Audit objective
An authorized reviewer must be able to reconstruct what the system knew, what it inferred, which tools it called, what controls were applied, what was communicated and who subsequently changed the decision.
The audit record should preserve decision-relevant artifacts, not hidden model chain-of-thought.
Minimum decision record
{
"case_id": "CASE-2026-000123",
"workflow_version": "case-workflow-v3",
"model_id": "provider/model-version",
"prompt_version": "agri-analyst-v8",
"input_evidence_ids": ["EV-101", "EV-102"],
"tool_call_ids": ["TOOL-501", "TOOL-502"],
"retrieved_source_versions": ["SRC-17:v4"],
"policy_rule_ids": ["RULE-44:v2"],
"structured_analysis_id": "AN-782",
"risk_class": "high",
"decision": "OFFICER_REVIEW",
"response_id": "RESP-991",
"reviewer_actions": [],
"timestamps": {}
}
Provenance graph
Every conclusion should link to one or more evidence nodes. Each evidence node records:
- origin type: farmer, media, sensor/API, internal document, external research, prior case or reviewer;
- immutable identifier and version;
- collection or publication time;
- geographic and temporal applicability;
- authority and approval status;
- extraction method;
- transformations and parent source;
- confidence or quality limitations.
Observability
Use OpenTelemetry-compatible traces across:
- mobile/API request;
- media upload and processing;
- STT and transcript normalization;
- context tool calls;
- retrieval and reranking;
- web research;
- model analysis;
- policy verification;
- TTS;
- persistence and human review.
Correlate all spans through case, session and request identifiers without putting sensitive content in telemetry labels.
Dashboards
Track:
- active and completed sessions;
- weak-network failures and retry success;
- STT latency and critical-entity error;
- image rejection and recapture rate;
- model, retrieval and research latency;
- source-authority distribution;
- unsupported claim blocks;
- policy-rule misses and conflicts;
- escalation volume, age and reviewer workload;
- cost per completed advisory;
- outcome feedback and correction rate;
- errors by district, crop, device and language group.
Alerts
Alert on safety-gate failures, sudden source withdrawal, rule expiry, provider degradation, audit gaps, unusual export activity, rising unsupported-answer rate, queue SLA breach, failed backups and abnormal district-level case patterns.
Forensic reconstruction
A replay utility should reconstruct the case using pinned source, rule, prompt and model versions where technically possible. When a hosted model version is unavailable, preserve its original structured output and document the limitation.
Live media audit
Record session ID, room ID, participant joins/leaves, track state, network mode transitions, recording state, selected-frame IDs, quality results, capture objectives and the exact frames supporting each visual observation. Do not imply raw-video retention when only metadata and selected evidence are stored.