Two must agree

A verdict can read high only when two validated checks of different kinds of evidence agree and their combined score is high enough. One kind agreeing can reach likely at most.

Each of those is needed and none is enough on its own. Two checks of the same kind, whoever makes them, count as one kind of evidence. A check that is not validated for the kind of media in front of it is shown as evidence and never counts toward high.

The same rule is applied to each region of a frame and each stretch of a track that a finding rests on, over that part’s own checks.

Media Verdict console

The rule

In the documents’ own terms: the fifth of the product’s guarantees and the definitions under it.

The lines 0.85, 0.60 and the inconclusive line at 0.40 are the default settings. They sit in a part of the installation’s settings in which a change needs an approver and a second person’s signed acknowledgement, and every change is a recorded version with its reason.

  1. “High” needs two validated sources of different kinds of evidence agreeing.

    “High” needs a combined score of at least 0.85 and at least two validated sources in different cue families agreeing. “Likely” needs a combined score of at least 0.60, at least one validated source agreeing, and no validated source scoring below the inconclusive line. Otherwise there is no certainty. A source agrees when its own calibrated score reaches the line for likely. No positive verdict is ever built on unvalidated sources alone. The only exception is trusted provenance: a valid content credential from a trusted signer, or a hit from a trusted watermark detector.

The one exception

A valid content credential from a trusted signer that records AI actions decides the verdict on its own, without two kinds of evidence. So does a hit from a trusted watermark detector.

In the drawing, file 1 carries a credential (2), its signer is found on the trusted list (3), and the verdict (4) follows from it alone: AI-generated when the credential covers the whole item, partially manipulated, with one region or stretch per recorded action, when it does not. The certainty is high, and the checks’ results are kept as evidence only.

File 5 carries a credential that is not trusted, or one that is invalid. It goes to the evidence (6) and decides nothing. A credential from a signer that is not trusted saying the item was generated, or a platform’s own AI label, is marked as declared by the uploader and can shift the confidence once, never across a certainty level. An invalid or unreadable credential shifts nothing; a phone video re-encoded by a messaging app arrives that way. A watermark of a kind that can be forged is evidence only.

Trusted signers come only from the published C2PA trust list, recorded with its source and date.

Media Verdict console
1 2 3 4 5 6
Illustration

Likely, high, inconclusive

The conditions each needs. For high, likely and no evidence found, all of a row’s conditions are needed; for inconclusive, any one is enough. Meeting them makes the outcome possible, and the procedure’s earlier rules can still decide first. For any positive, the agreeing sources must also include a kind of evidence that the question being answered accepts: a synthetic voice, for example, needs a voice check among them.

OutcomeNeedsIs stopped by
HighA combined score of at least 0.85. At least two validated sources in different kinds of evidence agreeing. On a low-quality item, one more independent kind beyond those. Or: a trusted credential or trusted watermark.Fewer than two kinds agreeing. A pair of validated checks whose errors have not been shown to be independent, which counts as one source until measured.
LikelyA combined score of at least 0.60. At least one validated source agreeing. No validated source scoring below the inconclusive line.A validated source below the inconclusive line. Validated sources that disagree by 0.45 or more, by default, with fewer than two agreeing.
InconclusiveAny of: no check ran; none that ran is validated for this kind of media; the validated checks disagree; the combined score is at or above the inconclusive line but below 0.60; a low-quality item without its extra kind of evidence.Nothing turns it into a negative. The record shows the reason in place of a certainty.
No evidence foundEvery validated source that looked scoring below the inconclusive line.No validated source having looked. A negative needs a validated check that looked and found nothing.

A worked illustration

The site’s standard example record, worked through the rule. The values are illustrative.

Media Verdict console
0.85 0.60 0.91 1 2 3 4 5
Illustration
  1. 1

    One region of a video frame

    The finding rests on one region, so a positive verdict on it is partially manipulated.

  2. 2

    Image detail, validated

    A check of image detail validated for this kind of media agrees.

  3. 3

    Frame continuity, validated

    A check of change over time, a different kind of evidence, also validated, agrees as well. That is two kinds.

  4. 4

    Combined score 0.91

    Above 0.85, the line for high. Had only one kind agreed, the region could have reached likely at most; on a low-quality copy a third kind would have been needed.

  5. 5

    The record

    Partially manipulated, certainty high, probability 0.91, calibrated, two kinds of evidence.

Ask us to work one of your own items through the rule, condition by condition.

Request a demo