Pipeline Stages

Every verification pipeline is composed of ordered stages. Each stage type has different characteristics, latency profiles, and output shapes. Understanding how stages work helps you interpret verdicts, estimate run duration, and choose the right pipeline tier for your use case.

Stage Types

Type	Who runs it	Latency	Requires crowd workers
AI	Local LLM	Seconds to minutes	No
Crowd	Vetted human workers	Minutes to hours	Yes
Expert	Specialist crowd workers	Minutes to hours	Yes
Synthesis	Local LLM	Seconds	No

Stages execute sequentially. Each stage receives the file's enrichment metadata, any supplied verification context, and the results of all prior stages. This means later stages can build on earlier findings rather than analysing in isolation.

AI Stages

AI stages call a locally-hosted LLM with a context-aware prompt assembled from:

The file's enrichment metadata (EXIF, codec info, loudness levels, etc.)
Any verification context key-value pairs attached to the run
The verdicts and summaries from all prior stages

The LLM output is validated against a strict structured JSON schema before being accepted. If the output fails schema validation, the stage is retried automatically up to a configured limit. On persistent failure the run transitions to failed.

A successful AI stage writes the following to its stage record:

verdict — stage-level verdict (authentic, manipulated, synthetic, inconclusive, or unverified)
confidence — integer 0–100
summary — one or two sentence human-readable description of the finding
findings — optional array of structured observations with severity values (info, low, medium, high, critical)

Crowd Stages

Crowd stages distribute an analysis task to eligible vetted workers. The pipeline pauses at a crowd stage until the response threshold is met.

Key parameters for every crowd stage:

Parameter	Description	Typical values
`minResponses`	Minimum worker responses required before the stage can complete	2–5
`consensusThreshold`	Percentage of responses that must agree on the same verdict	60–100%
`workerTags`	Required worker specialization tags	See table below

When a pipeline reaches a crowd stage, the overall run status transitions from running to awaiting_crowd. The pipeline remains paused until both minResponses and consensusThreshold are satisfied, at which point it automatically resumes.

Crowd stage latency depends on worker availability at the time of the request. Tier 2 pipelines (which include one crowd stage) typically complete within the duration estimate shown in the pipeline catalog. Tier 3 pipelines (which include an expert stage) may take longer.

Worker Tags

Each crowd stage specifies the workerTags required to participate. Workers must hold at least one of the listed tags to be eligible for the task.

Tag	Specialization
`general`	General-purpose content review
`image_forensics`	Image manipulation and forensic analysis
`video_forensics`	Video deepfake and temporal analysis
`native_speaker`	Language-specific audio review
`fact_checker`	Journalistic fact-checking and source verification
`identity_verification`	Background check and credential verification

Expert Stages

Expert stages are a specialized variant of crowd stages configured for high-stakes verification. They use stricter thresholds than standard crowd stages:

minResponses: 2 — exactly two expert workers must respond
consensusThreshold: 100% — both experts must agree; if they disagree the stage outcome is inconclusive

Expert stages appear in pipelines where errors carry significant consequences — for example, the Expert Review stage in verify-image-deep and the Expert Assessment stage in verify-identity-claim. Because the pool of qualified experts is smaller than the general crowd, expert stages may take longer than standard crowd stages.

Synthesis Stages

Synthesis stages are the final stage in every pipeline. They are run by the local LLM and have access to the complete record of all prior stages — every verdict, confidence score, summary, and findings array.

A synthesis stage produces:

The final pipeline verdict — the single authentic, manipulated, synthetic, inconclusive, or unverified value returned on the top-level run object
A confidence score — aggregated from all stage scores
A scorecard — a structured breakdown of confidence by analysis dimension (e.g. metadata_consistency, compression_artifacts)
A human-readable explanation — a prose summary suitable for display to end users or inclusion in audit reports

Run Status Lifecycle

A verification run moves through the following statuses from creation to completion.

pending — The run has been created and the job has been enqueued. No stages have started yet.

running — Stages are actively executing. AI stages complete within seconds to minutes. The run stays in this state for the duration of all AI-only stages.

awaiting_crowd — The pipeline has reached a crowd or expert stage and is waiting for the minimum number of worker responses. This status only appears in Tier 2 and Tier 3 pipelines.

completed — All stages have finished, the synthesis stage has produced a verdict, and the run record is fully populated. This is the terminal success state.

failed — A stage encountered an unrecoverable error (for example, the LLM returned invalid output after all retries, or a required context key was missing). The error field on the run describes the failure. This is the terminal failure state.

Poll the run endpoint to check the current status:

curl https://api.crowdee.ai/v2/projects/{projectId}/verification-runs/{runId} \
  -H "X-API-Key: crw_YOUR_API_KEY"