Detection Method · 03
Every issue has to earn its way onto the reviewer's screen.
An issue backlog that surfaces everything surfaces nothing. The detection engine can find thousands of knowledge risks in a mature plant's work-order history, and will, if unchecked. Sovel gates every surfaced issue through an ROI filter modeled directly on the logic SkySpark's Tariff Engine uses for energy-waste alerts — "only escalate issues whose expected cost clears a threshold" — adapted to knowledge-waste and risk cost rather than kilowatt-hours.
Why this is hard
Fig 1. The composite ROI gate. Each detected issue is scored across four axes (criticality, recurrence, narrative thinness, concentration). The escalation threshold is a configurable plane in that space. Issues beyond the plane (pink) enter the reviewer inbox; issues inside it stay in a ranked backlog.
How we detect it
The ROI gate is not a threshold on a single metric. It is a composite score that multiplies asset criticality, unplanned-labor recurrence, WO-narrative thinness, and expert concentration, then compares against a plant-configurable threshold. Issues below the threshold stay in the backlog, visible but quiet. Issues above it escalate into the reviewer inbox with the composite ROI estimate attached.
-
Composite cost estimation
Each surfaced issue carries an expected-cost estimate derived from four signals: criticality tier, labor-hour recurrence, narrative thinness (proxy for knowledge fragility), and expert-concentration score.
-
Plant-configurable threshold
The escalation threshold is set per plant and per reviewer team. The SkySpark pattern: 'notify only if cost > $X/week.' Sovel: 'escalate only if composite ROI clears the plant's configured filter.'
-
Visible backlog
Issues below the threshold stay visible in a ranked backlog. The reviewer can always drill down; the inbox just does not push them. Nothing is thrown away — low-ROI issues become high-ROI if they recur.
-
Telemetry feedback
Reviewer approvals, edits, and rejections feed the Correction Inference Engine, which over time personalizes the ROI weights to the judgment patterns that reviewer has demonstrated on that asset class.
What reviewers see
The Sovel reviewer inbox with the ROI gate engaged: three escalated issues with their composite ROI estimates, a compact backlog of sub-threshold issues available on demand, and the per-plant threshold configuration visible as a header control.
How we benchmark it
The ROI gate is evaluated end-to-end by measuring reviewer governance-velocity — issues-governed-per-reviewer-hour — against a baseline of ungated issue surfacing. The target is to raise velocity without reducing governance quality, measured as reviewer edit-rate on approved entries.
| Metric | Method | Dataset / Corpus | Result |
|---|---|---|---|
| Governance velocity | Issues governed per reviewer-hour, ROI-gated vs ungated | Internal pilot telemetry; n=4 reviewers | 2.1x lift (targeting ≥2x) |
| Edit rate | Percent of approved entries requiring reviewer edits before sign-off | Same pilot slice | No degradation (≤ baseline) |
| Coverage preservation | Percent of ground-truth issues reachable within 30-day reviewer cycle | Synthetic backlog seeded into FMUCD | 0.94 (gated) vs 0.96 (ungated) |
Governance velocity and edit-rate targets reflect the shipped pilot success metrics in the positioning spine. Coverage preservation will be re-measured at each retainer-month close with reviewer-approved redactions.
Positioning against adjacent tooling
Tariff-gated prioritization is a mature pattern in sensor-first FDD tooling. Sovel imports the pattern into knowledge-risk detection, where it is novel — the first time the pattern has been applied to WO narrative and reviewer-governed capture rather than to energy waste.
| Adjacent Tooling | Their lane | Sovel lane |
|---|---|---|
| SkySpark Spark rules + Tariff Engine | Programmable rules that gate FDD alerts by real utility-tariff cost. 'Notify only if cost > $100/week.' | Same gating logic, different signal. Gates knowledge-risk issues by composite ROI (criticality × recurrence × narrative thinness × concentration) rather than by tariff-cost. |
| Generic CMMS priority flags (P1/P2/P3) | Manual, operator-assigned priority at WO creation. Prone to inflation (everything becomes P1) and does not reflect knowledge-risk cost. | Composite, signal-derived score. Personalizes to each reviewer's demonstrated judgment over time via the Correction Inference Engine. |
| Sensor-first APM (Aspen Mtell, Augury, Tractian) | ROI gate on mechanical-failure cost: repair hours, parts, downtime. | ROI gate on knowledge-waste cost: recurrent unplanned labor, audit exposure, pre-migration risk, expert-dependency risk. |
Frequently asked questions
- Does the ROI gate hide real issues from us?
- No. Issues below threshold remain visible in a ranked backlog — the gate controls what the inbox actively escalates, not what the system knows. Below-threshold issues are one click away, and any issue whose recurrence or cost profile rises will escalate automatically.
- How do we set the threshold?
- The threshold is configured at the plant and reviewer-team level during onboarding. The default is calibrated from your first diagnostic pass — roughly the top-quartile composite-ROI score. You can tune up (narrower inbox, higher velocity) or tune down (broader inbox, higher coverage) at any time.
- Is this just a priority flag?
- Priority flags are manual, coarse, and do not adapt. The Sovel ROI gate is a composite score derived from four measurable signals, personalized per reviewer through the Correction Inference Engine, and governed by the same reason-coded audit trail as every other governed write.
- Why not just show reviewers everything and let them sort?
- Operator ROI filtering is an established pattern for exactly this reason: unfiltered backlogs produce reviewer fatigue and eventually get ignored. The "nickel to save a dollar" framing is how operators already reason about maintenance work. Gating honors that reasoning — it does not override it.
- Does the threshold vary by asset class?
- Yes. The composite is weighted per asset-criticality tier. A narrative-thin closure on a redundant unit is not the same signal as the identical closure on a single-train critical asset, and the gate reflects that.
Where this method came from
The ROI gate is not a Sovel invention. It is a direct port of the tariff-gating logic that SkySpark’s Spark rules and Tariff Engine have productized for sensor-first fault detection — the “only escalate issues whose expected cost > X” pattern that is the quiet workhorse of mature FDD deployments. Paul Kempf described the same instinct in operator voice on 2026-04-21: “a nickel to save a dollar, not a nickel to save six cents.”
What is novel here is the signal the gate operates on. Sovel is the first product to apply the pattern to knowledge-risk cost — narrative thinness, expert concentration, unplanned-labor recurrence — rather than to energy-waste tariffs. The logic is familiar; the substrate is new.
Where this method is going
The threshold is initially conservative — wide inbox, high coverage — for the first two retainer months. As reviewer telemetry accumulates, the Correction Inference Engine begins to personalize the composite weights, and the gate tightens without losing coverage. The path is: reviewer pattern → weight update → velocity lift → coverage preserved. The flywheel compounds across months.
Test the method.
Run the diagnostic on your own work-order export. 48-hour turnaround, no data migration, no seat licenses.