TL;DR for operators
An AI agent can correctly enforce every operational rule it knows and still execute invalid actions when relevant rules are missing. The distinction matters whenever an organization allows agents to change equipment states, initiate transactions, or execute workflow commands.
OntoAct1 addresses this problem by extracting conditional knowledge from historical records, checking diagnostic hypotheses against current observations, and verifying proposed actions before execution. Its results show both the value and the limits of that approach:
- Preventive checking works when rules are represented. In the Mars simulation, OntoAct redirected all 70 initial action candidates violating resource requirements correctly learned by both it and the IfR baseline. Yet 363 actual resource violations occurred elsewhere in its broader trace, involving commands without explicit typed resource-cost rules.
- Diagnosis becomes more selective. On medium Tacit-FLE maintenance tasks, OntoAct repaired 83.3% of faults versus 66.7% for IfR, with 2.4 versus 11.9 mean unsuccessful interventions per episode.
- Execution-time inference can become cheaper. On ALFWorld, OntoAct completed 132 of 134 tasks with 52.5% fewer evaluation LLM calls than WALL-E.
For maintenance assistants and enterprise workflow agents, the research supports investing in traceable operational knowledge, current-state evidence, and pre-execution verification. It does not establish that the resulting agents are safe for autonomous industrial deployment. Rule coverage, observation reliability, and knowledge-maintenance costs remain central adoption criteria.
The action passed the check. The requirement was missing.
Consider an agent preparing to execute a maintenance or production command. Historical records describe similar operations, and the agent has identified an apparently appropriate action. Before execution, it must determine whether the necessary conditions actually hold.
A missing prerequisite creates two different problems. The agent might know the requirement but fail to check it. Alternatively, the requirement might never have been incorporated into its operational knowledge.
The distinction is observable in OntoAct’s Mars experiments.
Among resource requirements correctly learned by both OntoAct and IfR, OntoAct redirected all 70 initially violating candidates before execution. IfR instead executed all 20 violating candidates found on its own trajectories, responding afterward.
Those numbers describe different candidate populations, not 70 matched interventions against 20 identical interventions. They establish that OntoAct consistently enforced the represented conditions in this particular analysis, rather than an independently measured reduction in all invalid actions.
The broader trace supplies the critical qualification: 363 actual resource violations occurred among OntoAct commands lacking explicit typed resource-cost rules (Section 4.3, Table 5; Appendix D.4).
The verifier worked on the requirements it could evaluate. Other requirements remained outside its coverage.
For an organization deciding whether an agent may execute commands automatically, checking the reliability of a verifier is therefore insufficient. The organization must also establish which operational constraints the verifier can represent and observe.
Turning incident records into conditions that can be checked
OntoAct addresses two problems that ordinary retrieval does not necessarily solve: extracting transferable knowledge from individual incidents and determining whether that knowledge applies to the current situation.
A maintenance record might associate a particular component failure with abnormal readings elsewhere in a production system. Reusing that experience requires identifying the component relationships and relevant conditions, rather than simply retrieving a document describing a similar incident.
OntoAct calls this ontological abstraction.
The framework begins with a predefined schema covering entity types, relations, observable attributes, and allowable rule forms. An LLM proposes typed assertions and conditional axioms from inherited records. Admission checks assess their structural validity, source provenance, applicable support, and contradictions before incorporating them into reusable knowledge.
Importantly, the system distinguishes historical information from current observations. A temperature recorded during a previous incident does not become today’s temperature merely because the earlier record was relevant.
This separation matters for enterprise knowledge systems in which operating procedures can remain applicable while equipment readings, inventory levels, and process states change continuously.
OntoAct then constructs diagnostic hypotheses connecting possible causes, expected observations, supporting evidence, and proposed repairs.
Rather than forcing every condition into true or false, its evaluator uses three states:
| Condition assessment | Meaning | Diagnostic consequence |
|---|---|---|
| Supported | Current resolved evidence satisfies the condition | Contributes to hypothesis confirmation |
| Contradicted | Current resolved evidence conflicts with the condition | Refutes the hypothesis |
| Unresolved | Adequate evidence is unavailable | Leaves the hypothesis open for inspection |
A hypothesis is observationally confirmed when all its expected conditions are supported, refuted when at least one is contradicted, and otherwise remains open.
Confirmation is provisional. Agreement with expected readings does not establish unique causation. Subsequent observations or unsuccessful repairs can overturn the assessment.
This mechanism changes how an agent chooses its next action. It can inspect an unresolved condition, discard a contradicted explanation, or proceed with a repair backed by a currently supported hypothesis. The framework requires diagnostic backing for repairs that need it, while permitting information-gathering actions before a diagnosis is confirmed.
Its symbolic verifier also checks proposed actions against represented prerequisites and prohibitions. Contradictions can block execution, but missing observations do not automatically prohibit every action.
The design combines selective evidence acquisition with preventive action checking. Its benefits depend on the correctness and completeness of both the learned conditions and the observations used to evaluate them.
Repair success improves when the agent narrows its candidates
The paper introduces Tacit-FLE, a synthetic maintenance benchmark built on Factorio Learning Environment plants.
Agents inherit distributed maintenance records and must locate and repair hidden faults. Its Chain, Conjunction, and Elimination tasks test whether agents can combine dependency relationships, complementary evidence, and inconsistent explanations.
Across 104 episodes covering small, medium, and large plants, OntoAct leads the reported overall repair-success comparison under the GPT-5.6 Luna condition.
| Tacit-FLE plant scale | OntoAct repair success | IfR repair success | OntoAct / IfR mean spurious interventions |
|---|---|---|---|
| Small (28 episodes) | 78.6% | 64.3% | 1.9 / 6.4 |
| Medium (36 episodes) | 83.3% | 66.7% | 2.4 / 11.9 |
| Large (40 episodes) | 90.0% | 65.0% | 1.6 / 17.8 |
Main benchmark comparison: Section 4.2, Table 1. Spurious interventions are unsuccessful executed replacements or refills.
The mechanism becomes clearer in the candidate-level diagnostics.
On medium tasks, OntoAct proposed an average of 3.36 distinct replacement targets per episode, compared with 5.14 for WALL-E and 9.00 for IfR. The true fault appeared among OntoAct’s first three proposed targets in 25 of 36 episodes, compared with 10 for WALL-E and four for IfR.
These candidate measures include unexecuted proposals, so they should not be interpreted as repair success. Their purpose is to assess diagnostic search quality before and alongside intervention.
Separate hypothesis-ledger diagnostics reinforce the account: across plant scales, OntoAct removed 95.6–99.6% of tracked off-target hypotheses while retaining 71.4–90.0% of target-aligned hypotheses.
The advantage is particularly pronounced on Chain tasks requiring reasoning across records. OntoAct does not uniformly outperform IfR on Elimination tasks.
For a maintenance-assistance product, these findings suggest a measurable design objective: prioritize the correct fault early and reduce unnecessary physical interventions. The experiments establish that improvement within synthetic plants, not a quantified reduction in real maintenance costs.
Better decisions can require fewer model calls
OntoAct also evaluates adaptation in Mars and household task execution in ALFWorld, testing capabilities beyond hidden-fault diagnosis.
On seven modified Mars worlds, OntoAct achieves an aggregate achievement Score of 15.92% versus 10.64% for retrieval-augmented ReAct*. This Score geometrically aggregates achievement-success rates; it is not the percentage of episodes completed.
The Mars comparison also reports approximately 45% fewer evaluation tokens per environment step for OntoAct: 0.84k versus 1.54k. The figure excludes separately incurred knowledge-construction costs.
In ALFWorld, OntoAct completes 132 of 134 unseen tasks versus WALL-E’s 129. Its average evaluation usage is 13.7 LLM calls and 32.16k tokens per task, versus WALL-E’s 28.8 calls and 47.70k tokens.
That corresponds to 52.5% fewer calls and 32.6% fewer evaluation tokens. The paper attributes much of the call difference to deterministic condition checks replacing additional LLM outcome-prediction calls.
These are distinct benchmark comparisons, not one combined estimate of agent performance.
Model-condition comparisons provide further sensitivity evidence. On medium Tacit-FLE, OntoAct’s repair success rises from 83.3% under Luna without reasoning effort to 97.2% with medium reasoning effort and 100% under Terra without reasoning effort. This demonstrates sensitivity to model capability, not the isolated contribution of stronger reasoning to the architecture.
The cost question also extends beyond inference during execution. Medium Tacit-FLE knowledge construction averages 48.8 LLM calls and 58.7k tokens per plant for OntoAct, exceeding several baselines in construction-token usage.
A business deployment must therefore account for knowledge extraction, admission checks, rule updates, and their reuse across tasks. Lower evaluation-token consumption alone cannot establish lower total operating cost.
The coverage problem remains before deployment
Three constraints materially limit how the results should be used.
First, OntoAct’s entity vocabulary, observation mappings, and permissible axiom templates are predefined. The system induces knowledge within that structure; it does not autonomously discover an unrestricted ontology. Its corpus-derived knowledge also remains fixed during evaluation, even as episode-level hypotheses change.
Second, the benchmarks are simulated. Tacit-FLE uses synthetic maintenance records, Mars uses modified game mechanics, and ALFWorld uses text-based household tasks. Some comparison methods are reconstructed or ported, and Mars averages use five evaluation seeds per world. The experiments do not establish real industrial reliability or independently reproduced performance from released research code.
Third, encoded verification is not comprehensive verification. The 363 uncovered Mars resource violations demonstrate that rule completeness is an empirical requirement, not something established by successful enforcement of known conditions. Incorrectly induced rules and unreliable observation mappings create additional failure paths.
For enterprise automation, Cognaptus therefore infers that evaluation should cover more than successful task completion. A credible pilot should measure represented-prerequisite coverage, accuracy and freshness of observations, unsupported actions, unnecessary interventions, and total knowledge-maintenance costs.
Actions whose critical requirements cannot be reliably represented or observed warrant independent controls rather than unrestricted automatic execution. That is an engineering implication of the identified failure modes, not a safety protocol validated by this paper.
Execution authority should follow evidence coverage
OntoAct’s contribution is to connect inherited operational experience to explicit, revisable hypotheses and machine-checkable action conditions. Across three simulated environments, that architecture accompanies stronger aggregate outcomes in selected tasks, more selective diagnosis, and lower evaluation-time inference usage.
Its most consequential result may also be its clearest limitation. Preventing every violation of a known requirement says little about requirements the system has failed to learn.
An enterprise agent should earn execution authority through demonstrated coverage of relevant constraints, trustworthy current observations, and validated intervention outcomes. OntoAct provides a concrete architecture and benchmark evidence for developing those capabilities. It does not eliminate the need to establish their completeness in the environment where the agent will operate.
Cognaptus: Automate the Present, Incubate the Future.
-
Donghyeok Lee and Bong Gyun Kang and Junyong Ahn and Sungroh Yoon (2026). OntoAct: Hypothesis-Based Neuro-Symbolic Agents through Ontological Abstraction. arXiv:2610.21983. https://arxiv.org/abs/2610.21983 ↩︎