Media assets: offline validation and storyboard scaffolds

reachweaver media-check validates a media manifest before a human publishes images, videos or subtitle files by hand. It reads real file headers with the Python standard library and reports blocking errors and advisory warnings as text or JSON.

ReachWeaver does not upload, render, transcode or publish media. No enabled adapter sends images or video to any platform: the enabled live adapters are text-only, and the mock-tested Mastodon image (#13), Instagram image (#20), YouTube (#22) and TikTok (#23) adapters (owner-approved 2026-10-09) stay disabled in the capability registry until a live receipt. media-check makes no network requests and never calls a generation tool or model. Image, video, caption or translation generation tools are optional and user-configured: you choose them, run them yourself and record them in provenance. No paid model or provider is required. Local tools (for example an image editor, ffmpeg or a local speech-to-text model) cost nothing to use with ReachWeaver. Any paid provider you choose bills you directly; ReachWeaver neither integrates with nor pays for one.

Quickstart

reachweaver media-check media/manifest.json          # human-readable VALID / BLOCKED
reachweaver media-check media/manifest.json --json   # machine-readable report

The exit code is 0 when there are no errors and 2 when there are errors. Warnings never block, but each one needs a human decision.

Manifest format (schema_version 1)

Asset paths are relative to the manifest file and must stay inside its directory. Absolute paths and .. escapes are rejected.

{
  "schema_version": 1,
  "default_locale": "en",
  "target_locales": ["en", "it", "fa"],
  "preserve_terms": ["ReachWeaver", "Kubernetes"],
  "assets": [
    {
      "id": "release-card",
      "path": "release-card.png",
      "kind": "image",
      "targets": ["linkedin", "bluesky"],
      "locale": "en",
      "alt_text": "Terminal output of ReachWeaver media-check reporting zero errors for a release card",
      "alt_text_localized": {"it": "Output del terminale: ReachWeaver media-check senza errori"},
      "caption": "v0.3 validates media before you post it.",
      "content_warning": null,
      "attribution": "",
      "license": "owner",
      "provenance": {"created_by": "owner", "tool": "", "source_url": ""},
      "rights_confirmed": true
    },
    {
      "id": "demo-short",
      "path": "demo.mp4",
      "kind": "video",
      "targets": ["youtube_shorts", "bluesky", "linkedin"],
      "subtitles": [
        {"path": "demo.en.vtt", "locale": "en"},
        {"path": "demo.en.srt", "locale": "en"},
        {"path": "demo.it.vtt", "locale": "it"}
      ],
      "license": "owner",
      "provenance": {"created_by": "tool", "tool": "ffmpeg 7 (local screen capture edit)", "source_url": ""},
      "rights_confirmed": true
    }
  ]
}

Field	Rule
kind	image, video or subtitle. Must match the detected file header.
targets	Platform ids from the constraint table below. An unknown id produces a warning and is not checked.
alt_text	Required for images. It must be non-empty, must not be the file name or a placeholder (image, image of, photo, TODO, …), and must fit each target’s verified limit. A leading “image of …” produces a warning.
subtitles	Videos require at least one SRT or WebVTT sidecar. "burned_in_captions": true downgrades a missing sidecar to a warning, because burned-in captions cannot be verified offline.
content_warning	Optional. If present, it must be a non-empty string.
license	An SPDX id (for example CC-BY-4.0), owner or licensed. CC-BY* requires attribution.
provenance.created_by	owner, tool or third_party. tool requires provenance.tool and produces a disclosure/likeness review warning. third_party requires a real license (not owner) and attribution.
rights_confirmed	Must be JSON true. Anything else blocks the asset.
realistic_synthetic	Optional. true if the image or video was made or meaningfully altered with AI and looks real (a real person saying or doing something they did not, a realistic scene that did not happen). Anything other than true/false blocks the asset; true adds a warning. Leaving it out changes nothing. ai-label-check uses it to tell you which platforms ask for an AI label (AI-CONTENT-LABELS.md).
locale / alt_text_localized	BCP-47-style tags. English is the default and the first locale. A declared target_locales entry without alt text (images) or subtitles (videos) produces a warning. A translation that drops a preserve_terms entry present in the source text produces a warning.
subtitles[].locale / subtitles[].alignment	A track whose locale differs from the asset’s is a translated track and is checked cue by cue against the asset’s source-locale track (below). alignment is source (default: the cues follow the source spotting) or independent (the owner re-timed the track on purpose; structural findings become warnings).
subtitle_timing_tolerance_ms	Optional, top level. Integer 0-5000, default 500: how far a translated cue’s start or end may sit from its source cue.

Secret-like keys (token, password, cookie, …) are rejected anywhere in the manifest.

Translated subtitle tracks

For each asset, the first well-formed subtitle track in the asset’s own locale is the source track. Locale tags are compared case-insensitively, as BCP-47 requires (RFC 5646 section 2.1.1, accessed 2026-10-09), so en-us is the source track of an en-US asset and IT in target_locales matches an it track. Every well-formed track in another locale is compared with it. Tracks with errors, including an unresolved placeholder, are reported as usual and take no part. If there are translated tracks but no usable source track, a warning says they were not compared.

Finding	Severity	Rule
Cue count differs	error	has 2 cues; the source track has 3.
Missing or extra cue	error	With equal counts, cues pair by position. Otherwise they pair by timing, and each unpaired source cue (missing a cue for source cue 2 (00:00:02.000 --> 00:00:04.500)) or translated cue (extra cue 4 ... has no source cue within 0.5s) is listed.
Timing drift	error	A paired cue whose start or end is more than the tolerance away from its source cue, with both timings and the drift.
Dropped preserved term	warning	A preserve_terms entry in the source cue that the paired cue lacks (case-sensitive, after removing <i>, <c.x>, inline timestamps and {\an8}-style tags). Terms absent from the source cue are not required.
Changed number	warning	Digit runs compared as a multiset, so 1,500, 1.500 and Persian ۱٬۵۰۰ all match and leading zeros are ignored (007 = 7), while 3 -> 4 reports missing 3; added 4. Runs of any length are compared as text, never through int(), and runs longer than 20 digits are shortened in the finding. A number written in digits on one side and spelled out (“tre”) on the other is reported for a person to judge.
Dropped or changed URL	warning	Every http(s):// or www. URL in the source cue must appear unchanged in the paired cue (trailing punctuation ignored).
Added URL	warning	adds URL ... that the source cue does not contain: a URL in the translated cue that its source cue lacks (or has fewer times). A link injected during translation is a phishing risk, so a changed URL is reported both as dropped and as added.

Structural findings are errors because a track out of step with the speech misleads every viewer, and they are objective. alignment: "independent" downgrades them to warnings for that one track. Content findings are warnings, matching the alt-text preserved-term rule: a translation may adapt wording, so a person decides. At most 20 findings per severity are listed per track; the rest are counted in one more finding of the same severity. The JSON report adds subtitle_translations per asset: path, locale, source path, cues, source_cues, paired, missing, extra, mistimed, tolerance, alignment and the error and warning counts. Translation quality itself is not assessed.

What is read from files

• PNG (IHDR), JPEG (SOF markers; EXIF APP1 detected), GIF (logical screen), WebP (VP8, VP8L, VP8X).
• MP4/MOV (ISO BMFF): container brand (qt means MOV), duration from moov/mvhd, display width and height from the video track’s tkhd (with 90/270-degree rotation matrices applied), and top-level box order (moov before mdat, reported as fast_start).
• MP4/MOV codecs, frame rate and audio (see below): the stsd sample-entry fourcc of the first video and first audio track, the average frame rate from mdhd timescale and stts sample deltas, and the audio sample rate and channel count from the audio sample entry.
• File size and extension-versus-header mismatch. A mismatch is an error, except MP4 versus MOV, which is a warning because both are ISO BMFF.
• SRT/WebVTT: UTF-8 encoding, WEBVTT header, numeric SRT indices, well-formed timestamps, end > start, monotonic start times, non-empty cue text, overlap warnings, and cues running past the video’s duration.

When a header cannot be parsed, the value is reported as "unknown" and the dependent checks are skipped with a warning. ReachWeaver never guesses a value. EXIF metadata triggers a warning because it can reveal location or device data and platforms may strip it.

Codec, frame-rate and audio parsing

For kind: video the report adds video_codec, video_fourcc, frame_rate_fps, variable_frame_rate, frame_rate_range_fps, audio_tracks, audio_codec, audio_fourcc, audio_sample_rate_hz, audio_channels and fast_start. Every value is read from the file or reported as "unknown".

Sample entry fourcc	Reported codec
avc1, avc3	h264
hvc1, hev1	hevc
av01 / vp09 / vp08	av1 / vp9 / vp8
mp4a	aac only when the esds objectTypeIndication is an AAC type (0x40, 0x66-0x68); mp3 for 0x69/0x6B; otherwise unknown
Opus / ac-3 / ec-3	opus / ac3 / eac3
anything else (for example encrypted encv)	unknown, with the fourcc still reported

• Frame rate is samples x timescale / total ticks over the video track’s stts, rounded to 3 decimals. That is an average.
• Variable frame rate is true only when stts holds more than one distinct sample delta. A single shorter final sample is ignored, because encoders commonly write one. frame_rate_range_fps gives the slowest and fastest per-sample rate. If the deltas cannot be read, it stays unknown and no VFR warning is raised.
• Audio sample rate and channel count come from the ISO/QuickTime v0 sample entry or the QuickTime v2 extended fields. The ISO AudioSampleEntryV1 rate sentinel and its srat box are not parsed, so they report unknown.
• audio_tracks: 0 is reported only when the whole moov was walked without a parse problem. Then the audio rules pass, because there is no audio to check.

Guards. The parser reads at most MAX_MOOV_BYTES (64 MiB) of moov and walks at most 10,000 top-level boxes, 200,000 boxes inside moov, 16 levels of container nesting and 1,000,000 stts entries. It stops at the first box whose declared size is smaller than its header or runs past its parent, and that covers absurd 64-bit sizes. Each guard that fires adds a header parsing stopped early (...) warning, and the affected values stay unknown. No exception escapes, and recursion never exceeds the depth guard.

For vertical video, the report adds a safe-area note only, because media-check reads no layout. To check where text and focal points sit, use the storyboard’s box check (below). Preview each upload in the native app either way.

Platform constraint table (verified 2026-10-08; github 2026-10-14, devto 2026-10-15)

Every row carries its source URL and its own verification date in reachweaver/media.py (CONSTRAINTS); the report lists every date in constraints_verified_dates. Where the cited documentation does not establish a value, the row is unknown: it is skipped with a warning, never invented. Limits change, so re-check before publishing.

Target	Kind	Verified constraints	Unknown (skipped)	Source
bluesky	image	any detected image; ≤ 2,000,000 bytes; alt ≤ 2000 (app constant, counted in code points)	aspect, min resolution	images lexicon, app constants
bluesky	video	MP4; ≤ 300,000,000 bytes; ≤ 600 s; captions must be WebVTT ≤ 20,000 bytes	aspect, min resolution, video/audio codec, frame rate	video lexicon
mastodon	image / video	documented defaults only, as warnings: images ≤ 16 MB, video MP4/MOV ≤ 99 MB; instance-configurable	alt limit, duration, aspect, caption files, video codec, frame rate (instance-reported video_frame_rate_limit; Instance entity)	posting docs, media API
instagram	image	JPEG; ≤ 8 MB; aspect 4:5 to 1.91:1; width 320–1440 (warnings: rescaled); alt ≤ 1000	none	IG User Media
instagram	video (Reels)	MP4/MOV; moov before mdat; ≤ 300 MB; 3–900 s; aspect 0.01–10; width ≤ 1920; alt text not supported; video H.264 or HEVC; 23–60 fps; audio AAC, ≤ 48 kHz, 1–2 channels	caption files	same
youtube_shorts	video	≤ 180 s; square or vertical (width/height ≤ 1). Warnings only (recommended encoding): H.264 video; AAC or Opus audio; 48 kHz; moov at the front	container, size, resolution, caption files, frame rate (any recorded rate is acceptable), channel layout	YouTube Help, recommended upload encoding settings
tiktok	image	JPEG/WebP; ≤ 20 MB	aspect, resolution (“up to 1080p” is ambiguous), alt	media transfer guide
tiktok	video	MP4/MOV (WebM not detected); ≤ 4 GB; ≤ 600 s; each side 360–4096 px; video H.264, H.265, VP8 or VP9; 23–60 fps	aspect, caption files, audio codec	same
linkedin	image	JPEG/GIF/PNG; < 36,152,320 pixels; alt ≤ 4086	file size, aspect	Images API
linkedin	video	MP4; 75 KB–500 MB; 3–1800 s; captions SRT (one file, English only)	aspect, video/audio codec, frame rate	Videos API
github	image (repository social preview)	PNG/JPEG/GIF; < 1 MB (decimal, conservative); width ≥ 640 and height ≥ 320. Warning only: about 2:1, because the docs recommend 1280 × 640 for best display	alt text (no field described)	GitHub Docs
devto	image (post cover)	Warning only: aspect near 1000 × 420 (about 2.38:1), the guide’s “best size”, not a limit	file types and file size (not stated)	DEV editor guide

Where a unit is ambiguous (KB versus KiB, decimal versus binary GB), the stricter interpretation is used and noted in the row.

Codec and frame-rate rules (verified 2026-10-08). A frame rate outside a documented range is an error (Instagram Reels and TikTok, both 23–60 fps). The range is checked against the average frame rate. When variable frame rate is detected, each target with a documented range also gets a warning that gives the per-sample span. Neither platform’s page addresses VFR, so VFR never blocks. YouTube’s encoding page publishes recommended settings, so its codec, sample-rate and fast-start rows are warnings and never block. That page also says other frame rates are acceptable, so YouTube has no frame-rate range. LinkedIn’s Videos API on Microsoft Learn states only MP4, 75 KB–500 MB and 3 s–30 min. LinkedIn’s separate Video Ads specifications list H.264/VP8, AAC/MPEG4 and under 30 fps, but they cover ads, not organic posts, so they are not enforced. Bluesky’s lexicon and app constants name container MIME types only. Rows these sources do not establish stay unknown and are skipped with a warning.

Storyboard scaffold: media-storyboard

media-storyboard turns one campaign brief into the files a person needs to make a vertical video and its images. The brief is the campaign file: the campaign contract or a signal-brief scaffold (SOURCES.md). It writes no copy, alt text or subtitle text, generates and edits no media, and calls no model, provider or network.

reachweaver media-storyboard --campaign campaign.json --out media/storyboard          # defaults below
reachweaver media-storyboard --campaign campaign.json --out media/sb2 --duration 45 \
  --video-targets youtube_shorts tiktok instagram bluesky --image-targets linkedin instagram
reachweaver media-storyboard --check media/storyboard/storyboard.json                 # after editing
reachweaver media-check media/storyboard/media-manifest.json                          # after producing media

Defaults: 30 s (10–180), --aspect 9:16 (1080x1920 frame), --video-targets youtube_shorts tiktok instagram, --image-targets linkedin, --locale en, --fit cover. --out must be a NEW folder. It is created 0700 and every file is a new 0600 file; an existing folder is refused. The command exits 0 when the scaffold is written, and the scaffold itself is expected to be BLOCKED.

File	Contents
storyboard.json	VerticalVideoStoryboard: campaign id and sha256, frame, scenes, evidence, safe zones, safe-area result, video_spec, image_specs, generation
storyboard.md	The same for a human: scene table, evidence claims in a fenced data block, safe areas, image specs, cost notes, checklist
subtitles.<locale>.srt	One cue per scene with valid, contiguous timing and placeholder text. A WebVTT copy is added when a target only accepts WebVTT (Bluesky)
media-manifest.json	Schema-1 manifest in the format above: image-<target> per image target and vertical-video.mp4 with the subtitle sidecars

Scenes. One hook (0–3 s: the visible outcome first), one beat per cited evidence record (at most three, each citing its evidence_ids) and one cta (the last 4 s). Each scene has on_screen_text, voiceover and visual placeholders, plus a text box and a focal box placed inside the combined safe area of the video targets. Boxes are fractions of the frame: x, y, w, h in 0..1.

Blocking by design. Alt text, captions, subtitle cues and scene copy are {{write: ...}} placeholders, and rights_confirmed is false. media-check therefore blocks the manifest until the media files exist, a person has written every placeholder in their own words and confirmed rights. Each image’s alt-text placeholder quotes the target’s sourced alt-text limit, or says that the cited source documents none. Alt text is never generated, because only someone who has seen the final image can describe it. media-check now also rejects a subtitle file that still contains a placeholder. The builder refuses to write a manifest that validates as generated, or one that fails for any reason other than missing files, placeholders or unconfirmed rights. Tests fill in a scaffold with fixture PNG and MP4 files and confirm that the existing media-check then passes it.

Specs. image_specs.<target> names the file, the container (LinkedIn PNG, Instagram and TikTok JPEG, Bluesky and Mastodon PNG) and a suggested canvas, and copies that target’s sourced rows from the constraint table. Each canvas is labelled within_sourced_limits (Instagram 1080x1350, 4:5) or unverified_convention (LinkedIn 1200x627, Bluesky and Mastodon 1200x675, TikTok 1080x1920), because those platforms’ cited documents state no aspect ratio. video_spec gives the frame, the duration and each video target’s sourced rows.

Crop and UI-overlay safe areas

check_safe_area takes the frame’s aspect ratio, each target’s overlay zone and a fit assumption:

• Overlay zone. The fractions of a full-screen 9:16 player covered by app UI (top bar, caption and username, action rail, progress bar), given as top, bottom, left and right.
• Fit. cover scales the frame to fill the player and crops the overflow evenly from both sides, which is the conservative default. contain letterboxes it.

The zone is mapped into frame coordinates. Any box outside the visible area is reported as cropped, and any box inside the visible area but under the UI as under_ui_overlay, with the overshoot per edge in frame fractions and pixels. A text box outside a safe area is an error, and a focal box is a warning. Targets with no zone data (Bluesky, Mastodon, LinkedIn) are skipped with a warning. combined_safe_rect is the intersection across all targets, so a single layout fits every selected platform.

Worked examples, which the golden tests also check:

• A 1:1 frame filled into a 9:16 player shows only x = 0.21875..0.78125. A text box at x = 0.05 is reported as cropped by 182 px.
• A 16:9 frame shows 31.6% of its width.
• In a 9:16 frame on TikTok, a text box ending at y = 0.9 sits 0.24375 (468 px) under the sourced bottom overlay (safe bottom edge 1 - 0.34375).

Built-in zones are sourced from first-party PAID-AD specs (accessed 2026-10-09). Each carries status: "sourced", the source URL, verified_date: "2026-10-09", placement: "ads", the source_document, the numbers as stated by the source, an evidence quote or description and a note. None of these sources states an organic-post overlay, so every check also adds a warning that organic overlays may differ. Overlays vary by app version, device, language and caption length, so preview each upload in the native app. Fractions derived from pixel values are rounded up to 6 decimals, so a zone is never smaller than its source.

Target	top	bottom	left	right	Source (paid ads)	What the source states
tiktok	0.125	0.34375	0.111112	0.277778	TikTok Ads Manager: TikTok Auction In-Feed Ads (“Last updated: June 2026”), its “In-Feed-Standard Version LTR” safe-zone template	The page text states no numbers; it says the safe zone depends on dimension, ad caption length and add-ons, and links the template. The template image is labelled on a 720x1280 base: 160 px top, 440 px bottom, 80 px each side, plus a 120 px action column inside the right margin over the bottom 720 px. A zone here is a rectangle, so right applies 80 + 120 px to the full height (stricter than the template above the column). Longer captions, anchors and Arabic (RTL) placements have their own templates.
instagram (Reels)	0.14	0.35	0.06	0.06	Meta Ads Guide: Awareness Video Ad Specs on Instagram Reels	“Consider leaving at least 14% of the top, 35% of the bottom, and 6% on each side of your asset free from text, logos, …”
youtube_shorts	0.15	0.35	0.044445	0.177778	Google Ads Help: Use square and vertical video to engage mobile customers, section “Safe zones for vertical video ads on YouTube”	The page text states no numbers. Its “YouTube | Vertical Video Ads Safe Zone” image (1080x1920, “Dimensions in pixels”) labels 288 top, 672 bottom, 48 left and 192 right. It covers vertical video ads across in-stream, in-feed and Shorts inventory.

To use other numbers, for example a template for longer TikTok captions or a measured organic overlay, pass --safe-zones FILE:

{"tiktok": {"top": 0.125, "bottom": 0.40, "left": 0.111112, "right": 0.277778,
            "source": "https://<official page you opened>", "verified_date": "2026-10-09"}}

Each value must be a fraction from 0 up to, but not including, 0.5. Only video targets from the constraint table are accepted. An entry is labelled owner_sourced only when it has a public https source and a YYYY-MM-DD verified_date; otherwise it stays unverified, and every check that uses it adds a safe zone is UNVERIFIED warning. --check takes zones from the built-in table or --safe-zones and never from the edited storyboard.json, so changing zone numbers inside the file cannot make a layout pass. --check also reports any scene field that still contains a placeholder, and validates the subtitle timing and placeholders. It exits 2 on any error.

Generation is opt-in, and what it costs

generation.enabled is always false. ReachWeaver does not generate or transform media, and has no generation integration to switch on.

• Manual or local tools cost nothing through ReachWeaver or any provider; the only costs are your hardware, electricity and time. Examples are a screen recording of the real product, an image editor, ffmpeg (trimming, burned-in captions, -movflags +faststart) and a local speech-to-text model for a first subtitle draft. Local model output still needs a human to check every word.
• A paid provider that you sign up for bills your own account at its current published price. ReachWeaver holds no key, has no integration and pays nothing. No prices are recorded here because they change, so read the provider’s pricing page and terms before use, including who owns the output and its likeness and voice rules.

Whichever tool you use, record it in the manifest: provenance.created_by: "tool" and provenance.tool: "<name and version>". media-check then asks for disclosure and likeness review.

Cut list: media-cutlist

Repurpose one master video for several platforms. You describe the master and mark its segments; ReachWeaver says, per platform, which segments to keep, how to crop and which limits apply. It never opens, cuts or encodes a file: the master’s numbers are yours, and it prints an ffmpeg command for you to run.

reachweaver media-cutlist examples/cut-list.example.json          # readable plan
reachweaver media-cutlist examples/cut-list.example.json --json   # for scripts

{
  "schema_version": "reachweaver.cut-list.v1",
  "source": {"file": "media/release-talk.mp4", "duration_s": 312.5, "width": 1920, "height": 1080,
             "focus_x": 0.42, "focus_y": 0.5, "has_audio": true},
  "segments": [{"id": "hook", "start_s": 4.0, "end_s": 16.0, "label": "The one-line problem", "priority": 1}],
  "targets": [{"platform": "youtube_shorts"}, {"platform": "tiktok", "max_s": 60}, {"platform": "instagram", "aspect": "4:5"}]
}

Field	Rule
source.file	The master’s path. It only appears in the suggested command; the file is not opened
source.duration_s, width, height	The master’s length and frame size, as your player or ffprobe reports them
focus_x, focus_y	0 to 1 (default 0.5): where the subject is, so a crop keeps it in frame
segments[]	Up to 100, no overlaps: id (lowercase slug), start_s, end_s, optional label, priority 1 (keep first) to 5. No segments means the whole video
targets[]	Up to 12 of youtube_shorts, tiktok, instagram, linkedin, bluesky, mastodon (the platforms with a video row in the constraint table); optional aspect (W:H) and max_s, which can only shorten a cut

For each target:

• Keep: segments are taken by priority, then by time, while the cut fits the platform’s maximum duration, and are then played in time order. The maximum and minimum come from the constraint table with their official source and checked date. A limit the source does not establish (for example Mastodon’s duration) is reported as unverified with a warning, and nothing is cut to fit it.
• Aspect and crop: your aspect, else 9:16 for Shorts (inside the official square-or-vertical range; the choice is ReachWeaver’s) and for TikTok (a hypothesis: the cited sources set no aspect), else the master’s own. The crop is the largest window of that aspect centred on the focus point. The output is scaled down (never up) so its short side is at most 1080 px (1080x1920, 1920x1080, 1080x1080), within the platform’s sourced maximum width or dimension (Instagram 1920 px wide, TikTok 4096 px), with even sides as H.264 4:2:0 needs. A smaller picture keeps its size with a warning, and one below the platform’s minimum (TikTok 360 px) is flagged. An output name that would start with - gets a ./ prefix so ffmpeg never reads it as an option.
• Subtitles and size: the subtitle format (vtt for Bluesky, srt for LinkedIn) and the maximum file size, when the source establishes them.
• Suggested command: ffmpeg with the trim/atrim, setpts, concat, crop and scale filters (ffmpeg filters, accessed 2026-10-17), H.264 with AAC audio and -movflags +faststart, written to NAME.PLATFORM.mp4 next to the master, with every argument shell-quoted. ReachWeaver and its tests do not run it: check it with your ffmpeg version and watch the result.

Then run media-check on the cut before attaching it to a draft. The cut list adds no media to a post and changes no approval.

Rights on the attached image (issue #48, draft for owner review)

media-check validates the rights of a manifest, but until now an image attached to a draft carried none of it. media-attach now accepts an optional rights record that is stored with the image, shown by show and the confirmation display, and bound by the approval hash:

reachweaver media-attach POST_ID --file photo.png --alt "A cat on a sofa" \
  --rights-confirmed --license CC-BY-4.0 --created-by third_party \
  --attribution "Photo: A. Photographer" --rights-source-url https://example.org/photo \
  --alt-translation it="Un gatto su un divano" --alt-translation fa="یک گربه روی مبل"

• Rules (the same as the manifest’s): --rights-confirmed is required with any rights option (you hold the rights, consent and likeness permission); --license is an SPDX id, owner or licensed; --created-by is owner, tool (then --tool names it, and a disclosure note is printed) or third_party (then a real license and --attribution are required, and a source URL is advised); CC-BY* needs attribution. Incomplete input refuses the attachment and changes nothing.
• Translations are your own alt text per BCP-47 language, validated like alt text. They are stored and bound so you can review and edit them with the post, but never sent: a platform takes one description, the --alt one.
• Binding: an image with a rights record is bound under approval/media-2 (the rights are inside the digest, so changing the license, attribution, provenance or a translation after approval invalidates it). An image without one keeps approval/media-1 byte for byte, so every existing approval still verifies. Separately, the AI-label declarations (ai-declare) and the approve --ai-label answer of a post with media are bound under the add-on tag approval/media-3 (AI-CONTENT-LABELS.md).
• Not published: rights text is never part of what a send publishes. Nothing is fetched or checked against a remote source; the record is the owner’s statement, not proof.

Mock upload adapter: reachweaver/media_upload.py (issue #18)

MockMediaUploadAdapter(platform).preflight(manifest, base_dir) (or .preflight_file(path)) simulates what an upload adapter would reject, for one platform, with no network, no credential, no database access and no upload. It reuses the media-check validator, keeps only assets whose targets include the platform (subtitle sidecars always travel with them), drops other platforms’ findings, and turns each message into {code, asset_id, message, action}. The result carries simulated: true, uploaded: false and accepted (true only when there are no errors). accepted is not consent to publish; approval, dry-run and the send gate are untouched. It is a library API with no CLI subcommand.

Code	Meaning
UNKNOWN_PLATFORM, NO_ASSETS_FOR_PLATFORM, MANIFEST_INVALID	Platform id has no verified table; no asset targets it; manifest-level problem
ASSET_PATH_INVALID, FILE_HEADER_UNRECOGNIZED, KIND_MISMATCH	Missing, absolute or escaping path; unreadable header; declared kind differs from header
CONTAINER_NOT_ALLOWED, FILE_TOO_LARGE, FILE_TOO_SMALL, DURATION_OUT_OF_RANGE, ASPECT_RATIO_UNSUPPORTED, RESOLUTION_OUT_OF_RANGE	Sourced platform limits
CODEC_NOT_ALLOWED, FRAME_RATE_OUT_OF_RANGE, AUDIO_SPEC_UNSUPPORTED	Container-metadata codec, frame-rate and audio rules
ALT_TEXT_MISSING, ALT_TEXT_PLACEHOLDER, ALT_TEXT_TOO_LONG, ALT_TEXT_NOT_SUPPORTED	Accessibility
CAPTIONS_MISSING, SUBTITLE_FORMAT_UNSUPPORTED, SUBTITLE_TOO_LARGE	Captions
RIGHTS_NOT_CONFIRMED, LICENSE_INVALID, PROVENANCE_INVALID, ATTRIBUTION_REQUIRED, PLACEHOLDER_UNRESOLVED	Rights fields, checked in memory only
MEDIA_INVALID	Any other validator error (fallback; read message)

Codes are derived from validator message text, so a message the classifier does not recognise degrades to MEDIA_INVALID rather than being dropped. Subtitle-track structural errors (cue syntax, translation drift) currently fall into MEDIA_INVALID. Provenance and license are not persisted anywhere; storing them in the database is a deferred, security-reviewed item.

Limits of this slice

• Codec, frame-rate and audio checks read container metadata only. ReachWeaver never decodes a stream, so it does not check codec profile or level, bitrate, GOP structure (Instagram asks for closed GOP), progressive versus interlaced scan, chroma subsampling, edit lists, HDR or colour profile. Only the first video and first audio track are checked, and fragmented MP4 (moof) timing is not read. No crop-preview rendering, no thumbnail checks and no OCR of burned-in text. media-storyboard checks the boxes you declare, not the pixels: it cannot see where text actually lands in a rendered file. Its built-in safe zones come from paid-ad specs; organic overlays may differ.
• No automatic fact-check of captions, and no detection of likenesses, copyrighted footage or sensitive content in pixels. rights_confirmed and provenance are owner declarations that a human must stand behind.
• Translation quality is not assessed. Checks flag missing locales, dropped preserved terms, and, for translated subtitle tracks, cue count and timing drift (errors) and dropped terms, changed numbers, and dropped or added URLs per cue (warnings).
• Tests build tiny PNG, JPEG, GIF, WebP and MP4 headers in temporary directories, so no binaries are committed. The MP4 fixtures include stsd/mdhd/stts and audio sample entries (H.264 and HEVC at 30 fps with AAC, malformed sizes, 5,000-level nesting, moov after mdat). The WebP and MP4 fixtures are structurally valid headers, not decodable media.