Verdicts & Confidence Scores

Every completed verification run produces a verdict — not just a probability score, but a structured, defensible finding that your team can act on and your auditors can inspect. Verdicts are accompanied by a confidence score, a stage-level audit trail, and an optional human-readable explanation.

Verdict Values

Crowdee uses five verdict values. Each has a precise meaning to avoid ambiguity in downstream workflows.

Verdict	Meaning	When used
`authentic`	No evidence of manipulation; all signals consistent	AI and crowd stages agree the content appears genuine
`manipulated`	Evidence of tampering, deepfake, forgery, or fabrication	One or more stages detected significant anomalies
`synthetic`	Content was generated or substantially rewritten by AI	A synthetic-screening stage found strong AI-generation evidence (see Synthetic Data Detection)
`inconclusive`	Mixed or insufficient signals	Crowd consensus not reached, or AI signals conflict across stages
`unverified`	Pipeline could not complete analysis	Missing enrichment, unsupported format, or a stage failure

These values are mutually exclusive — a run produces exactly one verdict. The confidence score conveys how strongly the pipeline arrived at that verdict.

Confidence Score

The confidence score is an integer from 0 to 100. It is computed by aggregating the individual confidence percentages from AI stages and the consensus ratios from crowd stages. Higher scores mean the pipeline's stages were in stronger agreement.

A score of 87 on an authentic verdict means the pipeline arrived at that finding with strong, consistent evidence — it does not mean there is an 87% probability that the content is authentic. Treat the confidence score as a signal quality indicator, not a frequentist probability.

As a rough guide:

Range	Interpretation
90–100	Very strong agreement across all stages
75–89	Strong agreement; minor uncertainty in one or more stages
60–74	Moderate agreement; consider a deeper pipeline or manual review
0–59	Weak or conflicting signals; results should be treated with caution

Stage-Level Audit Trail

Every stage in a pipeline appends its own result to the run record. This gives you full visibility into how the final verdict was reached.

Each stage result includes:

Field	Type	Description
`slug`	string	Identifier for the stage (e.g. `exif-analysis`)
`type`	string	Stage type: `ai`, `crowd`, or `synthesis`
`verdict`	string	Stage-level verdict
`confidence`	integer	Stage-level confidence score (0–100)
`summary`	string	Human-readable description of what the stage found
`findings`	array	Optional structured findings (see below)

When a stage produces findings, each finding has a severity field to indicate its importance:

Severity	Meaning
`info`	Informational observation; no impact on verdict
`low`	Minor anomaly noted; unlikely to indicate manipulation
`medium`	Noteworthy anomaly; warrants attention
`high`	Significant anomaly strongly associated with manipulation
`critical`	Clear indicator of manipulation or forgery

The full stage log is available in the run response under the stages array.

Example Verdict Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "authentic",
  "confidence": 87,
  "pipeline_id": "verify-image-metadata",
  "stages": [
    {
      "slug": "exif-analysis",
      "type": "ai",
      "verdict": "authentic",
      "confidence": 91,
      "summary": "EXIF data consistent with claimed camera model"
    },
    {
      "slug": "format-analysis",
      "type": "ai",
      "verdict": "authentic",
      "confidence": 83,
      "summary": "No evidence of format-level manipulation"
    },
    {
      "slug": "verdict-synthesis",
      "type": "ai",
      "verdict": "authentic",
      "confidence": 87,
      "summary": "All signals consistent; no manipulation detected"
    }
  ]
}

Acting on Verdicts

Use the table below as a starting point for your downstream workflow logic.

Verdict	Recommended action
`authentic`	Proceed with content; archive the run ID for your audit trail
`manipulated`	Flag or quarantine the content; escalate for editorial or legal review
`synthetic`	Label the content as AI-generated before publication, or reject it per your editorial policy
`inconclusive`	Request manual review, or re-run using a deeper Tier 2 pipeline
`unverified`	Check the file's `enrichmentStatus`; retry enrichment if failed, or contact support

Do not automate consequential actions (publication, deletion, legal escalation) on inconclusive verdicts. Inconclusive means the evidence was insufficient to reach a finding — not that the content is safe. Always route inconclusive results to a human reviewer.