sentryskills-output

SentrySkills Output Privacy Guard

Role

As mandatory gate before final output to prevent "explanatory answer leakage."

Input

1. Text to be output
2. sensitivity_state
3. trust_annotations
4. Whether multi-source verification completed

Output

1. leakage_detected
2. redaction_applied
3. confidence_level
4. safe_response
5. output_decision (allow|downgrade|block)
6. output_rule_action and optional output_model_action

Mandatory Checks

1. Credentials, keys, privacy fields
2. Restatement and reconstruction leakage of sensitive context
3. Whether single-source tool conclusions are over-deterministically expressed
4. Multi-source verification status

Disposition Rules

1. Sensitive leakage detected: redact or block.
2. Multi-source verification not completed: downgrade, output uncertainty notice.
3. Multi-source consistent and no sensitive leakage: can allow.
4. Output rules run before output model analysis; model results may tighten but not relax rule decisions.
5. Prompt-injection phrases in output text are not blocking by themselves unless they create leakage, unsafe reconstruction, or executable dangerous content.

Output Template

## Output Guard Result
- output_decision: <allow|downgrade|block>
- leakage_detected: <true|false>
- redaction_applied: <true|false>
- confidence_level: <low|medium|high>
- safe_response: <final text to user>

Recommended Script Calls

1. If trust_annotations contains single-tool source conclusions, first call ../shared/scripts/verify_multi_source_template.py.
2. When conclusion has not reached multi_source_verified, output_decision should not default to allow.