Narrow and Deep: An Ontology Tower as the Knowledge of an LLM Agent for an Industrial Equipment System

1. Introduction

• The content and form of knowledge for an agent operating a particular plant have not been measured, motivating a structure that can supply plant knowledge in controlled parts.Existing studies judge a combined package of tag meanings, equipment couplings, procedures, and past incidents against a baseline without it.Source: §1 · p. 2
• Broad ontologies name points across many sites, but no equipment ontology combines one system’s physical derivations and operating history with its points.A single-system agent needs knowledge of how values are derived, the assumptions involved, and what has happened during operation.Source: §1 · p. 2
• Prior findings disagree on which model sizes benefit most from retrieval, and none studied curated equipment knowledge on a physical plant.The paper therefore measures whether small open-weight models gain as much as the largest and whether the form of supply matters.Source: §1 · pp. 2–3
• The study proposes an ontology tower that adds physically derived quantities and journal lessons as knowledge nodes, then tests prompt-projected text across tasks and models.The evaluation uses a real low-humidity air-handling plant, preregistered tasks replayed from its records, and live runs through an invariant safety layer.Source: §1 · p. 3

2. Method

• The method presents the tower, its two forms of depth, its prompt supply and safety layer, and tests using replayed tasks, lesson-form comparisons, and live runs.The evaluation setup and tests are summarized in the section overview.Source: §2 · p. 4

2.1. Narrow and deep knowledge of one equipment system

• An ontology tower describes a few dozen system entities in depth, including controller bindings, units, derivations, assumptions, and operator knowledge.Its node levels are computed from declared dependencies rather than assigned by the modeler.Source: §2.1 · p. 4

2.2. The tower and its height

• The studied tower has 162 nodes and 217 edges; its levels are computed from declared dependency paths above measurement, with level 0 holding tags, assumptions, and knowledge nodes.The tower includes 42 tags, 34 assumptions, 33 bindings, 24 knowledge nodes, 16 derived quantities, and 13 plant elements. Figure 1 depicts node counts at each level, two dependency chains, and interchange-schema node and edge kinds.Source: §2.2 · pp. 4–6, Figure 1 · p. 5

2.3. Depth through physics

• The studied tower contains 16 derived quantities at levels 1–4, including two consistency checks at the top.Source: §2.3 · p. 6
• Derivations encode engineering relations such as psychrometric conversions, mass and energy balances, and fan affinity laws.For example, heater power, measured airflow, and humidity ratio combine into specific dehumidification energy.Source: §2.3 · p. 6
• Assumption links show which derived values rely on each belief and therefore need recomputation if measurement disproves it.Atmospheric pressure enters four derived quantities, while the two flow-sensor assumptions enter six.Source: §2.3 · p. 6

2.4. Depth through operating history

• An operating lesson can record that a normal-status flag hides an abnormal desiccant-unit cycle, linked to the relevant nodes.Source: §2.4 · p. 7
• Knowledge nodes hold operating lessons, equipment automatic logic, and operating envelopes as text, while remaining at level 0.Their links do not count toward tower levels, so they deepen what is known about a point rather than its distance from measurement.Source: §2.4 · p. 7
• Each tower extraction incorporates journal entries written since the previous extraction, making its knowledge specific to the operated system.Of four incident diagnosis tasks, two lessons were already in the plant’s ontology files and two were never incorporated.Source: §2.4 · p. 7

2.5. Supplying the tower to an agent

• The prompt projection presents sensor identity, verification and faults, equipment couplings and logic, knowledge nodes, and control-point envelopes in about 45,000 characters.It is generated from the plant’s descriptive records; 16 derived quantities and their assumptions are in the tower but not in the projection.Source: §2.5 · p. 7
• Query tools let the agent look up a node, follow its edges, and list assumptions, making queried evidence traceable to the returned node.Figure 2 shows the plant, queried tower nodes, and the call sequence beside diagnosis findings and their supporting calls.Source: §2.5 · pp. 7–8, Figure 2 · p. 8

2.6. Bounding the agent’s actions

• Every setpoint command passes through the same invariant safety layer, which checks run-specific tag and range permissions, recent human writes, and plant change guards before writing and recording it.The guards cover absolute limits, step size, hourly and cumulative change limits, and waiting time between changes.Source: §2.6 · pp. 8–9

2.7. Test plant

• The test plant is a daily operated, PLC-controlled low-humidity air-handling system with commercial desiccant, air-handling, and chiller equipment.Its dew-point sensors are preferred over relative-humidity sensors, which are unreliable below about 5% relative humidity.Source: §2.7 · p. 9

2.8. Preregistered evaluation

• Nine record-based tasks comprise five fixed-time reading tasks and four incident-based diagnosis tasks, scored by element share and avoidance of each task’s primary trap.Trials used read-only plant tools, were limited to 60 steps and 1,500 seconds, and two diagnosis incidents were reconstructed from journal values because time-series data were unavailable.Source: §2.8 · pp. 9–10
• The evaluation compares cumulative knowledge conditions and variants that remove or replace knowledge forms, including tower knowledge nodes, journal records accessed through tools, and derived quantities.Table 1 maps evaluation conditions to the tower knowledge they carry.Source: §2.8 · p. 10, Table 1 · p. 11
• Four open-weight models from Qwen3.5-9B to GLM-5.2 (about 750 billion parameters) were evaluated in a 1,304-trial preregistered design with three hypotheses.A blinded LLM judge scored trials; a second judge agreed on primary-trap avoidance with κ = 0.819 over 121 sampled trials, and hypotheses used task-intercept mixed models with Holm correction.Source: §2.8 · pp. 10, 12

2.9. Experiment on the form of the operating lessons

• Each diagnosis task’s own lesson was supplied in three forms while the rest of K4 remained unchanged: as a tower node, a prompt record, or a journal record available through search.For lessons not previously incorporated, a separately run model created the tower node from the incident record.Source: §2.9 · p. 12
• The comparison was run across four diagnosis tasks, four models, and three repetitions, for 144 trials in one batch.For previously incorporated lessons, the prompt and searchable-document versions were made by removing the lesson passages from those sources.Source: §2.9 · p. 12
• The three forms isolate where the task’s incident lesson is kept while holding the other K4 knowledge constant.In tasks without an existing tower lesson, the node version was added from the incident record.Source: §2.9 · p. 12

2.10. Live runs

• The dew-point task sought to raise return-air dew point from below −24.0 °C into −24.0 to −21.0 °C using a humidity setpoint limited to 3%–7%.The temperature task sought a 0.6 °C chamber-temperature rise into a band 0.3–0.9 °C above the preceding hour’s mean, using a supply-air setpoint limited to 15–21 °C.Source: §2.10 · p. 12
• The chamber temperature rises about 0.15–0.2 °C per degree of supply-air setpoint, with an approximately 3-hour time constant; the guard’s 3 °C cumulative limit proved binding in the first temperature run.The permitted 6 °C setpoint range allows about a 1 °C rise, while the guard limit allows about 0.5 °C.Source: §2.10 · p. 13
• Fourteen runs were made—eight dew-point and six temperature—and target-band entry was defined by the controlled variable’s 30-minute moving average.Runs ended on agent completion, time expiry, five consecutive refusals, or a plant safety stop; factual claims were judged against each turn’s evidence.Source: §2.10 · p. 13

3.1. Projected knowledge in the prompt

• 0.605 to 0.802 (+19.8 percentage points): projected tower text raised primary-trap avoidance, supporting H1 (odds ratio 5.45; Holm-adjusted one-sided p = 6.0 × 10−7).Rates pooled over four models and nine tasks; the 95 % confidence interval for the change was +3.0 to +38.4 percentage points.Source: §3.1 · pp. 13–14, Figure 3 · p. 14
• Element-score gains from K1 to K2 were +0.235 for Qwen3.5-9B, +0.178 for gemma-4-31B-it, +0.213 for EXAONE-4.5-33B, and +0.219 for GLM-5.2, with no detected increase by model capacity.With projected text, Qwen3.5-9B scored 0.737 against 0.672 for GLM-5.2 without it; moderate model differences were not excluded.Source: §3.1 · p. 15, Figure 4 · p. 15

3.2. Form and interface of the supplied knowledge

• Projected text scored 0.777 pooled, compared with 0.685 when shuffled, 0.717 as tower JSON, 0.597 for documents behind search, and 0.638 for tower query tools.Tool comparisons were descriptive: models chose whether to call, and tower conditions also included derived quantities.Source: §3.2 · pp. 15–16, Table 2 · p. 16
• In the rotor duty-cycle diagnosis, GLM-5.2 avoided the trap in four of five projected-text trials versus three of five with tag names, while tool use exposed the unit’s state machine.Across the evaluation, prompt text raised every model’s score, whereas tower tools helped only models that called them and read the returned information.Source: §3.2 · p. 16, Figure 5 · p. 17

3.3. Operating lessons in the prompt and in the journal

• Incorporated lessons lowered scores when removed, and lessons or records placed in prompts outperformed the same records left in the journal behind search.For the two tasks with incorporated lessons, removal reduced scores from 0.885 to 0.642 and from 1.019 to 0.856; journal search was called in 4 of 80 diagnosis trials. Absolute scores varied between batches, so only within-batch comparisons were interpreted.Source: §3.3 · pp. 16–18, Table 3 · p. 18

3.4. Live runs through the safety layer

• 25 of 61 commands reaching the safety layer were refused, and every run ended with the baseline restored and read back.Refusals included 15 commands for setpoints not allowed in the run and nine exceeding cumulative change limits.Source: §3.4 · pp. 18–19
• 12 of 14 live runs brought the controlled variable into its band; projected text and query tools reached the band equally often.The dew-point moving average entered the band 15 to 21 minutes after the first accepted command in all eight runs; the temperature task entered it in four of six runs.Source: §3.4 · p. 19, Figure 6 · p. 20
• Factual claims unsupported by that turn’s evidence comprised 24 % to 68 % with GLM-5.2 and 4 % to 19 % with gpt-5.6-sol, and reported changes sometimes had not occurred.In at least four runs, responses reported setpoint changes that had not been made.Source: §3.4 · pp. 19–20

4.1. Projected knowledge across model sizes

• Projected text raised the element score similarly across model sizes, and the 9-billion-parameter model scored above the largest model without it.This gain occurred for every model only when structured knowledge was placed in the prompt as text.Source: §4.1 · pp. 20–21

4.2. Depth through operating history

• In the two tasks with incorporated lessons, every model used the lesson; lessons left in the journal were found almost only by the largest model.Prompted records performed comparably to incorporated lessons when the task’s incident was known, but a new situation would require searching the full, growing journal.Source: §4.2 · p. 21

4.3. Depth through physics and the record of actions

• The derived quantities and their assumptions were not required by evaluation tasks or included in projected text, so their value to agents remains untested.Tasks combining derived quantities, such as an energy balance across the desiccant rotor, would test this depth.Source: §4.3 · pp. 21–22
• Tower queries can trace statements to their source nodes, while an independent safety layer bounds commands and its log records actions.In live runs, agents reported unsupported readings and actions; some setpoint changes they reported had not been made.Source: §4.3 · p. 22

4.4. Extension to many similar systems

• For similar systems, recurring parts could be gathered into a common tower, but extraction and journal-lesson incorporation must be automated for the approach to scale.Hand-writing a tower for each system would be no better than the text it replaces.Source: §4.4 · pp. 22–23

4.5. Limitations

• The tests covered one plant, nine tasks, and four models; projected-text gains were measured on tasks whose answers it largely stated.Comparisons beyond the three hypotheses, the lesson-form experiment, and live runs were exploratory, with few trials per setting; physical derivations were not tested.Source: §4.5 · p. 23

• 0.605 to 0.802: projected tower text raised primary-trap avoidance across four open-weight models in a preregistered evaluation (odds ratio 5.45, Holm-adjusted one-sided p = 6.0 × 10−7).The projected text raised element scores by a similar amount in models from 9 to about 750 billion parameters.Source: §5 · p. 23
• Operating lessons were used when incorporated into the tower or placed in the prompt as records, but seldom when left for agents to retrieve from the journal.Removing an incorporated lesson lowered the element score by about 0.20; the record restored it only in the prompt.Source: §5 · pp. 23–24
• 12 of 14 runs brought the controlled variable into its target band, and no setpoint changed except through the invariant safety layer.Source: §5 · p. 24

• Younghwan Joo is credited with conceptualization, methodology, software, validation, formal analysis, investigation, data curation, visualization, and original-draft writing.Source: CRediT authorship contribution statement · p. 24
• Sung-il Kim is credited with supervision, project administration, and funding acquisition.Source: CRediT authorship contribution statement · p. 24
• Both authors are credited with writing-review and editing.Source: CRediT authorship contribution statement · p. 24