Oversight Ontology

OWL 2 DL ontology implementing the Domain Constitution — formal conditions under which human corrections on LLM-agentic output constitute valid oversight signal for RLHF training.

Overview

The domain constitution defines five conditions (C1–C5) that must hold simultaneously for edit-traces to qualify as valid preference data:

Condition	Name	Description
C1	Shared Persistent State	Agent and human operate on a shared, evolving computational state
C2	Compositional Task Layering	Sessions compose into dependency chains
C3	Grounding in Observable Reality	Success criteria are predicates over production metrics
C4	Information Asymmetry	Human corrections are based on information not accessible to the agent
C5	Consequential Grounding	Output ships to production with measurable consequences

The ontology formalizes these conditions in SHOIQ description logic (OWL 2 DL profile), enabling automated classification of workflow sessions as:

• FullOversight (gamma = 5) — all conditions satisfied
• PartialOversight (gamma in {3, 4}) — most conditions satisfied
• InvalidOversight (gamma <= 2) — insufficient conditions

Files

File	Description
OversightOntology.owl	OWL 2 DL ontology (RDF/XML). 25 classes, 17 object properties, 12 defined classes.
create_ontology.py	Generates the ontology programmatically using owlready2.
verify_ontology.py	Verifies TBox consistency, instance classification, condition independence, subsumption, and monotonicity using HermiT reasoner on 5 hand-crafted individuals.
full_abox_verify.py	SQL-based classification on 2,892 sessions + HermiT verification on stratified sample (50/50 match). Requires PostgreSQL connection to rlhf-signals database.

Key Results

• TBox consistency: SATISFIABLE (0.26s, HermiT)
• Condition independence: all 5 conditions are formally independent (no condition is entailed by the other 4)
• Subsumption: ValidOversight is a strict specialization of OntoChatGPT_Control — every valid oversight instance satisfies ontology-controlled output conditions (C1, C3), extending them with C2, C4, C5
• Monotonicity: adding assertions can only increase the oversight grade
• Full dataset classification: 24 FullOversight / 1,970 PartialOversight / 898 InvalidOversight (from 2,892 sessions)

Usage

pip install owlready2
python create_ontology.py    # generates OversightOntology.owl
python verify_ontology.py    # runs HermiT verification

Related Papers

• Ovcharov, V. (2026). Edit-Trace Oversight: Scalable Alignment Signal from Agentic Workflows. arXiv preprint (under review).
• Ovcharov, V. (2026). From Ontology-Controlled Systems to Oversight-Controlled Training: Formal Foundations for Human-LLM Alignment Signal Validation. Target: Cybernetics and Systems Analysis (Springer).

Intellectual Lineage

This work extends the principle of ontology-controlled systems (Palagin, 2006) — where formal ontological structure governs system runtime behavior — from the level of system output to the level of human oversight over system output. The evolutionary line:

1. Ontology-Controlled System Architecture (Palagin, 2006)
2. Ontology-Controlled NL Text Processing (Palagin et al., 2012, 2020)
3. Ontology-Controlled LLM Output — OntoChatGPT (Palagin et al., 2023)
4. Ontology-Controlled Human Oversight — Domain Constitution (this work, 2026)

License

MIT