Governance & analytics
AI can propose; humans decide. And everything that matters leaves a trace.
Provenance & the approval queue
Every rule and skill carries its provenance: human (created in the admin) or AI (proposed by an agent). When Require AI content validation is on (the default), AI-proposed content lands in a queue in Settings → MCP & agents with Approve / Reject buttons — it is not served to agents or used by the gate until approved.
The audit log
Settings → Audit log records who did what, when — user (by email), agent (by token) or system. Events covered:
- Rules & skills: create, update, delete, restore, AI proposals
- Workspaces: create, delete (with the chosen fate of their rules)
- AI provider config: update, delete
- MCP tokens: create, revoke
- Governance: setting changes, approvals, rejections
- Playbook exports and plan-limit hits
Actor and target names are snapshotted, so entries stay readable even after a user or rule is deleted.
The dashboard
| Metric | What it tells you |
|---|---|
| Compliance rate | Share of analyses without violations, over 24h / 7d / 30d / custom, filterable per workspace. |
| Violations by severity & top rules | Where the playbook actually bites — and which rules drift most. |
| Pre-commit fix rate | Per rule: how often agents tripped on it, and how often they fixed it before committing. The shift-left metric. |
| Rule reliability | Per rule, all-time: share of feedback marking a violation false positive. A rule crossing the evidence threshold is flagged “Needs attention” — it's producing enough noise to be worth rewording or rescoping. |
| Token usage | Input / cached / output tokens per provider over time — your BYOK bill, visible. |
| Skill usage | Which how-tos agents actually load. |
False-positive feedback
On any analysis, each violation can be marked false positive (or re-confirmed) with one click. Confirmed false positives are fed back into the next analysis for that rule and file — the model sees “a human already reviewed a similar case here and rejected it” as context, which measurably reduces repeat noise on genuinely ambiguous cases. It does not blanket-disable the rule on that file: a real, different violation still gets caught. Feedback is also tracked per rule for the long run: the dashboard's Rule reliability card surfaces the false-positive rate per rule and flags the ones that have crossed the evidence threshold. From there, editing a flagged rule offers a Suggest rewrite button: it reads the rule's whole feedback history — every false positive and every confirmed violation, never a single case — and proposes a reworded statement that would have avoided the noise without missing what was already confirmed real. It never applies on its own; the suggestion lands in the edit form for a human to review, adjust and save like any other change.
Plan limits
Limits apply at creation time only — a downgrade never locks you out of existing data. Current limits: Free 1 workspace · Team 3 workspaces, 10 members · Agency unlimited workspaces, 50 members · Enterprise unlimited. Analyses are unlimited on every plan.