Feedback Insights

CodeStax records reviewer actions on PR-review findings and summarizes them for the current organization. The current feature is an analytics view: it helps teams identify categories that reviewers frequently accept, reject, or mark as false positives.

It does not currently retrain a model, tune prompts, change confidence scores, deprioritize rules, or automatically suppress future findings.

Recorded actions

On a PR-review finding, a reviewer can record one of the supported actions:

• Accepted — the reviewer confirms that the finding is useful or valid.
• Rejected — the reviewer does not accept the finding.
• False positive — the reviewer identifies the finding as incorrect for the reviewed code.
• Helpful — recorded by clients that support the helpful action.

Feedback is organization-scoped and remains associated with its review and finding.

What the dashboard reports

Navigate to Dashboard → PR Reviews → Feedback Insights to view:

• Organization-wide totals by finding category.
• Accepted, rejected, and false-positive counts.
• Acceptance rate, calculated as accepted feedback divided by all recorded feedback in the category.
• The most accepted and most rejected categories.
• Overall false-positive rate.
• A bounded table of the most frequent category, severity, and action groups.

Category totals and rates use the organization’s full recorded feedback history. The event table is intentionally limited for display and does not change those totals or rates.

Interpreting the results

Feedback aggregates describe reviewer behavior; they do not prove that a detector is accurate or inaccurate. Small samples can move rates substantially, so the dashboard marks low-sample data and teams should inspect the underlying findings before changing review policy.

Useful questions include:

• Which categories receive repeated false-positive feedback?
• Do reviewers consistently reject one severity/category combination?
• Is feedback volume broad enough to represent the team, or concentrated among a few reviews?
• Do changes to deterministic rules or prompts improve later reporting trends after an independently validated release?

Current limitations

• No automatic model retraining or online learning.
• No automatic prompt or confidence adjustment.
• No automatic finding suppression or rule deprioritization.
• No per-pattern confidence model or reset controls in this view.
• The event table is a grouped summary, not an audit log of individual feedback records.

Use the findings, analyzer coverage, policy result, and repository tests as the decision evidence. Treat Feedback Insights as a reporting aid, not an autonomous tuning system.