Sora content rules vs Sume generation_rejected: read the job events

When a prompt that Sora accepted fails on Sume, the job ends as failed with an error category such as generation_rejected, and GET /v1/jobs/{id}/events tells you what happened in order. Do not assume the two systems reject the same prompts: they are different models with different providers, and I found no page that states parity.

OpenAI's video guide, read 2026-10-08, lists its content restrictions: copyrighted characters and music, real people, and content unsuitable for audiences under 18. Sume's docs list job error categories, not a content policy, so test your own prompt set.

Where a rejection lands

Sume documents these job error categories: validation, auth, quota, queue, generation_unavailable, generation_rejected, generation_timeout, runtime_unavailable, worker_timeout and internal. For generation_rejected the docs say to examine the events and correct the unsupported input. The job object from GET /v1/jobs/{id}/status carries an error with category, stage, retryable, retry_after_seconds, public_reason and next_action; the OpenAPI document gives generation_rejected, image_content_rejected and content_policy_rejected as example public_reason values.

Read the timeline

The events endpoint returns a snapshot with a data object holding request_id, job_id and an events array. Each event has id, type, source, status, message, created_at and data. This script prints the timeline for one job; set JOB_ID.

import json, os, urllib.request

def main():
    req = urllib.request.Request(
        "https://api.sume.com/v1/jobs/" + os.environ["JOB_ID"] + "/events",
        headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]})
    body = json.load(urllib.request.urlopen(req))
    for event in body["data"]["events"]:
        print(event["created_at"], event["type"], event["status"], event["message"])

main()

A practical test set

Before you flip traffic, replay the prompts your users sent in the last month, with real-person names and brand names removed from the log, and count how many end as generation_rejected. That number belongs in your replacement scorecard next to latency and cost.

Show users a plain message for this category and keep the raw event text in your logs only. Never auto-retry a rejected prompt; it will be rejected again and, if the job reserved credits, it adds noise to your spend review.