Skip to content

feat(analysis): A-4 regression gate AND + richer baseline - #73

Merged
leo-gan merged 2 commits into
masterfrom
feat/a4-regression-gate
Jul 25, 2026
Merged

feat(analysis): A-4 regression gate AND + richer baseline#73
leo-gan merged 2 commits into
masterfrom
feat/a4-regression-gate

Conversation

@leo-gan

@leo-gan leo-gan commented Jul 25, 2026

Copy link
Copy Markdown
Owner

Summary

  • Regression gate default is AND: fail only when the point estimate is more than the threshold slower and the bootstrap CI lower bound still supports that increase (fixes noisy OR false alarms).
  • Baseline schema v2: median, CI, run count, optional samples for Cliff’s δ; keys include batch N + type-config hash.
  • CLI summary counts + reports/regression_report.json; --regression-combine for and|or|….
  • Methodology docs explain the problem and policy in plain language for first-year readers.

Validation

  • Tests: analysis 73 pass; python 29; js 9; rust 13; java OK; cpp roundtrips OK
  • Benchmarks: full mode, all enabled langs (config change forces all)
    • Stems (latest): csharp 2026-07-24-201328, python 201715, rust 201921, c 201935, javascript 201952, go 202008, java 202031, cpp 202051, swift 202115
  • Error CSVs: regression gate clean (0 new error keys)
  • Analysis + dashboard: results/plots regenerated; sync-data.py OK for all 9 languages

Notes

  • Re-save baseline with --save-baseline after merge when you want a clean gate against the new policy (old baselines still load; many host-noise diffs vs pre-shuffle baselines are expected).
  • Escape hatch: regression.combine: or or --regression-combine or for the old, noisier fail mode.
  • Optional dry-run against the existing reports/baseline.json still reports many regressions — that reflects stale multi-session baselines, not a bug in AND logic.

leo-gan added 2 commits July 24, 2026 20:10
Default fail requires both a practical percent increase and CI support
(fixing the old OR false-alarm path). Baseline schema v2 stores median,
CI, n, and optional samples for Cliff's δ diagnostics. Document the
policy in plain language for first-year readers in the methodology.
Refresh language Results, violin plots, and dashboard payloads from the
prepare-pr full runs used to validate the regression-gate change.
@gemini-code-assist

Copy link
Copy Markdown
Contributor

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

@leo-gan
leo-gan merged commit 5bcc739 into master Jul 25, 2026
11 checks passed
@leo-gan
leo-gan deleted the feat/a4-regression-gate branch July 25, 2026 03:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant