Candidate offers

Candidate offers

The offer gate exists because the choice between fixes is a decision about consequences, and consequences are not visible in a description of an approach. Two candidates described in prose are compared on how well each was written. Two candidates with a blast area each are compared on what they disturb, which is the thing the chooser actually cares about.

Five fields, every candidate

Field	What it holds	Failure if omitted
blast area	the surfaces the change would touch, each with its break time (compile / runtime / silent), plus the deploy ordering and its reason	the cheap-looking candidate wins, and the cost surfaces after it is chosen
what it does not fix	the part of the report this candidate leaves standing, in the reporter's own terms	the fix ships, the report reopens, and everyone believes it regressed
reversibility	what a revert restores, what it does not, and what data written in between survives it	"revertible" is assumed, and a rollback becomes an incident
size	a rung, not an estimate: one file / one surface / several surfaces / crosses a boundary	precision nobody has is offered, and it will be quoted back
evidence	what the candidate rests on from G1 — which finding, at which class	a candidate built on the premise that was refuted at G1

Call the mapping skill once per candidate. It is more work than mapping the favourite, and it is the only version of this gate that produces a comparison rather than a recommendation with decoration. Where a map genuinely cannot be produced for one candidate, say so in that candidate's row: "effects reasoned, not enumerated."

The four candidate classes that must be considered every time

These are the ones a pipeline drops, because none of them is satisfying to say and all four are frequently correct.

1. Already fixed, or retracted

The behaviour changed at a revision after the report was written, or the premise it rested on no longer holds. Requires: the revision that changed it, what it changed, and a re-run of the reproduction table at HEAD. The work is dating, not building.

Say it in the reporter's terms and without triumph — "this was happening; it stopped at this revision; here is what changed" — and check whether the symptom stopped or only the mechanism they named stopped. Those come apart more often than you would like.

2. Do nothing, deliberately

Not "won't fix" and not silence. A do-nothing candidate carries:

• what the report costs if left — frequency, who hits it, what they do instead;
• what would make it worth revisiting: a threshold, a date, a dependency that lands;
• what to tell the reporter, which is the part that makes this an honest outcome rather than an ignored ticket.

A pipeline that cannot produce this candidate will produce a change for every report it is ever handed, including the ones whose correct answer was "this is working as designed and here is why".

3. The precedent this repository already used

The same problem, solved elsewhere in the tree, possibly under another word. A four-step search, in cost order:

1. Name it three ways — the reporter's words, the domain word this codebase uses, and the identifier form (symbol, slug, route segment, column name). Most already-solved problems are missed because they were searched under one of the three.
2. Search the whole tree by name, unscoped. The precedent is usually in a package you would not have thought to look in.
3. Search the history — a closed report, a reverted attempt, a commit message naming the same symptom. A reverted attempt is especially valuable: somebody already learned why the obvious fix does not work.
4. Search work in flight — open branches and open pull requests. The tree is the product minus everything currently being built.

Matching an existing solution costs less to review, less to maintain, and arrives with a blast area someone has already walked. Diverging from one is a legitimate choice — say why in one sentence, because the next reader will ask.

4. The sibling requirement already in the ticket

Reports and requests rarely carry one requirement. When the one that stalled is ambiguous, expensive, or blocked on a limit, look at what sits next to it in the same ticket: an adjacent requirement, usually better specified, frequently delivering most of the value the reporter actually described, at a fraction of the cost and with nothing left to interpret.

That sibling is a real candidate, ranked in the same table as the others — not a consolation prize produced after the ambiguous one fails, and not a scope reduction smuggled in as a simplification. It carries the same five fields as everything else, including what it does not fix, which is where the part of the ask it drops has to be stated plainly and in the reporter's own terms.

Two rules keep it honest:

• The reporter or owner chooses it, not the agent — swapping one requirement for its neighbour changes what gets delivered, so it goes through the same choice as any other candidate, above the light band.
• The stalled requirement stays on the table. Offering the sibling is not a verdict that the first one is infeasible; where the block was a limit, that claim needs the limit decomposed at G1 before it can be made at all.

The comparison table

Rank by consequence, not by elegance. One table, candidates as rows:

candidate | fixes | does NOT fix | surfaces touched (break times) | reversibility | size | evidence

Then one sentence: which you would take and why. The recommendation is welcome; it is not the gate. The gate is that somebody else could reach a different conclusion from the same table and be right.

Two to four candidates. Fewer than two is not an offer. More than four is usually two real ones and a decorative spread — and it pushes the decision back to the reader by volume.

Recording the choice

The chosen candidate is recorded with who chose it and when: the reporter, the owning person, or the agent under the light-band default. This single field decides what the resolution note says at the end and whether a surprise at G4 is a conversation or an accusation. Where the confirmation policy made the choice the agent's, record the one-line notice that was given.

Anti-patterns

• The strawman spread. One real candidate flanked by two nobody would pick. It looks like a choice and functions as a recommendation.
• Candidates that differ only in effort. Small / medium / large versions of one approach. That is a sizing question inside one candidate, not three candidates.
• The candidate whose blast area is "unknown". Either map it or say it is reasoned; an empty cell reads as a small blast area.
• Ranking by how clean the diff would be. The chooser is not going to read the diff. They are going to live with what it disturbs.
• Dropping "what it does not fix". The most useful field on the table, and the first one to disappear under time pressure — because writing it means admitting the fix is partial, which is exactly what the reporter needs to know before it ships.