Datasets:
Positioning of SREB — corrected
This note exists because our first formulation was wrong and was caught during literature review on 2026-10-19. It is kept in the repository so the mistake is not repeated and so reviewers can see the reasoning.
What we claimed first (wrong)
"SREB is the first dataset of executable evidence for SQL rewrites."
That is false. Two predecessors exist:
| Work | What it already does |
|---|---|
| SQL-RewriteBench (arXiv 2607.09251, 2026) | 180 instances, correctness-gated, full-denominator evaluation, a database and result evidence per instance |
| provensql (Zenodo 10.5281/zenodo.21854387, preprint 08.2026) | for a DIFFERENT verdict returns a concrete executable counterexample database instance as a replayable witness; 213 annotated rewrites, 511 mutations |
So "executable counterexample" is not our novelty. Writing it as ours would be dishonest and a reviewer would find these two papers in minutes.
What we can honestly claim
To our knowledge, SREB is the first public dataset that links model-proposed rewrites to the state violating the schema constraints and to the differing rows, at the level of individual model samples rather than method scores.
The distinction is narrow but real:
- SQL-RewriteBench scores methods on instances; it is an evaluation harness.
- provensql is a verdict tool for individual rewrites: it answers "correct or different" and can produce a witness for that one rewrite.
- SREB records what a model proposed across a corpus, keeps the raw candidate, the refuting state profile and the witness together, and separates refutation rate from rewrite propensity.
What that means for the paper
- The claim in
PAPER-OUTLINE.md§"claim in one sentence" must be replaced. - Related Work must cite both predecessors explicitly, not hide them.
- The contribution should shift from "we have evidence" to "we can say how adequate the evidence is" — i.e. whether a finite bank of states catches the semantic errors that actually occur, and which state profiles are responsible for catching them.
The open question this exposes
If counterexample evidence is not new, the defensible contribution is adequacy: given a bank of allowed states, what fraction of wrong rewrites does it refute? That is a mutation-adequacy question for SQL queries, and it is the direction the paper should take instead of claiming priority on "evidence".