Firecrawl wins this run.
Best gated accuracy and coverage. The verdict changes when the persona prioritizes speed over editability, so the disagreement remains visible.
Task library/Structured extraction
Extract a 500-page product catalog
Return normalized inventory, variants, price status, source freshness, and failed-page evidence.
Synthetic evaluation using demonstration outputs. Not real customer demand, adoption, or revenue.
OUTPUT VIEWER / RUN 04Firecrawl
{
"sku": "R-2048",
"variant": "bone / 42",
"price": { "value": 185, "currency": "CAD" },
"freshness": "2026-07-12T14:18Z",
"source_status": "observed"
}fieldspagesvariants
What changed the decision
Scores follow the artifact. Every judgment stays attached to a visible state or recorded validator.
Task completion
The winning output completed the primary journey without recovery help in three repeated runs.
Persona fit
Founders chose the fastest usable path; specialists rewarded control and evidence depth.
Failure replay
The lowest-scoring output hid one critical action and lost state after the second step.
Continue the evidence trail.
Inspect the full task library or compare this winner against another model.