How is the 'no ungrounded statements' guarantee enforced - and what does it cost?

[witthayaevans]

draft.py claims every statement carries the exact source spans it came from, and that ungrounded statements are never written. Is that a convention or actually enforced? And what is the downside - are there facts the draft simply refuses to include?

[michellsilva]

It is structural, not a convention, and yes - it deliberately drops things.

1. Enforcement is in the type, not the writer. A Claim cannot exist without at least one Provenance: the dataclass raises UngroundedStatement in __post_init__ if sources is empty, and the renderer only ever prints Claim objects. There is no code path where a sentence reaches the output without citations, because there is no object to carry it. A convention would be a comment asking future contributors to behave; this is a constructor that refuses.

2. The cost is omission. Any fact that cannot be tied to a span - an inference, a summary, a "we believe the root cause was X" without a source line - is not written at all. The tool chooses a thinner draft that is always defensible over a richer one that is sometimes wrong. For a postmortem that circulates to people who were not in the incident, that trade is usually right: every line can be checked by following its citation.

3. Determinism is part of the same design. Line-oriented output diffs cleanly, and identical input produces byte-identical output, so you can review a regenerated draft as a diff instead of re-reading it. The citations are rendered from p.span() - a stable reference into the source, not a page number that shifts when the input changes.

4. What to do with the gap. If reviewers need a claim the tool refuses to make, the fix is to add the evidence (export the dashboard, attach the runbook line) or to write the sentence outside the generated section - not to loosen the guarantee. Keeping the generated draft strictly grounded is what makes the whole document trustworthy; the moment one uncited paragraph is normal, readers stop checking any of them.

So the trade is: fewer statements, all checkable, reproducibly generated. That is the intended product, not a limitation to work around.