The draft signal: a contribution that asks for a verdict, not a review

ADR-052: The draft signal: a contribution that asks for a verdict, not a review

Status Active · Filed 2026-08-08 · Issue #29

Context

The pattern appeared ad hoc in ADR-039: the Fire proposal shipped as a draft PR, Proposed on purpose, with a writeup arguing the trade in both directions — “the decision is the merge verdict.” Now the DP→GP migration ships the same way: #55 defines the run itself as the machinery’s acceptance test, so the PR is the experiment and merging it is the judgement. Two uses is this record’s own threshold for writing the norm down — and without a written norm, the draft flag stays ambiguous between “not finished yet” and “finished, awaiting a verdict,” which are opposite states.

Underneath is a distinction between two kinds of PR the record produces:

• Review-shaped: the decision was already made — in an issue, a conversation, a prior ADR — and the PR executes it. Any Proposed documents aboard are riding along to be read; the presumption is merge, and review checks execution.
• Verdict-shaped: the contribution is the question. The choice could not be weighed from a plan; someone had to build it to price it. The presumption is neither merge nor close — that open verdict is the point.

Decision

The draft flag marks a verdict-shaped PR, and it obliges three things:

1. A Proposed decision in the diff — the question being asked, filed as a document, alternatives included. The verdict needs something to land on: merge flips it Active; close files it Rejected, body intact, because a recorded “no” is what stops the question being re-asked.
2. A writeup that argues both directions. What this buys and what it costs, priced from the finished diff (ADR-039’s worth-it writeup is the form). A draft whose writeup only advocates is a review-shaped PR wearing the wrong flag.
3. Rejection stays cheap. Nothing else may stack on the draft while the verdict is open — no follow-on branches, no references from other work that would make closing it expensive. The moment “no” stops being a practical answer, the draft signal is a lie.

When to reach for it: the worth-it experiment (the trade is only priceable built — ADR-039); the first live run of new machinery, where the run is the acceptance test and hand-edits beyond the declared interface would fail it (the DP→GP migration); any change whose author genuinely expects “no” to be a reasonable verdict and wants that priced in.

When not to: agreed work — a draft on work whose decision is already made just slows the loop; and record-only batches, where the documents are the deliverable and the ordinary Proposed-until-merged cycle already carries the verdict. GitHub’s “not finished yet” use of the flag remains what it is; the signal defined here is a draft plus a Proposed decision plus a both-directions writeup, and it reads “finished — judge it.”

Alternatives considered

• No signal — every PR presumes merge. The cost is social, which makes it invisible until it’s paid: closing a finished, working PR feels like waste, so marginal experiments get merged by default and the record accretes decisions nobody would have made from a clean sheet.
• Decide first, then build — settle the ADR in an issue, only build what’s accepted. Right for most work (it’s the review-shaped path), wrong for exactly the cases this signal exists for: a trade like “is Fire worth a dependency” or “is the migration machinery done” has its real price visible only in the diff.
• A label or title convention instead of the draft flag — carries the same information but doesn’t do anything; the draft flag is the one spelling the platform enforces, since it physically disables the merge button until someone deliberately promotes the PR.

Consequences

• The Proposed status gains a second, sharper meaning on draft PRs: not “awaiting the routine verdict” but “the verdict is the open question.” The pending-decisions report already surfaces both, which is what keeps an unjudged draft from being forgotten.
• Closing a draft now carries an obligation: land the decision as Rejected (a record-only commit) rather than letting the branch die silently — otherwise the next person rebuilds the experiment to rediscover the no.
• The DP→GP migration PR is this signal’s first deliberate use; its merge or close is simultaneously the verdict on ADR-040’s rung-2 machinery being done.