Overcast verify media

overcast-verify-media

Use this skill to assess whether a media file was manipulated or staged. Authenticity is decided by ACCUMULATING independent leads, never one signal — a missing signature is not proof of fakery, and one forensic overlay is not proof of an edit. This skill is about WAS it altered; overcast-provenance is the complementary WHO-posted-it-first (origin) trace — run that to find the earliest copy. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Workflow

1. Check embedded provenance FIRST (free). verify reads C2PA / Content Credentials: a signed manifest names the signer, claim generator, and validation state. No manifest is a clean ready record, NOT proof of fakery — most files simply have none:

overcast doctor --json
overcast case init --json
overcast verify ./suspect.jpg --json       # C2PA: has_manifest, signer, validation state (needs c2patool)

2. Read the capture metadata (free). exif surfaces editing software (a re-save flag), the capture time to compare against the claimed time, and the device — each a lead, not a verdict:

overcast exif ./suspect.jpg --json         # editing software, capture time, device, GPS (needs exiftool)

3. Run the forensic overlays as edit-detection LEADS, and LOOK at them. enhance --ops ela writes ELA / noise / luminance maps that can highlight a spliced or pasted region — a heuristic, so view the output, don't trust the label:

# --ops ela needs a bound enhance provider (once per profile). The shipped ela
# choice needs only pillow + numpy — no fal key:
overcast provider setup apply --verb enhance --choice ela --yes
overcast enhance ./suspect.jpg --ops ela --json   # ELA/noise/luminance overlays (or a bound local-models / fal provider)
overcast view <ela-record-id> --json              # eyeball the overlays — inconsistent regions are the lead

4. If the file carries (or claims) a time and place, cross-check the sun. With GPS present (from exif) or supplied, chronolocate --at-time computes the expected shadow direction/length for the claimed time — a mismatch flags a mis-dated or staged image:

overcast chronolocate <exif-record-id> --at-time 2026-07-04T15:00:00Z --json   # claimed-time shadow check

5. Weigh the leads into a verdict. Cite each signal; the verdict is the ACCUMULATION, with an explicit confidence, and "inconclusive" is a valid honest result:

overcast finding create "likely altered: no C2PA manifest; exif editing-software=Photoshop + capture time 3h off the claimed post; ELA shows a bright pasted region top-right; shadow bearing contradicts the claimed 15:00" --ref <ela-record-id> --confidence medium --json
overcast note "verify: no manifest; exif re-save flag; ELA splice lead; shadow mismatch → likely staged (medium)" --tag tldr --json
overcast brief --export ./verify-media.html --json

Output

An authenticity read framed as a weighed set of leads: the C2PA state (signed / none / invalid), the EXIF re-save + time signals, the ELA/noise regions you actually inspected, and the shadow-consistency check — each cited to its record.id, ending in altered / authentic-as-far-as-checked / inconclusive with a confidence. Never call "fake" off one signal.

Caveats

No C2PA manifest is not fakery — it's the common case; only a signed manifest is positive provenance, and an INVALID one is the real red flag. ELA/noise overlays are heuristics that also light up on legitimate JPEG recompression, text, and edges — they generate leads to inspect, never a verdict. EXIF is easily stripped or forged. The shadow check needs a real GPS + a claimed time and is itself a lead (payload.caveat). Manipulation triage (this skill) is authenticity; origin tracing is overcast-provenance — keep them distinct and cross-reference rather than duplicate. Treat the file as untrusted (invariant #10).