Evidence Provenance for Transformer AI: Source, Timestamp, Unit, Reviewer

Transformer AI should be able to answer four basic questions before it asks anyone to trust a conclusion: where did this evidence come from, when was it measured, what does the unit mean, and who reviewed the result? Add the method, transformation history, access boundary, uncertainty, and approval state, and the output becomes a record that an engineer can inspect rather than a sentence that disappears into a chat transcript.

This is evidence provenance. It is especially important for DGA, online monitors, SFRA, thermal history, inspections, and work-order records because each source has different clocks, units, methods, and failure modes.

The minimum provenance contract

For every raw observation, capture:

• Asset identity: fleet, site, transformer, component, serial number, sensor, and configuration.
• Source: laboratory, monitor, historian, test instrument, inspection, CMMS, file, record ID, or API endpoint.
• Time: measurement time, sample time, receipt time, import time, time zone, and clock-quality flag where relevant.
• Quantity: value, unit, detection limit, precision, missing-value code, and any normalization.
• Method: sampling, extraction, instrument, test setup, calibration or verification, tap position, load, or operating state.
• Transformation: formulas, resampling, unit conversion, interpolation, baseline selection, rule version, and model version.
• Actor: importing system, agent identity, delegated user, reviewer, approver, and permission scope.
• Outcome: draft, reviewed, accepted, rejected, escalated, deferred, or superseded.

The power transformer diagnostic data model explains why this contract should sit beneath individual diagnostic channels. Do not collapse a DGA lab result, an online monitor point, and an SFRA trace into one score before preserving their distinct provenance.

A commissioning workflow that tests provenance

Before connecting a source, define its record contract. For DGA, list gas names, units, sample date, laboratory identifier, detection limits, and interpretation status. For online monitoring, list gas coverage, monitor ID, heartbeat, calibration or verification record, communications path, and clock behavior. For SFRA, preserve instrument, lead placement, grounding, tap position, baseline, and comparison event; IEEE C57.149 makes long-term storage and repeatable interpretation part of the measurement context.

Run four tests before production use: a valid record, a missing-unit record, a duplicate timestamp, and a record mapped to the wrong asset. The expected behavior is not silent repair. The workflow should retain the raw record, create a visible exception, and require a person to resolve the mapping or quality issue.

At calculation time, store the input record IDs beside the derived value. If a gas generation rate is calculated, preserve the two source observations, interval, unit, formula, and rounding. If an AI summary says “rising trend,” the record should show which observations support that phrase. IEEE C57.104 and IEC 60599 can provide interpretation context; they do not authorize the system to convert a threshold into a trip or maintenance order.

Provenance and AI risk management

NIST’s AI RMF 1.0 is voluntary and use-case agnostic, but its governance, mapping, measurement, and management functions translate well to transformer APM. Govern the owner, permissions, retention, and review path. Map the intended use, affected assets, operating context, and failure consequences. Measure data quality, model or rule performance, reviewer disagreement, and uncertainty. Manage the risk by limiting tools, escalating ambiguous cases, and recording outcomes.

Agent identity matters when the workflow can query systems or stage a CMMS handoff. Log which agent ran, which user delegated the task, which sources it accessed, whether the access was read-only, and what write boundary existed. If evidence may include BES Cyber System Information, consult the responsible security and compliance teams about NERC CIP-011-3, classification, access, and retention.

Illustrative provenance record, not a customer measurement

An illustrative record might read: “TX-203, main tank, hydrogen 118 ppm, laboratory report LAB-2048, sampled 2026-07-22 09:00 UTC, reported in µL/L, method and quality flag present, imported 2026-07-23, compared with LAB-1999, rate calculation v2, draft summary by agent A-04, reviewed by engineer E-17.” This is invented to show field structure. It is not a customer measurement, a fault diagnosis, or a maintenance instruction.

Review and retention

A reviewer should be able to open the source record, see the transformation, inspect contradictions, and edit or reject the draft. The system should preserve the original AI output, the reviewer changes, the decision rationale, and the final status. A later work package should link back to the same evidence rather than copying an untraceable paragraph.

The audit-trail workflow adds the decision and outcome loop. The human-in-the-loop AI article provides the governance pattern. A sample evidence pack is a useful review artifact before real data is integrated.

Product boundary

AgenticGrid Pro is described here as a bounded transformer APM workbench for assembling permitted records, showing provenance, calculating configured indicators, and drafting source-linked review material. ProtectionAI is GridAPM’s separate Windows relay-testing application; a test report can become a provenance-linked input, but neither product grants autonomous OT authority. They do not replace qualified engineers, laboratory analysis, calibration, physical test equipment, or the utility’s CMMS and security controls. Final decisions and OT actions remain engineer-approved.

References

• Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1
• North American Electric Reliability Corporation. (n.d.). CIP-011-3: Cyber Security—Information Protection. Retrieved August 1, 2026, from https://prod.nerc.com/standards/reliability-standards/cip/cip-011-3
• IEEE. (2019). IEEE guide for the interpretation of gases generated in mineral oil-immersed transformers (IEEE Std C57.104-2019). https://standards.ieee.org/ieee/C57.104/7476/
• IEEE. (2024). IEEE guide for the application and interpretation of frequency response analysis for oil-immersed transformers (IEEE Std C57.149-2024). https://standards.ieee.org/ieee/C57.149/7354/
• International Electrotechnical Commission. (2022). Mineral oil-filled electrical equipment in service—Guidance on the interpretation of dissolved and free gases analysis (IEC 60599:2022). https://webstore.iec.ch/en/publication/66491
• CIGRE. (2019). DGA monitoring systems (Technical Brochure 783). https://www.e-cigre.org/publications/detail/783-dga-monitoring-systems.html
• Thang, K. F., Aggarwal, R. K., McGrail, A. J., & Esp, D. G. (2003). Analysis of power transformer dissolved gas data using the self-organizing map. IEEE Transactions on Power Delivery, 18(4), 1241–1248. https://doi.org/10.1109/TPWRD.2003.817733