[nick-developer]

Requested by Cod3r · project thread

This PR runs relsynth autotune on AdventureWorks and fixes the autotune and scorecard issues that run exposed. The report is docs/research/autotune-adventureworks.md.

Result of the AdventureWorks search

• The search ran in two 5-hour CI jobs (64 runs; seeds 1-3 for the search, 11-12 held out). Every hard constraint and every G1-G14 identity held on all five seeds.
• The only setting accepted was segments.default_cap: 16, and the held-out seeds confirm it. Excess failures per seed fell from 43.7 to 42.0 on the search seeds and from 48 to 43 on the held-out seeds.
• No setting moved most of the remaining misfit, so it needs engine or spec work:
  • novel_combination_rate: 22 excess failures per seed. Synthesised dimension rows break the support of pairs two hops away (for example FactResellerSales employee territory with reseller geography, 0.87 new combinations). A link_affinity spec experiment did not change it.
  • Reseller/product pair repeats.
  • Freight/TaxAmt and UnitsBalance marginals.
  • Order codes typed as categoricals.
• The doctor's set of 13 fixes cut excess failures by 7 per seed but broke a privacy check, so it was rejected. This PR now tries its parts one block at a time.

Changes

• relsynth autotune --workers N: the seeds of a candidate run in parallel processes. The records and the search path are the same as with one worker (test).
• tools/autotune_bench.py, bench/autotune_domain.py and .github/workflows/autotune.yml run autotune on a bench database on CI. The workflow starts on dispatch or on a push with [autotune], and a push with [autotune-resume:<run>] continues an earlier search. The report and a per-candidate summary are printed to the log.
• Scorecard: a check rule the source itself breaks gets a three-standard-error allowance instead of an exact limit. On AdventureWorks, 9 vs 12 of 60,855 rows was counted as a hard violation.
• structure.structural_prior is now bounded by 0.999 in the registry, because the spec refuses 1. Before this, 6 runs failed. A new test checks every tunable bound against spec validation.
• The report counts a trade-off only when the change is beyond seed noise.
• A rejected doctor set is now retried block by block.
• The autotune history is keyed by an engine signature (relsynth source hash plus numpy/pandas/scipy/numba/duckdb/SQLAlchemy/pyarrow versions). In runs 3 and 4, a pandas 2.3 incumbent from run 1 was compared with pandas 3.1 candidates, so every move looked like a DimProduct DCR privacy break.
• docs/settings-reference.md is regenerated, and the CHANGELOG has an Unreleased entry.

Checks

• ruff check src tests tools bench is clean. The mypy ratchet reports 0 new errors.
• pytest tests/unit -m "not slow": 638 passed, after merging main (2.0.0).
• The latest push starts a fresh AdventureWorks search (run 5) on CI under one environment. The report will be updated with its result.

https://claude.ai/code/session_01PSVqwFW6fhyVohUnuKd2fu

---

Generated by Claude Code

[nick-developer]

The dco check fails because two commits on this branch have no Signed-off-by line: 46af8f3 (fix: history keyed by engine and library versions) and 77d2f21 (docs: runs 3-4). I forgot -s on them. The code is fine. Every other check passed on those commits.

The only fix is to rewrite those two commits with a sign-off and force-push the branch. I have a local copy ready with the same file tree as 2ae8294, and your merge of main is kept. A force-push needs the owner's go-ahead, so I have not pushed it. Run 5 of autotune is still running on 46af8f3 and is not affected.

---

Generated by Claude Code