TS Imagine’s trading AI draws a hard line between reasoning and calculation

The architecture matters more than the chat

TS Imagine has described the control architecture behind TSIQ, its AI system for capital-markets workflows. The company says it separates a deterministic plane—trading and risk engines plus certified reference and market data—from a probabilistic plane containing models and agents that reason and interact with users. An ontology, semantic layer and agent-control framework form the seam between them. TS Imagine

That division is more consequential than another conversational interface. According to the company, the AI does not independently invent or calculate the numbers used by the workflow; it relies on the verified core. TS Imagine says that core covers more than 31 million financial instruments. TS Imagine

A metric name is not a metric definition

The company illustrates the grounding problem with average daily volume. “ADV” can refer to a local listing, a primary listing or consolidated volume; it can use several time windows; and it can include different classes of reported liquidity. Each calculation may be mathematically valid while still being wrong for a particular order or review. TSIQ’s proprietary ontology is intended to bind those choices to certified data structures so the agent receives a defined market, window and liquidity type rather than guessing from a label. TS Imagine

For data-agent builders, the practical lesson is to keep business meaning executable. A glossary or embedding can help retrieval, but a production agent also needs machine-readable definitions tied to the data and calculation path it is allowed to use. TS Imagine says it spent five years, more than 50,000 person-days and $100 million on TSIQ’s data foundation, including three years of internal use before client release; those are vendor-reported figures, not an independent performance benchmark. TS Imagine

Audit the action, not only the answer

TS Imagine also says every TSIQ action is recorded and every generated order can be traced to the conversation and context that initiated it. Customers can choose how much authority to delegate, with approvals and user permissions governing action. TS Imagine

That suggests a concrete deployment checklist beyond text-to-SQL accuracy: preserve the resolved metric definition, source data, calculation version, caller identity, permission decision, model interaction and resulting action in one trace. The model can remain probabilistic while the data contract and authorization boundary stay deterministic.

The post does not publish an error rate, latency study or independent evaluation, so it should not be read as evidence that TSIQ eliminates ambiguity in practice. It is useful instead as a first-party production design: models reason over a constrained, auditable core rather than becoming the core themselves. TS Imagine