Price false positives and false negatives separately
A hardcoded 0.85 confidence threshold treats blocking a paying customer and missing a policy violation as equally expensive. In real trust-and-safety queues, a leaked scam listing often costs ten times more than holding a borderline post for a 30-second human check.
Adjust the violation base rate, error costs, and moderator hourly rate above to see how shifting threshold changes both your human review queue and total monthly operating cost.
Confidence threshold guide · OpenAI Moderation API comparison
Why custom policy criteria beat fixed taxonomy endpoints
Generic moderation endpoints only score universal harm categories and miss marketplace-specific rules like off-platform payment links, competitor coupon scraping, or B2B compliance disclosures. Pairing a noul policy check with a three-way routing choice lets your team enforce its exact written policy and route uncertain items to moderators.
| Threshold band | Typical queue behavior | Human review load | Best fit |
|---|---|---|---|
| 0.70 – 0.79 | High automation, higher boundary error exposure | Low (8% – 15%) | Low-stakes community tags |
| 0.80 – 0.89 | Balanced automation with explicit uncertainty triage | Moderate (18% – 30%) | Marketplace UGC & support |
| 0.90 – 0.98 | Strict safety gate; borderline items always escalated | Higher (32% – 55%) | Fraud, billing & compliance |
Validate your chosen cutoff on a labeled holdout set
Use the curl command generated above to test your policy on 100 to 500 historical items your moderators already adjudicated. A $29 Top-up pack covers 1,800 live evaluations with no recurring subscription, so you can measure empirical precision before automating a single user action.
Try live in Playground · Start with $29 Top-up or $39 Builder
Moderation threshold calculator FAQ
What happens when Jev omits a confidence score on an item?
When confidence is null or below your threshold, Jev sets needsReview to true and never invents a fallback allow verdict.
Does this calculator use live production benchmark data?
No. The calculator models expected cost mathematically from your inputs; always measure actual precision and review rates on your own labeled traffic.
