docs(seal-reward): state the OpenAI RFT path as the grader docs allow

[abrichr]

What changed

docs/commercial/seal-reward.md, section "OpenAI Agent RFT grader", is now "OpenAI grader shape". The old text said OpenAI documents one custom-grader contract and that a hosted job can't reach the worker. The new text names the six grader types OpenAI documents, quotes the no-network rule for the python grader, states that the guides describe no tool endpoint that carries a result into the graded sample, notes the fine-tuning wind-down, and presents POST /v1/graders/openai as the local mirror of grade(sample, item) for a self-hosted loop. The TRL, verl, and Prime Intellect sections are unchanged.

OpenAI documentation read on 2026-09-01

Graders guide, https://developers.openai.com/api/docs/guides/graders

Your uploaded code must be less than 256kB and will not have network access.

The python source code must contain a grade function that takes in exactly two arguments and returns a float value as a grade.

OpenAI is deprecating graders as part of the evals and fine-tuning workflows they support.

The guide documents string_check, text_similarity, score_model, python, and multi. The API reference (https://developers.openai.com/api/docs/api-reference/graders) adds label_model. No type takes a URL or endpoint field.

Reinforcement fine-tuning guide, https://developers.openai.com/api/docs/guides/reinforcement-fine-tuning

OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months.

Configure your grader to assign rewards based on the contents of the tool calls made by the model.

The guide has no page for Agent RFT (https://developers.openai.com/api/docs/guides/agent-rft returns 404) and no sentence about a tool endpoint the model calls during a rollout.

Opened by an agent session, not the founder.

🤖 Generated with Claude Code