Skip to main content
The Improvement history is the record of every improvement ever proposed, rejected, deployed, or rolled back. It’s how you know what your Flow’s quality trajectory has been — and what to try again vs. avoid.

​
The history view

Reliable → Optimization → History lists every improvement:
  • When — created / approved / deployed / resolved timestamps.
  • What — one-line title.
  • Lever — which category of change.
  • Cluster — the signal cluster it addressed.
  • Status — proposed / simulated / approved / deployed / rolled back / resolved.
  • Result — did it work? (pass rate delta on the cluster, holdout impact, regression count).
Filter by Flow, by status, by lever, by author.

​
Improvement detail

Click any improvement for the full record:
  • The problem — cluster description, sample signals.
  • The diagnosis — root cause note.
  • The proposals — all candidates considered.
  • The simulation — before/after numbers.
  • The approval — approver, timestamp, comments.
  • The deploy — target version.
  • The follow-up — did signals in this cluster reappear? When?

​
Success rate

At the top of History:
  • Deployed improvements: 47.
  • Still resolved (no regression): 42.
  • Regressed within 30 days: 3.
  • Rolled back: 2.
Success rate = deployed that stayed resolved / total deployed. A healthy Flow has 85%+ success rate. Low success rate suggests:
  • Simulation is under-representative — the dataset misses too much.
  • Proposals are too aggressive — pattern-match to specifics, don’t generalize.
  • Regressions aren’t well-guarded — critical examples missing from the regression suite.

​
Improvement patterns over time

Chart view: improvements shipped per week, grouped by lever.
  • Rising rate — active fix cadence. Healthy for growing Flows.
  • Falling rate — either fewer issues (good) or team paying less attention (bad — check signal queue for stale confirmed signals).
  • Concentration on one lever — you’re maxing out that lever. Consider whether other levers deserve attention.

​
Comparing improvements

Pick two improvements → Compare. Shows:
  • Diffs side-by-side.
  • Their simulation deltas.
  • Whether both were on the same cluster or different.
Useful for understanding “this improvement fixed the pattern; this one didn’t” — the diff often shows why.

​
Improvement recipes

If a specific lever + change pattern keeps working, save it as a recipe. Recipes are reusable proposal templates:
  • “Add tag boost recipe” — when a cluster suggests a tag preference, generate a preset change automatically.
  • “Add rule recipe” — when a cluster shows a consistent behavior gap, generate a memory rule.
Recipes accelerate future improvements: the Improvement flow starts from a recipe instead of proposing from scratch. Save any deployed improvement as a recipe: right-click → Save as recipe.

​
Rollback history

Every rollback is prominent in the history:
  • Which improvement was rolled back.
  • When.
  • Why (rollback reason, mandatory).
  • What was deployed after (usually a corrected version).
Rollback isn’t a failure — it’s a safety valve. But repeated rollbacks on the same cluster mean you haven’t yet found the right fix.

​
Export

Full history exportable as CSV or JSON. Useful for:
  • Weekly team review — what shipped, what stuck.
  • Quarterly quality report — improvement volume, success rate, cost trends.
  • Compliance audit — evidence of change control processes.