Delivery Contract

Why control flow is not evidence

An agent pipeline reports success when its last function returns. That says nothing about whether a paper exists. Observed outcomes that all reported "done": a 12 KB PDF containing a template skeleton, a paper whose every citation rendered as ?, a full draft written with zero literature behind it because the research step had failed silently an hour earlier.

The fix is to stop asking the code whether it finished and start asking the artifacts what they are.

Run these before saying done

Each check reads the delivered files. Hard means the paper does not ship. Soft means it ships with the problem stated out loud.

Evidence-based done

Do not report success on any path that has not produced:

1. a PDF over 50 KB that a parser can open,
2. zero unresolved citations,
3. a non-empty bibliography.

If a run ends without these, it failed, whatever the log says. Say so, say why, and say what the user can do about it. A truthful failure is worth more than a confident lie, and a user who is told "the literature step failed, so this draft has no related work" can act on it.

Disclosure

If an agent wrote it, the paper says so. This is not optional and it is increasingly a submission requirement. One line in the acknowledgements is enough:

This paper was prepared with the assistance of an AI research agent.

Two things to get right. It must survive into the rendered PDF, not just the source, so check the extracted text. And it must not leak internal machinery, tool names, stack traces or file paths into an acknowledgements section a reviewer will read.

Report the outcome honestly

State what passed and what did not:

Delivery: 8 checks passed, 1 soft violation
  page_budget: 10 pages against a 4-page budget (references span 5 pages)

Then let the user decide. Hiding a soft violation to make the summary look clean is the same instinct that produces invented numbers, just wearing better clothes.