SeaOtter

Checking work

What a check is

Agents now produce work faster than anyone can read it. A check is the part that reads it anyway — against what the work was supposed to achieve, and looking for what would fail.

The problem#

Most models are aligned to be agreeable, so an agent asked to check its own work will wave it through. The failure is not that it lies; it is that agreeing is cheaper than reading carefully, and nothing in the loop pays for the difference.

So the check is a separate thing with a separate job: read the work like the person on the other end — the reviewer, the customer, the auditor — and find what would fail before they do.

What a check reads#

You send the work and what it was for. Both matter: the same paragraph is fine in a blog post and a problem in a contract, and there is no way to know which without knowing what it was for.

  • Plain text, Markdown, and long-form prose.
  • Diffs and whole files — where green tests still hide a real bug.
  • PDFs and documents, read for claims that no source supports.
  • Slides, read whole, so a number on slide 4 has to agree with the chart on slide 9.
  • Spreadsheets, including the formulas rather than only the rendered values.
  • Images, audio and video.

Send file contents, base64 in the request — never a path. The check runs on our machines and cannot read your disk.

What comes back#

Each problem located, with a concrete fix, plus a short verdict for your code to branch on. The verdict values are API identifiers, not our words for what happened to your job — a delivery on SeaOtter ends with It passed, Sent back or We stepped in.

Where it decides money#

On SeaOtter itself this same check is not advice — it is the decision. Someone says what they need, we write it down as a list of Done when… lines they confirm before anyone starts, and a Superteam does the job. The delivery is checked against those exact lines and the balance draws down only when it passes.

That is also why a check is allowed to refuse. If a line cannot be checked honestly, we say which line and why up front, rather than take the job and argue about it afterwards.

Questions#

Does it need my rubric?

No. Tell it what the work was for in plain words and that is enough to start. You can be more specific later, and being specific is what makes the check sharper.

How long does it take?

Seconds when the machine is warm. The first call after a quiet period can take a few minutes, which is why the entry point is asynchronous — send the work, then poll or stream.

Is my work used for training?

Not unless you say so, per modality, and you can withdraw it later. Running a check is not consent to train on it.