Local Data Sentinel

Use the deterministic engine as the source of truth. The optional OpenVINO model may explain calculated Evidence, but must never decide anomalies, thresholds, groups, or causes.

Usage

Only call scripts\run.ps1; never call internal Python modules directly. Input must be one local .csv or .xlsx path. --sheet is optional for XLSX; without it every non-empty sheet is analyzed. Set LOCAL_DATA_SENTINEL_MODEL_DIR only when the model is outside an automatically discovered local model directory.

& "<skill-root>\scripts\run.ps1" analyze --file "$PWD\data.csv" --no-ai-summary
& "<skill-root>\scripts\run.ps1" analyze --file "$PWD\book.xlsx" --sheet "Sheet1" --ai-summary --device CPU
& "<skill-root>\scripts\run.ps1" analyze --file "$PWD\data.csv" --ai-summary --offline-strict
& "<skill-root>\scripts\run.ps1" status
& "<skill-root>\scripts\run.ps1" shutdown

If a first-run online model operation exits with code 3, call scripts\run.ps1 --continue after the download finishes. Exit codes: 0 success, 1 input/general error, 2 client/server communication error, 3 model download pending. Never retry malformed/unsupported input silently.

Interpreting results

Read the final stdout JSON. Use summary_json for the concise result, evidence_json for traceable claims, and report_md or report_html for the user-facing report. Summarize with Evidence IDs. State that candidate explanations are associations, not causes. The Agent does not recompute anomalies itself.

Successful output includes summary.json, evidence.json, report.md, report.html, the row/group/candidate CSV artifacts, charts, and run_manifest.json. On failure, return the final stdout error JSON and its actionable message; do not fabricate an analysis result.

Boundaries

• Keep raw CSV/XLSX and generated artifacts local; never upload them or use a cloud fallback.
• The optional local AI dependency is an OpenVINO GenAI IR model. The LLM receives only deterministic evidence_records; it never receives the source table, raw rows, source path, or report context.
• In --offline-strict, do not download a model. If the local model is absent, deterministic analysis still completes.
• Do not present the optional AI summary as FACT. Unknown Evidence references invalidate it and trigger deterministic fallback.
• Unsupported or malformed input must be reported clearly; do not hide failures.