Managed Judgments
Managed judgments help SchemaBounce interpret bounded structured state when a fixed rule cannot capture every valid name or business meaning. Depending on the feature, that state can be schema metadata, selected event fields, or an item list supplied to an authorized agent tool. Judgments do not generate free-form content, grant permissions, approve changes, or execute actions by themselves.
This capability is in private preview. Sensitive-column detection is the first implemented feature. Other judgment-assisted features shown in workspace settings are planned and remain unavailable until their product implementation is complete.
The short version
SchemaBounce combines three layers:
- Deterministic rules handle clear cases.
- A managed judgment evaluates the feature's allowlisted state with constrained answers.
- Product policy turns the result into a confidence band and decides whether to flag it, ask a person, or remain silent.
Permissions, workspace settings, budget policy, and human approval remain deterministic SchemaBounce controls.
What customers see
Each feature states:
- what data it can inspect;
- what it never receives;
- what the result can change;
- how uncertainty is shown;
- what happens when the engine is unavailable;
- whether SchemaBounce or the workspace pays for the inference.
Results use bands such as Flagged and Review. Standard product and MCP views do not expose exact probability distributions or provider model identifiers.
Current capability status
| Feature | Status | Purpose |
|---|---|---|
| Sensitive-column detection | Private preview | Identify likely personal, financial, health, authentication, and other sensitive fields from schema metadata |
| Semantic route fields | Planned | Evaluate selected event or item fields into semantic columns, filters, and tests |
| Reflex gates | Planned | Evaluate declared event content before an automation wakes |
| Agent judgment tool | Planned | Let an authorized agent evaluate or rank a bounded item list |
| Schema fingerprints | Planned | Add semantic context to structural schema comparison |
The settings screen reads capability status from the server. A saved switch does not make a planned feature available.
Safety boundary
Every feature has its own data allowlist. Sensitive-column detection and schema fingerprints are metadata-only. Semantic route fields and reflex gates may inspect only the event fields selected in their configuration. The agent tool may inspect only the bounded state supplied to that authorized call. Credentials and secrets are never allowed.
See Data and privacy for the full boundary and MCP access for agent access rules.
Failure behavior
Each feature defines a deterministic fallback. Sensitive-column detection continues with built-in rules when the judgment engine is disabled, unconfigured, or unavailable. The product reports the reduced mode instead of pretending a judgment ran.
Human control
An agent may read an authorized result. Future configuration features may let an agent preview and submit a draft. Agent-authored drafts go to the Inbox and require a person to approve them in the console. A person explicitly accepting a masking suggestion may use the normal confirmed transform-create path. Agents cannot approve or execute their own configuration proposals through MCP.