails silently.

Reachability

npm run check:reach walks every App Router page and every href in the codebase, and fails if any page has no inbound link. A route nothing links to is a feature nobody will find — indistinguishable, from the user's side, from not building it. 65 pages, all reachable.

Mobile

Bugs found and fixed this pass

• No navigation at all below 768px. The desktop nav is hidden md:flex; nothing

replaced it. A phone user could reach the Create/Watch toggle and nothing else.

• Horizontal overflow on the Graphics Model Test. A grid whose only column

definition sat behind lg: fell back to an implicit auto column, which will not shrink below min-content — so GOOGLE_AI_API_KEY is not configured widened the page to 526px on a 375px screen. check:mobile now fails on any grid without an unprefixed base column.

• undefined !== false corrupted the activation sort. notBuilt() left

blocksLaunch undefined while f() set it to false, so every needs-key item sorted below the not-built ones — exactly backwards from the order an operator works through.

• Two wrong assertions in my own mobile checker, both fixed rather than muted:

it matched max-w-[1800px] as a fixed width, and flagged grid-cols-3 as rigid when Tailwind compiles it to minmax(0,1fr), which shrinks fine.

Refactor

verifyCredentials() was extracted from signIn() so the mobile API can issue a Bearer token instead of a cookie while sharing one implementation of every gate — demo-account blocking, ban, suspension, 2FA. Two copies would drift, and the copy that drifted would be the one letting a banned user in.

A mobile token is a Session row: one device list, one revoke path, and a password reset or ban kills it. Asserted directly in verify-platform.ts.

Verification

827 assertions across ten suites, plus design-system, reachability and mobile checks — npm run verify. Live: 79 routes, 93 content assertions, 45 mobile checks.

Addendum A1 — ArcReel & ecosystem review

Full detail, including the design rationale and the bugs this pass uncovered, is in addendum-a1.md.

Corrections to earlier entries in this document

• Provider adapters. Nine adapter names in the registry had no implementation

behind them and silently resolved to the mock adapter. All nine are now implemented, and isModelLive() requires a registered adapter as well as a key so the state cannot recur. Any earlier row implying xai, pika, ideogram, recraft, adobe, google, vidu, selfhost or utility models worked on a key alone was wrong.

• Post-processing utilities (upscale, HDR master, interpolation, face

restore, lip-sync post) now refuse with a stated reason when no processing backend is configured, rather than resolving to a mock that returned the input.

Captions, dubbing, voices and the product page

Full detail in captions-dubbing-voices.md.