Auto-Learning Analytics
Auto-learning turns your corrections into enforced rules: the hook detects correction-shaped prompts, and /massu-rule approve promotes them. But once a rule is promoted, two questions matter: are people actually approving the candidates that get proposed, and do the approved rules actually prevent the bug class they were promoted to enforce? The Auto-Learning Analytics dashboard answers both — for your whole team, in one place.
It is a Team tier feature, because the value is org-scoped: a single seat has no funnel to aggregate. See License Tiers for the full ladder.
Why This Matters
Without a funnel view, auto-learning is a black box:
- You can see individual candidates locally, but not whether the team is accepting or ignoring them
- A promoted rule that keeps getting bypassed looks identical to one that works — until the bug recurs
- There is no signal for "this rule is noise, dismiss it" versus "this rule is load-bearing, keep it"
The dashboard makes the whole lifecycle observable, using only metadata — rule bodies and prompt text never leave your machine.
The Promotion Funnel
Every rule candidate moves through five stages:
| Stage | Meaning |
|---|---|
| Proposed | The hook detected a correction-shaped prompt and emitted a candidate |
| Shown | An operator opened the candidate's preview |
| Approved | The candidate was promoted to an enforced rule |
| Dismissed | The candidate was rejected (and its signals down-weighted) |
| Revoked | A previously-promoted rule was withdrawn |
The dashboard shows the counts at each stage plus your team's acceptance rate (approvals as a share of decisions) over time — so you can tell whether candidates are converting or piling up unread.
Rule Effectiveness
For each promoted rule, the effectiveness table shows how many times its bug class recurred after promotion:
0recurrences means the rule is doing its job- A non-zero count means the rule was promoted but the bug class came back anyway — a candidate to strengthen or replace
This is the same recurrence signal the framework already tracks locally; the dashboard simply surfaces it org-wide and ranks the rules that need attention first.
Privacy
The funnel is built from metadata-only events — the stage transition, a score, a destination category. The correction prompt, the rule body, and any free-text reason are never transmitted. Events are captured only on Team seats and are scoped to your organization.
How To Use It
- Open the dashboard and select Auto-Learning in the sidebar.
- Read the funnel: a low acceptance rate usually means candidates are too noisy — dismiss aggressively to down-weight their signals.
- Sort the effectiveness table by recurrences to find rules that were promoted but did not hold, and strengthen them.