Datasets:
Cross-DBMS verification: PostgreSQL 16.15
Status: the evaluation mechanism was re-run on PostgreSQL for the 12 seed families on 2026-10-19. Full-corpus PostgreSQL evidence is not published in v0.2.0; this document records the portability check and its limits.
What was verified
The published MySQL evidence for the 12 seed families of the library schema was recomputed on PostgreSQL 16.15 from the same inputs:
| Input | Value |
|---|---|
| Corpus | data/corpus/families.json (12 families, unchanged reference SQL) |
| State bank | data/states/states.json, the same 28 states |
| Model answers | the saved runs in data/runs/ (three seed models, unchanged) |
| DBMS | PostgreSQL 16.15 (previous series preflight used the same version) |
| DDL | the corpus DDL, executed without any adaptation |
| Reference executions | 12 families × 28 states = 336 |
| Candidate checks | 12 families × 28 states × 3 models = 1 008 |
For every state the reference and the candidate were executed read-only in a clean
private schema, and the resulting row lists were compared; the per-(model, family, state)
outcome was then compared with the refuting-state sets published in data/evidence/.
Result
- DDL adaptation required: none. The library DDL (
INTEGER PRIMARY KEY,VARCHARlengths, inline foreign keys) and all 12 reference queries executed on PostgreSQL unchanged. - Status mismatches with the MySQL evidence: 0 of 1 008 candidate checks.
The raw check record is stored in cross_dbms_seed_check.json. No per-state rows or pins are duplicated there; the record is a verification log, not a second evidence bank.
Why the comparison is meaningful, and what it does not show
- Equivalence is checked within one engine: the reference and the candidate run on the same DBMS, so engine-specific row order rules do not affect the verdict. MySQL sorts NULLs first in ascending order and PostgreSQL sorts them last, so the ordering of rows in a witness can differ between engines; the refuting state and the multiset difference do not.
- The check covers the 12 seed families only. It shows that the mechanism ports, not that every family in the corpus behaves identically on PostgreSQL. In particular, queries with engine-specific functions would need dialect review before a full PostgreSQL run.
- The dataset format stores one
dbmsper evidence file. Adding a second engine as first-class evidence requires a format extension (separate evidence per engine or adbmsfield per record) and is left for the next version.