# Continuous Learning Insights

> The Learning tab shows what your workspace has learned from reviewer edits and test runs—acceptance rate, edit distance, time to green, flake rate, and a generation-quality score.

Source: https://totalshiftleft.ai/help-center/product-documentation/continuous-learning-insights

## Overview

The **Learning** tab of Rule Intelligence is a read-only dashboard headed **What your workspace learned**. It shows how the platform improves over time by observing your reviewer edits, approvals, deletions, and test-run outcomes — and how those signals change what the generator produces next. Explore the capability at [continuous learning](/features/ai-test-generation/rag-continuous-learning).

## Before you begin

- Open the workspace (project editor → **Requirements** tab, or the standalone Rule Intelligence hub → select a project), then click the **Learning** tab.
- The subtitle reads *"From your team's edits, deletions and test runs over the last N days. Learning changes what the generator produces next."* (the window is typically 30 days).
- **Nulls are normal on day one.** Signals accumulate over time, so metrics without data yet show *collecting data…* rather than a number.
- An **AI provider** (**Settings → AI**) is optional here — learning runs either way, but an AI judge refines ambiguous signals when one is configured.

## Step 1 — Read the KPI cards

Four cards summarize how well generated tests land:

| Card | What it measures |
|------|------------------|
| **Acceptance rate** | Generated tests kept (not deleted or heavily rewritten). Turns red below 50%. |
| **Edit distance** | Average structural change when users edit a generated test. |
| **Time to green** | Median time from generation to the first passing run. |
| **Flake rate** | Tests with both passing and failing runs in the window. Turns red above 10%. |

## Step 2 — Check the generation-quality badge

Next to the header, a **Generation quality NN/100** badge shows the latest golden-set structural evaluation for the current prompt version. It's green when the run passed and red when it didn't. This score is install-wide (not project-scoped), so its absence on a fresh install is expected.

## Step 3 — Understand curation quality

The **Curation quality** card tells you how your learned labels were produced:

- **Heuristic + AI-refined** — *"N ambiguous signal(s) judged by <model>"* when an AI judge has refined signals.
- **Heuristic-only** — shown when no AI provider is configured, with a pointer to *Settings → AI*. Learning still works; the judge just isn't disambiguating suspected product bugs vs. bad tests, or corrections vs. preferences.
- **Heuristic** (default) — the judge refines ambiguous signals during nightly curation when it has something to decide.

## Step 4 — Read "Learned so far"

The **Learned so far** card shows count pills for what curation has produced:

- **endpoints curated**
- **field-name corrections**
- **suppressed test categories**
- **flaky tests identified**
- **objective corrections**
- **team preferences**
- **suspected product bugs** (only when any exist)

Before any curation has run it reads: *"Nothing curated yet. Edit, delete, and run generated tests — the nightly curation job turns those signals into generation improvements automatically."*

## Step 5 — Read the per-endpoint breakdown

When signals exist, a table lists up to 50 endpoints with columns **Endpoint**, **Signals**, **Corrections** (shown as `wrong→correct`), **Suppressed categories**, **Flaky tests**, and **Last curated**. If there are more than 50, a note reads *"Showing first 50 of N endpoints."*

## Step 6 — Run curation on demand

Curation runs automatically on a nightly schedule, but you can force it now:

1. Click **Run curation now** (the button shows **Curating…** while it runs).
2. On success you get a summary such as *"Curated N new signal(s) across N endpoint(s),"* and — when the AI judge participated — *"…N refined by AI judge."*
3. The dashboard refreshes with the new aggregates.

## How it improves generation

The signals shown here — field-name corrections, suppressed categories, flaky-test flags, and reviewer preferences — feed back into future generation so it drifts toward what your team actually accepts. Applied and dismissed suggestions from [Fix tests with AI](/help-center/product-documentation/fix-tests-with-ai) are learning signals too.

## Notes

- This tab never changes tests directly — it's an observation dashboard plus the manual **Run curation now** trigger.
- Empty metrics mean "not enough data yet," not an error; use the workspace normally and revisit after a few review-and-run cycles.

## Related articles

- [Understanding Rule Intelligence](/help-center/product-documentation/understanding-rule-intelligence)
- [Coverage and fidelity](/help-center/product-documentation/coverage-and-fidelity)
- [Fix tests with AI](/help-center/product-documentation/fix-tests-with-ai)
- [Generate endpoint test cases](/help-center/product-documentation/ai-generated-tests)