Back to AI Engine Overview Permanent Permanent Permanent Until meeting date + 90 days Permanent
PermanentMeeting date + 90 daysPermanentPermanentPermanentPermanent1 year rolling30 days
AI Audit Trail
Decision logging, data retention, and accountability for all AI operations
Every AI decision that affects compliance, finances, or public records is logged and traceable. This page documents what data is retained, how long it's kept, and how to trace any AI-generated output back to its source inputs and reasoning.
Audit Principle
For any AI-generated output visible to users, an auditor should be able to answer: What data went in? What model produced it? What rules constrained it? What human approved it? If any of these questions can't be answered, the AI function needs additional logging.
Audit Categories
Compliance AI Audit
Every conflict analysis, prediction, and compliance assessment is fully traceable.
Traceable Fields
conflictAnalysis.reasoning
Full AI reasoning chain for each conflict identified
conflictAnalysis.severity
AI-assigned severity level with justification
conflictAnalysis.confidence
Confidence score for the analysis
conflictPrediction.predictedConflicts
Pre-meeting conflict predictions with reasoning
conflictPrediction.officerContributions
Contribution data used for prediction (snapshot)
Available Audit Actions
View full reasoning chain for any conflict analysis
Compare AI prediction vs. actual meeting outcomes
Trace which contributions triggered which conflict flags
Export compliance audit reports for FPPC review
Sample Audit Log
| Timestamp | Function | User | Input | Output | Status |
|---|---|---|---|---|---|
| 2:23:17 PM | conflictAnalysis | system | Contribution #4821 vs Agenda Item #312 | severity: medium, confidence: 0.78 | |
| 2:15:42 PM | suggestFundingSources | citizen_1847 | Solution #156 (Park Renovation) | 4 suggestions, max confidence: 0.85 | |
| 1:58:03 PM | refineArgument | citizen_2341 | Argument #892 (Housing Density) | credibility: 72, tags: [housing, zoning] | |
| 1:45:19 PM | generateCommunityReport | citizen_1204 | Proposal #67 (Water Conservation) | voiceScore: 81, evidence: moderate | |
| 1:22:55 PM | conflictPrediction | system | Meeting 2026-04-08 (5 agenda items) | 2 potential conflicts identified | |
| 12:58:11 PM | reshapeToProposal | citizen_3102 | Raw input (287 chars) | Structured proposal contribution | |
| 12:34:28 PM | meetingAnalysis | admin | Meeting 2026-03-25 transcript | 12 topics, 3 compliance flags | |
| 11:47:03 AM | findCollaborations | citizen_1847 | Proposal #71 (Transit) | 3 collaboration matches |
Data Retention Policy
Compliance analysis results
Legal audit requirements — FPPC may review at any time
Conflict predictions
Predictions expire after meeting; retained for accuracy calibration
Funding suggestions
Financial advisory accountability — track suggestion accuracy over time
Proposal AI refinements
Citizen content integrity — always show original alongside AI version
Community Reports
Community intelligence is a public good — preserved for historical reference
Meeting analysis
Government accountability — supplements official records
AI invocation logs
Operational monitoring — older logs archived to cold storage
Raw LLM responses
Debugging and quality assurance — not needed long-term
Pre-Deployment Audit Checklist
All AI functions have system prompts that constrain output format
Financial functions use JSON Schema response_format (strict mode)
Compliance analysis includes full reasoning chain
Community Report stats are overridden with real database counts
AI-refined content stored alongside citizen originals (never overwrites)
Funding suggestions include confidence scores and risk factors
No AI function can auto-publish, auto-file, or auto-apply
Rate limiting configured for all AI endpoints
AI-generated content labeled in UI
Audit log captures timestamp, user, function, input hash, output hashplanned
Cold storage archival configured for logs older than 1 yearplanned
Automated accuracy calibration for conflict predictionsplanned