code/data/scenarios/README.md

data/scenarios/

Last updated: 2026-05-07

Smart Grid transformer maintenance scenarios, following the AssetOpsBench scenario format. Each scenario is a multi-turn agentic task where an LLM agent must use the IoT / FMSR / TSFM / WO MCP tools to diagnose, forecast, or remediate a transformer fault.

Format

Scenarios follow AssetOpsBench's existing utterance schema with required keys:

• id — unique identifier
• type — domain label (IoT, FMSR, TSFM, WO, or empty for mixed/general)
• text — user instruction for the agent
• category — task category label
• characteristic_form — objective expected answer pattern for grading

For Smart Grid authoring in this repo, we keep additional optional keys:

• asset_id — fictional transformer ID (T-001 to T-020)
• expected_tools — expected MCP tools in rough order
• ground_truth — checkable target answer/action
• difficulty — easy / medium / hard
• domain_tags — exercised domains (IoT, FMSR, TSFM, WO)

See the upstream AssetOpsBench structure in src/scenarios/local/vibration_utterance.json and aobench/scenario-server/src/scenario_server/handlers/*.py (which consume id, type, text, category, characteristic_form).

Targets

• W2 (Apr 7-13): 15+ validated scenarios (reviewer)
• W4 (Apr 21-27): 30+ scenarios (reviewer + team) — stretch goal per mid-point report

Conventions

• File naming: <domain>_<NN>_<short_slug>.json
  • e.g. fmsr_01_dga_arcing_diagnosis.json, tsfm_03_rul_forecast_weekly.json
• Multi-domain scenarios: multi_<NN>_<slug>.json
  • e.g. multi_01_full_fault_response.json (IoT sensor alert → FMSR diagnosis → TSFM RUL check → WO creation)
• Before committing, validate against the AssetOpsBench scenario schema and confirm the referenced asset_id exists in data/processed/asset_metadata.csv.
• Ground truth must be objectively checkable — if scoring depends on subjective judgment, add a scoring rubric field.

Validation

Run the validator from repo root before committing scenario changes:

python data/scenarios/validate_scenarios.py

This catches schema violations and negative-fixture regressions before you get to the heavier harness path. For the full harness workflow, see ../../docs/eval_harness_readme.md.

Status (May 7, 2026)

Canonical package contains 36 validated scenarios: 31 hand-authored scenarios plus 5 generated scenarios promoted after validation and manual review.