Blind Creator Pre-Flight — Support and How It Works

The six checks

1. Framing — is the subject in the shot, and where? SAM 3.1 locates the subject and any sign where it can. Where SAM finds no match, Muse Spark’s judgement can estimate the positions and the result says it is an estimate. Where neither supplies positions, the take’s recorded facts stand.
2. Crop safety — would the platform’s crop cut anything that matters? The check tests the subject and sign against the crops a post is shown through, including the vertical 9:16 crop and the portrait 4:5 crop.
3. Sharpness and quality — is the take usable? This is measured from the take’s own pixels.
4. Text in the image — what does any visible text say, in reading order, and is it legible?
5. First impression — what does a sighted viewer see first, and does that match what you meant the post to show?
6. Audio — what does the take sound like? The design judges speech from a transcript made by Muse Voice Transcribe. A still photo has no audio track. A silent video reports that there is nothing to transcribe, because the silence is part of the take — it is not a defect and not a failed check.

Alongside those six, a private-information warning can fire with any set of results — including an all-pass set. It names anything in the take that looks private, such as a readable address, a document, or a screen notification, before you post it publicly.

How to use Pre-Flight

The current build is a demonstration build. It works on written example takes, called seeded takes, each with a known state — clean, or carrying one named flaw — so the checks can be demonstrated and tested. There is no upload feature in this build.

1. Pick a seeded take from the list. Each one states the creator’s intent for the post.
2. Run the pre-flight check. Read the six results and any private-information warning.
3. If a check says fix, the repairs for it appear in the Repairs section. Repairs are deterministic — re-cropping, re-framing, exposure, or choosing a cover frame — and they are performed on a copy. The original take is never changed, you can undo in the current session, and the checks run again on the repaired take so you can hear what changed.
4. Read your proof copy: a private, speech-and-braille-ready text version of what you are about to publish, built from what the checks actually found. You can download it as a plain text file.
5. Edit the drafted alt text into your own words and approve it. The current draft is assembled from your stated intent and measured facts. Editing it clears any earlier approval, so approval always covers the exact words as they stand.
6. If a private-information warning is showing, read it and tick the acknowledgement to say you understand what posting would make public.
7. Read the readiness result. A take is ready when every check is settled, any warning has been acknowledged, and the alt text has been approved.

An optional public description — the same verified facts, written for your audience — is offered for a photo. It is not needed for readiness, and it is used only with your approval of its exact words. The Library Sweep is a simulated demonstration of a wider idea: knowing what is in a photo library, in words and counts, before any app sees it. Culling in the demonstration deletes nothing anywhere, and the page says so each time.

What the verdicts mean

• Pass — the check ran and found no problem.
• Fix — the check ran and found a problem. Where a deterministic repair exists, it is offered; the checks run again after any repair.
• Uncertain — the check could not settle the question, and it says why. Uncertain is always an allowed answer in Pre-Flight: the app never guesses a pass. A check whose model call is switched off, has failed, or came back unreadable reports uncertain with that reason.

Every verdict names how it was reached — measured from the take, measured by SAM 3.1, estimated by Muse Spark’s judgement, judged by the model check, or determined by the take itself. An estimate is never presented as a measurement.

The private-information warning

The warning is separate from the six verdicts, and it can fire even when every check passes. It names the private thing plainly — for example, a posted letter with a readable home address visible on a desk — and says what posting would make public. A take with a firing warning is not ready until you have acknowledged the warning yourself. Approving the alt text never substitutes for acknowledging a warning.

The warning is a safeguard, not a guarantee that everything private has been found. Review the proof copy and the media description before posting.

Accessibility notes

Accessibility is the requirement in Pre-Flight, not a feature. It was built screen-reader-first:

• Every flow in the demonstration completes with a screen reader and a braille display alone.
• Every status exists as plain text. A check’s meaning is always carried by its verdict word — pass, fix, or uncertain — and by its source, never by colour, icon, or position alone.
• Results are placed in a polite live region, and the proof copy can be downloaded as a plain, braille-ready text file.
• Nothing is published, attached, or approved on your behalf. Approval covers the exact current words, and editing approved words clears that approval.

This website follows the same rule: semantic headings, a skip link, alt text on images, visible description captions on the media page, and pages that stay readable with the stylesheet off.

Known limits in this build

• Seeded example takes only. ST-V1 and ST-V3 are specifications without media files yet, so their checks report that there is nothing to measure or judge.
• No seeded take ships a separate speech audio file, and the running server does not decode video. The transcription code path is real and tested with test doubles, but the seeded speech branch is not demonstrated by a seeded take.
• The backend can prepare a described-audio draft through a separate route, but the current web interface does not present that draft. Do not expect a described-audio control in the demonstration.
• The timestamped public description for a video is not built yet, so Pre-Flight reports it unavailable rather than inventing a timeline.
• When live model calls are switched off, the model checks report uncertain and say why. The measured checks still run from the take’s own media and recorded facts.
• The current client does not use Muse Wonder. Muse Wonder was evaluated during model selection, but the production judgement and semantic calls use Muse Spark 1.3.

Get support

For questions, problems, or accessibility barriers in Pre-Flight or on this page, contact Digital Dragons Inc. If you can, tell us which take, which check, and what your screen reader or braille display reported. Please do not send passwords or private media unless we ask for a specific item and explain why.

• Email: nick@digitaldragonsinc.com

Visit the main Digital Dragons Inc website