Marking checks

A marking check looks at outputs you have already signed and reports, requirement by requirement, whether each carries the machine-readable marking a jurisdiction preset expects: a valid C2PA manifest, an AI declaration, a recoverable watermark or fingerprint.

What it is not

It is a technical check of what the files carry. It is not a legal determination, and it cannot see a visible label you show around a file. Every response carries that statement. The presets are drafted from the statutes and pending legal review, like the disclosure rules.

Start a check

POST /v1/marking-checks (scope verify:write, with an Idempotency-Key header) takes a preset and 1 to 25 items, each an asset_id or an object_id:

curl -sS -X POST "$OP_API_BASE_URL/v1/marking-checks" \
  -H "Authorization: Bearer $OP_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"preset": "eu_ai_act_art50", "items": [{"asset_id": "ast_…"}, {"object_id": "aob_…"}]}'

The answer is 202 with a marking_check_id (mchk_…). Poll GET /v1/marking-checks/{marking_check_id} (scope verifications:read) until state is completed or failed, with the same backoff as a verification. An unknown preset, too many items, or an asset or object of another account is 422 MARKING_CHECK_INVALID; an unknown id is 404 MARKING_CHECK_NOT_FOUND. Replaying the same Idempotency-Key with the same body returns the same check, even when the two requests arrive at once; a different body under the same key is 422 MARKING_CHECK_INVALID.

Each item runs one verification, recorded as one verification charge in your usage ledger (see Billing: charges are not yet reported to the payment provider and are invoiced from a later date).

Requirements

Requirement	Passes when
c2pa_manifest	A C2PA manifest is present and its signature and hard binding validate.
ai_declaration	The manifest declares a generative digitalSourceType.
recoverable_watermark	Our watermark was recovered from the file and resolves to one of your assets, or a watermark was recorded at publication for this exact output.
recoverable_fingerprint	A stored fingerprint of one of your assets was matched against the file, or a fingerprint was recorded at publication for this exact output.
machine_readable_ai_marking	ai_declaration passes, or our watermark was recovered from the file, or a watermark was recorded at publication for this output and c2pa_manifest passes (the hard binding shows the bytes are unchanged).

Recovered or recorded

The names recoverable_watermark and recoverable_fingerprint overstate what is usually checked. Recovery runs only when the manifest does not validate. For an unaltered output whose manifest validates, these requirements pass on what Original Pictures recorded when it published that output, not on a watermark decoded from, or a fingerprint matched against, the file. Each passing requirement says which in evidence:

• "recovered": the signal was decoded or matched from the checked file.
• "recorded": the requirement rests on a binding recorded for this exact output at publication.
• null: the requirement was read from the manifest itself, or it failed.

The detail sentence says the same in words. machine_readable_ai_marking accepts a recorded watermark only together with a validating manifest: the manifest’s hard binding then shows that the file is byte-for-byte the one that was watermarked and signed. Without it, a recorded watermark proves nothing about the file and does not count. Checks stored before evidence existed report it as null.

Presets

An item passes when every required requirement passes. Recommended requirements are reported, but do not fail the item.

Preset	Covers	Required	Recommended
c2pa_baseline	A C2PA manifest only	c2pa_manifest	none
eu_ai_act_art50	EU AI Act Art. 50(2) machine-readable marking	c2pa_manifest, machine_readable_ai_marking	ai_declaration, recoverable_watermark, recoverable_fingerprint
us_ca_sb942	California SB 942 / AB 853 latent disclosure	as eu_ai_act_art50	as eu_ai_act_art50
cn_labeling_measures	China AI-generated content labelling: implicit label	c2pa_manifest, ai_declaration, machine_readable_ai_marking	recoverable_watermark, recoverable_fingerprint
kr_ai_framework_act	Korea AI Framework Act Art. 31 marking	as eu_ai_act_art50	as eu_ai_act_art50
in_synthetic_information	India synthetically generated information: metadata identifier	as eu_ai_act_art50	as eu_ai_act_art50

Because an AI declaration is accepted only on a sign-only call (Declare AI generation), there are two ways to pass the AI-marking presets: sign with an ai_declaration and "watermark": false (fingerprints allowed), or sign with "watermark": true. The China preset requires the declaration itself.

The result

{
  "marking_check_id": "mchk_…",
  "preset": "eu_ai_act_art50",
  "state": "completed",
  "summary": { "passed": 1, "failed": 0, "pending": 0, "errored": 0 },
  "items": [
    {
      "index": 0,
      "asset_id": "ast_…",
      "object_id": "aob_…",
      "verification_id": "vjob_…",
      "outcome": "pass",
      "failure_reason": null,
      "requirements": [
        {
          "requirement": "c2pa_manifest",
          "level": "required",
          "outcome": "pass",
          "detail": "A C2PA manifest is present and validates.",
          "evidence": null
        },
        {
          "requirement": "recoverable_fingerprint",
          "level": "recommended",
          "outcome": "pass",
          "detail": "A fingerprint was recorded for this output at publication; it was not matched against the file.",
          "evidence": "recorded"
        }
      ]
    }
  ],
  "statement": "A technical check of the machine-readable marking these files carry. It is not a legal determination and does not cover visible labels shown around the files."
}

An item’s outcome is pass, fail, pending or error (with a failure_reason); each requirement’s is pass, fail, pending or not_checked.

Gate publication on a check

A hosting platform can upload a file, sign it, and refuse to publish it until the marking check passes. The check reads only files already stored in your account, so the order is: upload, sign, then check the signed output.

# 1. Upload and sign as usual (see the quickstart); keep the signed output's asset id.
ASSET_ID="ast_…"

# 2. Start the check.
CHECK_ID=$(curl -sS -X POST "$OP_API_BASE_URL/v1/marking-checks" \
  -H "Authorization: Bearer $OP_API_KEY" \
  -H "Idempotency-Key: publish-$ASSET_ID" \
  -H "Content-Type: application/json" \
  -d "{\"preset\": \"eu_ai_act_art50\", \"items\": [{\"asset_id\": \"$ASSET_ID\"}]}" \
  | jq -r .marking_check_id)

# 3. Poll until it is terminal (back off between polls).
while :; do
  RESULT=$(curl -sS "$OP_API_BASE_URL/v1/marking-checks/$CHECK_ID" \
    -H "Authorization: Bearer $OP_API_KEY")
  STATE=$(echo "$RESULT" | jq -r .state)
  [ "$STATE" = completed ] || [ "$STATE" = failed ] && break
  sleep 2
done

# 4. Gate on the item outcome.
if [ "$(echo "$RESULT" | jq -r '.items[0].outcome')" = pass ]; then
  echo "publish"
else
  echo "$RESULT" | jq -r '.items[0].requirements[]
    | select(.outcome != "pass") | "\(.level) \(.requirement): \(.detail)"'
  exit 1
fi

Use the same Idempotency-Key for a retry of the same publication so a retry never runs (or bills) a second check. Treat an error outcome as “not checked”, not as a pass. A pass means only that the file carries the machine-readable marking the preset lists; it says nothing about a visible label your page shows, and it is not a legal determination.