Newsdesk

governed generative video for newsrooms
every video ships with its receipt

It refuses to say what it cannot prove.

Newsdesk turns a reporter's verified facts into a short vertical news video — script, pictures, narration, subtitles — for about a dollar, in about ten minutes. It will not put a sentence on screen that can't be traced, word for word, to a fact a named journalist entered. When it can't prove a line, it refuses — for free — and says what to do about it.

Start a storySee the desk
~$1.30 a finished video$0 spent on a refusal23–27 words per narrated line8 free redrafts before surfacing3 walls between a fact and a frame
The journalist's path

How a story becomes a video

Start with facts — yours.

Paste a link to a story you reported and Newsdesk proposes facts, each carrying the exact quote it came from — a proposal whose quote is not on the page character-for-character is dropped, not shown. You confirm every fact yourself, or type them by hand. Every fact needs a source before a story can exist.

Pick the look.

A brand kit is a visual style your newsroom owns — the house mixed-media collage, or the paper-diorama documentary world. The through-line object is one physical thing that appears in every scene and carries the escalation, so six clips read as one film.

It writes the script — and checks every sentence.

Six ten-second blocks. Each narrated line must fit the narrator's measured pace (23–27 words), and every claim in it must trace to one of your facts with the supporting words matching verbatim. Failed drafts are refused and redrafted, up to eight times, at no cost — and a surviving refusal names the sentence, the rule, and what to do next.

It renders the pictures and the voice.

Money is spent only after the script survives every check: six clips and a measured narration take per block. The Run Board shows each stage live, with every cost on screen.

A human signs it.

Nothing publishes without a named editor approving on the review screen. The approval — with the approver's name — is written into the record before assembly will run.

The receipt travels with the video.

Every published video carries a manifest: every fact and source, every refusal and retry, every dollar, the approver's name, and a cryptographic fingerprint of the file itself.

Why you can trust it

Three walls

WALL 1

No fact without a source

The same parser checks a story whether it arrives from the browser, a file, or a link. A fact with no source is refused at the door, by name.

WALL 2

The policy gate, before any spend

Every scene is checked against written policy — no camera-passable photorealism, no real person's likeness, no on-screen text that traces to nothing. The gate runs with no network and no credentials: “$0 spent on a refusal” is structural, not a promise.

WALL 3

No publish without a name

Assembly will not run for a story with no named approver. The name is in the receipt.

When it says no

A refusal is the product working

The machine never judges whether your reporting is true — that's your byline, not its. It judges whether it can prove each sentence against your own entered words. When it can't, it refuses before spending anything, names the exact sentence and rule, and ends with the remedy:

What to do: retry — refusals are free and the model redrafts. If it keeps refusing, trim F3 (27 words) to one spoken sentence.

The alternative is a tool that would have aired the sentence anyway. Corrections have your name on them; refusals don't.