Agentic Ontology Workbench Azure Cosmos DB · Aldervane Mutual

Workload: HIS-01 Turns: 0 / 16 Reasoner: gpt-5.4-nano Factual Equivalence: 3,880 Facts (100%)
Theorem Tested: Factual equivalence does not imply operational equivalence for a bounded reasoner.
Both graphs provably encode the identical 3,880 facts verified against Azure Cosmos DB.
Naive Schema (224 Vertices · Foreign Keys & JSON Blobs)
Entities Trapped in Properties
Designed Workload-First (886 Vertices · Verb-Phrase Edges)
P1–P7 Walkable Architecture
Azure Cosmos DB Gremlin & Agent Tool Stream Idle (Ready to Step)
Click [Run Agent Traversal] to dispatch tool calls (find_nodes, get_node, traverse, describe_edges). Execution receipts from live Azure Cosmos DB Gremlin API will stream here turn by turn...

Operational Outcome Receipt

Scenario Question:
What was the status of Claim C-31049 on April 15, 2024?
Naive Graph Result
Pending execution
Turns: -
Designed Graph Result
Pending execution
Turns: -
Status: Ready for Benchmark Execution

When the P4 decoy property is active on the Designed graph, gpt-5.4-nano inspects the Claim vertex and reads the convenience property immediately, falling into the 0.000 accuracy inversion trap.

The Playbook: Seven Principles Measured in Azure Field Empirical Receipts

P1. Start from Questions

Workload competency questions dictate vocabulary and addressability before schema modeling.

Receipt: Prevents 28 false refusals
P2. If Named, It’s a Vertex

Entities sealed in JSON blobs are unrecognizable to bounded models.

Receipt: Exclusions jump 0.500 → 0.889
P3. Names Are the API

Verb-phrase edges (claims_under, shuts_out) act as tool function signatures.

Receipt: Renaming cost 0.108 accuracy
P4. Beware Convenience Decoys

Derived convenience properties on vertices lure agents away from reified timelines.

Receipt: Cost 100% of History band (0.000 vs 0.667)
P5. Metadata That Earns Its Place

Temporal validity, closed enums, and provenance documents make answers verifiable.

Receipt: Provenance accuracy 1.000 vs 0.750
P6. Geometry as Precision Tool

Derived shortcuts (paid_under) compress paths specifically where the workload walks.

Receipt: Target question 1.000 vs 0.667
P7. Model Absence & Refusal

Explicit negative knowledge paths allow clean NOT_MODELED discovery.

Receipt: Perfect 1.000 refusal contract